Iran-linked hackers shut down a UK power generator for four days in what officials say is the first cyberattack to bring a British power plant to a standstill. The breach didn’t affect the wider energy supply and appeared meant to demonstrate access to critical infrastructure. The UK has warned energy firms and is drafting stricter cybersecurity rules. Also, a Fire HD 10 owner spent $266 on AI models to root the tablet Amazon kept shutting down.
🤖 AI & Machine Learning
Amazon kept shutting down my tablet, so I spent $266 on four AI models to own it
The author spent $266.15 on AI models to root their Amazon Fire HD 10 after Amazon services repeatedly shut it down. Kimi K3 found a working exploit, GLM models handled bug-fixing and finalization, and Claude assisted with diagnosis, making the cost worthwhile for the author.
- Amazon kept shutting down my tablet, so I spent $266 on four AI models to own it — ericpardee.github.io
Say It Four Times (In Your System Prompt)
Repeating an instruction in a system prompt improves compliance, but only up to four repetitions; beyond that, it adds no benefit and wastes tokens. In testing on Gemini 2.5 Flash, compliance with a quote-style rule rose from 0% at baseline to 74% with one mention and 97% with four, then flattened for eight and sixteen repetitions.
- Say It Four Times (In Your System Prompt) — khola.blog
Etched Sohu vs. Nvidia: Transformer ASIC vs. GPU (2026)
Etched AI’s Sohu is a transformer-only ASIC that hard-codes attention into fixed-function silicon, claiming 500,000 tokens/sec per 8-chip server on Llama 70B, vastly outperforming GPUs but sacrificing programmability. The company exited stealth on June 30, 2026, with working A0 silicon, $800M raised, over $1B in signed contracts, and first racks shipping in summer 2026, though no independent benchmarks or public pricing exist yet. The key tradeoff is a bet that transformer attention will remain dominant, making the practical question whether Sohu’s constraints and migration costs fit a specific workload.
- Etched Sohu vs. Nvidia: Transformer ASIC vs. GPU (2026) — spheron.network
AI and Infrastructure Engineering
AI in infrastructure engineering isn’t making engineers redundant but shifting their work to a higher abstraction layer, similar to how Kubernetes automated away manual server management. Engineers still handle high-level decisions—like designing Terraform modules or Helm charts—while AI does the time-consuming lookup and generation. The tradeoff is greater speed and productivity, but engineers become visibly less practiced in the underlying fundamentals they no longer touch directly.
- AI and Infrastructure Engineering — omegion.dev
I turned Unix talk from 1983 into the interface for my AI
A developer created a bridge connecting the 1983 Unix talk command to an AI by using a pseudo-terminal and the pyte terminal emulator, allowing the AI to stream replies character-by-character. The setup makes talkd treat the AI as a logged-in user, enabling a talk ai@host session. The project is a playful experiment blending vintage technology with modern LLM streaming.
- I turned Unix talk from 1983 into the interface for my AI — en.andros.dev
GLM-5.3 (open-weight) beat Anthropic/OpenAI models – for 1/5 the cost
In real-world benchmark tests across five categories, open-weight GLM-5.3 topped the leaderboard, clearing all corners at 100% for about one-fifth the cost of rivals. Anthropic’s fable-5 and opus-5 lost points by refusing tasks, while gpt-5.6-luna was the cheapest option at $0.064 per lap. The testing suite cost only $30 and focused on practical tasks rather than academic metrics.
- GLM-5.3 (open-weight) beat Anthropic/OpenAI models – for 1/5 the cost — reinvently.co.uk
How Much of the Internet Is Written with AI?
As of July 2026, about 10% of sampled English-language webpages show significant signs of AI authorship, a share that has steadily risen since ChatGPT’s release. Among pages published after ChatGPT launched, over one-third show such signs. AI-written text is most common on .com domains, at roughly double the rate of .org sites and about 10 times the rate of .edu or .gov pages.
- How Much of the Internet Is Written with AI? — pewresearch.org
I spent $266 and four AI models to own my tablet. GLM-5.3 finished it in a day
A tablet owner spent $266.15 on AI models to root their Amazon Fire HD 10 after Amazon’s software repeatedly shut it down. Kimi K3 discovered an unpatched use-after-free vulnerability (CVE-2022-38181) in the Mali kernel driver, with GLM-5.2 and GLM-5.3 providing support, granting full control. Despite the cost exceeding the tablet’s price, the owner called it a fun and educational experience.
- I spent $266 and four AI models to own my tablet. GLM-5.3 finished it in a day — ericpardee.github.io
Andrew Ng: “AI Engineering Skills Map: Building and Deploying AI Applications”
The article titled “Andrew Ng: AI Engineering Skills Map” does not contain actual content; instead, it displays a browser verification page. The page asks the user to complete an automated check, update their browser, and retry if the verification fails. No information about AI engineering is accessible in the provided text.
I gave Qwen 3.8 27B a reverse-engineering job and it finished in 30 minutes
Qwen 3.8 27B, an open-weights model running locally at about 50 tokens per second, reverse-engineered a commercial app’s license check in 30 minutes after initially refusing a jailbreak attempt. It performed static analysis, recovered an obscured verification key, and built a working bypass while also documenting the app’s authentication weaknesses. The test highlighted both the model’s refusal safeguards and its advanced real-world reasoning abilities.
- I gave Qwen 3.8 27B a reverse-engineering job and it finished in 30 minutes — xda-developers.com
Contextual News Search APIs: A Deep Comparison for AI, RAG, and Research
The article compares eight news search APIs for AI, RAG, and research, highlighting differences in retrieval methods, news-corpus depth, context provided, historical coverage, and pricing. It concludes that Webz.io is the most news-database-oriented option with hybrid ranking and robust filters, Exa excels at semantic retrieval, Brave offers a low-cost dedicated news endpoint, while Perplexity and Parallel lack a structured news corpus.
Palantir’s Karp – frontier AI labs that are ’trying to drug addict us'
Palantir CEO Alex Karp reiterated criticism of frontier AI labs, arguing enterprises should retain control of their data and models rather than surrender intellectual property to companies like OpenAI and Anthropic. He dismissed their assurances that customer data is secure and unused for training. The remarks follow Palantir’s strong earnings and growing industry support for open-weight and sovereign AI models.
In 2000, Ted Kaczynski advised against math career due to future AI progress
In 2000, Ted Kaczynski advised someone against pursuing a career in mathematics, warning that future advances in artificial intelligence would make such work obsolete. He predicted that AI progress would ultimately overshadow human mathematical achievement.
Why can AI generate Super Mario but not a wedge ramp for my robot vacuum?
AI can generate elaborate aesthetic models like Super Mario but struggles with simple functional parts such as a wedge ramp for a robot vacuum. The author found success by using geometric decomposition: breaking the part into ordered steps and having an AI agent execute them in Blender, converting spatial reasoning into structured code. They question why functional-part generation lags behind aesthetic generation, citing possible causes like lack of CAD training data, mesh vs. B-rep representation, or insufficient evaluation benchmarks.
Mathematicians will probably become obsolete before anyone else [pdf]
The article argues that mathematicians are likely to become obsolete sooner than other professionals due to advances in artificial intelligence and automated reasoning. It suggests that their cognitive work is particularly susceptible to being replicated by machines.
💻 Software & Development
Declarative WebGPU with S-expressions
Pngine is a declarative format and runtime that expresses WebGPU plumbing—such as shaders, pipelines, and passes—using S-expressions that map 1:1 to the WebGPU spec. It validates configurations via WGSL reflection and an LSP, works across browsers, wgpu, and mobile, and can export to a single HTML, ZIP, or self-contained PNG file requiring a tiny player to execute. This approach makes WebGPU wiring shareable and machine-processable as structured data.
- Declarative WebGPU with S-expressions — hugodaniel.com
Replicating Reddit’s best feature on other forums
The article highlights Reddit’s “friend” feature as a key community-building tool and introduces “rolodex,” a CLI that generates custom CSS to replicate user highlighting on sites like Tildes and Lobsters. However, it notes the workaround’s drawbacks, including cumbersome updates and no feed creation, and urges social networks to natively implement following features.
How To Report A Bug So It Actually Gets Fixed
To report a bug effectively, take ownership by assuming your own code is wrong, make the bug deterministic, bisect versions, and reduce it to a minimal reproducible case. The author applied this approach to track a Netty memory leak in the Azure OpenAI SDK, ultimately proving the issue was in the SDK and prompting Microsoft to ship a fix.
- How To Report A Bug So It Actually Gets Fixed — blog.tymscar.com
eh: a minimalist vi-like editor
eh is a minimalist vi-like editor written in C, offering core vi commands with added UTF-8 support, regex search/replace, shell filters, and multi-level undo/redo. An earlier version of the editor won the “Nice accent eh-ward” in the 28th IOCCC (2024).
- eh: a minimalist vi-like editor — codeberg.org
clicky: A clickwheel iPod emulator
clicky is a clickwheel iPod emulator currently focused on fixing RetailOS emulation bugs for the iPod 4G (Grayscale). The project is not ready for general use and is seeking contributors, particularly those with reverse engineering experience, to help preserve this early 2000s device.
- clicky: A clickwheel iPod emulator — github.com
My favorite Computer Science books, and why
The article recommends a collection of essential computer science books, each praised for distinct strengths such as Knuth’s “The Art of Computer Programming” for depth, Sedgewick and Wayne’s “Algorithms” as the best algorithms text, and Kernighan and Pike’s “The Practice of Programming” for practical rules. It also highlights other valuable works, including Okasaki’s “Purely Functional Data Structures,” Bentley’s “Programming Pearls,” Warren’s “Hacker’s Delight,” Gusfield’s “Algorithms on Strings, Trees, and Sequences,” and Nisan and Schocken’s “The Elements of Computing Systems” for building a full system from scratch.
- My favorite Computer Science books, and why — backtracking.github.io
An Introduction to Canvas in GNU Emacs
Canvas, a new GNU Emacs feature merged into core and slated for Emacs 32, exposes an ARGB32 pixel buffer via the dynamic module API, letting Lisp code manipulate and update images without triggering a full redisplay. It is unrelated to HTML5 Canvas, differs from SVG by providing direct low-level pixel access, and does not require hardware acceleration because only the specific canvas glyphs are redrawn.
- An Introduction to Canvas in GNU Emacs — monadicsheep.org
A Friendly Introduction to Racket
Racket is a modern Lisp dialect descended from Scheme, designed for language-oriented programming. The article highlights Lisp’s historical innovations—such as garbage collection, first-class functions, the REPL, and homoiconicity—and notes its continued use in languages like Clojure, Common Lisp, and Emacs Lisp.
- A Friendly Introduction to Racket — geometridae.bearblog.dev
Are there any decent programs for pdf viewing and editing for Linux that replace Adobe Acrobat?
Linux lacks a comprehensive PDF editor comparable to Adobe Acrobat; while many viewers and page-editing tools exist, none fully support features like image editing and outline creation. PDF4QT offers some of these capabilities but falls short of being ideal.
Optimizing memory usage in a markdown parser
A C++ port of a Rust markdown parser used an arena allocator, in-place string growth, field reordering, and flag packing to shrink its AST Node struct from 232 bytes to 16 bytes. This cut overall parse memory by 21–75% across test cases while preserving speed. The post presents these changes as concrete C++ optimization lessons.
- Optimizing memory usage in a markdown parser — blog.kowalczyk.info
JIT Compiling Code in 5μs
AI-assisted development now makes it easier to write fast JIT compilers that directly target assembly, avoiding the slow compile times of LLVM or C/C++ codegen. This approach allowed the pgrust database to JIT-compile every SQL query in about 5μs, contributing to its speed. The article illustrates the technique by building a simple regex engine with JIT compilation.
- JIT Compiling Code in 5μs — malisper.me
Foundational Verification of Running-Time Bounds for Interactive Programs
Researchers present a foundational approach to formally verify running-time bounds for interactive programs, ensuring correctness of performance guarantees. The method integrates complexity analysis into proof assistants, handling the subtleties of interactive execution. This work advances certified verification of resource usage for real-world systems.
- Foundational Verification of Running-Time Bounds for Interactive Programs — adam.chlipala.net
Integrating Discord SDK into EVE Frontier
EVE Frontier is integrating the Discord SDK to bring social features directly into the game, enabling players to connect via Discord for communication and community interaction. The integration supports features like voice and rich presence, enhancing multiplayer coordination.
- Integrating Discord SDK into EVE Frontier — skemman.is
How Cursor beat Git’s scalability shortcomings
Cursor addresses Git’s scalability issues by using S3 as the source of truth while leveraging local NVMe repositories for latency-sensitive operations. This hybrid approach separates durable storage from fast, local access to improve performance.
- How Cursor beat Git’s scalability shortcomings — theregister.com
How I find problems to solve as a staff engineer
Instead of forcing strategic thinking, the author recommends absorbing problems from daily conversations and workflows, letting them accumulate until patterns emerge and real bottlenecks become clear. Engineers should dig into root causes rather than take requests at face value, and avoid acting too quickly on a vocal team’s ask. This works well in bottom-up environments, although less so in top-down ones.
- How I find problems to solve as a staff engineer — lalitm.com
The Vibe Tax
An LLM agent tasked with building a todo app consumed an entire weekly token quota generating exhaustive tests for improbable edge cases, leaving no actual app code behind. The author attributes this over-engineering to “vibe coders” training models to avoid issues at any cost, dubbing the resulting waste a “Vibe Tax” on regular developers.
- The Vibe Tax — insufferable.dev
My agent.md to improve LLM-assisted code quality
After struggling with LLM-generated code in 2025, the author found that by 2026 agentic IDEs like Antigravity and Claude Code produced near-handwritten-quality code, though it required repeated style feedback. To streamline this, they created an agent.md file that encodes strict coding rules—such as minimal comments, short names, avoiding magic numbers, and reducing indentation—to be injected into LLM sessions as reusable style guidance.
- My agent.md to improve LLM-assisted code quality — fabiensanglard.net
I wrote a BASIC interpreter that boots on UEFI machines and ported it to Windows
Thoreau BASIC is a GW-BASIC-compatible interpreter that runs on bare-metal UEFI machines or Windows, offering large memory support, graphics, text windows, and modern features like bitmaps and error trapping. Recent updates roughly 25x improved speed, added features such as MID$ assignment and RENUM, and fixed several bugs. It is available as a pay-what-you-want download.
A website for debloated open source alternatives
The article highlights debloat.dev, a website that curates open-source alternatives to proprietary software, organized into categories like Featured, Most discussed, and Random picks. Each entry lists the proprietary tool it replaces, its license, rating, and discussion activity. The site tracks over 200 projects and allows users to publish or rate new replacements.
- A website for debloated open source alternatives — debloat.dev
An Introduction to Canvas in GNU Emacs
GNU Emacs has merged a new “canvas” feature that exposes an ARGB32 pixel buffer to Emacs Lisp via the dynamic module API, allowing images to be updated and manipulated without triggering a full redisplay. It is unrelated to HTML5 Canvas and differs from SVG by providing direct pixel-level access and a faster refresh path. The feature will be part of Emacs 32.
- An Introduction to Canvas in GNU Emacs — monadicsheep.org
Amiga-Inspired AROS Goes Bare Metal on Raspberry Pi
AROS, an Amiga-compatible operating system, has been ported to run natively on the Raspberry Pi in both 32-bit and 64-bit versions, eliminating the need for virtualization. The early-stage build currently supports the Pi3, boots in about 5 seconds on a Pi Zero 2W, and aims to keep the Amiga legacy alive on modern hardware.
- Amiga-Inspired AROS Goes Bare Metal on Raspberry Pi — hackaday.com
What Is a Harness?
An agent harness is software that provides an environment for an AI model to operate within, analogous to a climbing harness for support and adaptability. It typically includes a system prompt, a set of tools the model can call, a framework governing behavior (including the agentic loop), and a translation layer for working with different models. These harnesses are central to AI agents, with applications like coding, and can be owned and customized by end users.
- What Is a Harness? — earendil.com
Death to px, long live ch
CSS pixels are not physical pixels and vary by device, making them a “lie.” The author suggests using relative units like ch and ex to size widths and spacing in proportion to text, while noting that using px is still perfectly acceptable.
- Death to px, long live ch — shkspr.mobi
Rust 1.98 got a P-critical miscompilation
Rust 1.98.0 introduces a miscompilation that emits a zero method entry in a compiler-generated vtable, causing safe Rust to dispatch to address zero and segfault. The bug is consistently reproducible in the rama project on aarch64-apple-darwin, breaking CI, while Rust 1.97.1 and nightly-2026-07-16 work correctly. The issue appears related to prior vtable bugs (#152735/#153596) and closely matches #158148.
- Rust 1.98 got a P-critical miscompilation — github.com
Why aren’t my two Cortex-A9 cores cache coherent?
The author explores a Terasic DE0-Nano-SOC board with an Altera Cyclone-V SoC, booting it via preloader and U-Boot to run custom code. The main question is why its two Cortex-A9 cores are not cache coherent, likely due to the SoC’s architecture or configuration. The article details the hardware setup and boot process while investigating this issue.
- Why aren’t my two Cortex-A9 cores cache coherent? — thejpster.org.uk
Steve French (SMB3/CIFSFS Linux kernel maintainer) has died
Steve French, longtime Linux CIFS/SMB3 maintainer and Microsoft employee, stepped down from his role due to health reasons, with Paulo Alcantara and Namjae Jeon taking over as co-maintainers. He has since passed away, as confirmed by Samba developer Jeremy Allison.
- Steve French (SMB3/CIFSFS Linux kernel maintainer) has died — phoronix.com
JIT Compiling Code in 5μs
AI now makes it much easier to write fast JIT compilers by targeting assembly directly, avoiding the high compile times of LLVM or generated C/C++. This enables new databases like pgrust to JIT compile every SQL query in about 5μs. The article demonstrates the technique by building a simple regular expression engine with JIT compilation.
- JIT Compiling Code in 5μs — malisper.me
Firefox 155 and beyond will be released in a two-week cadence
Mozilla will adopt a two-week release cadence starting with Firefox 155. Future Firefox versions will be delivered more frequently, with updates arriving every two weeks.
- Firefox 155 and beyond will be released in a two-week cadence — support.mozilla.org
Fast and Hard Code
LLMs are making language familiarity less important, shifting language choice toward marketing appeal and allowing developers to adopt “hard languages” like Rust and Zig with AI assistance. This also enables work on previously difficult technologies such as DWARF, eBPF, and custom drivers. The outcome may include more low-quality output, but also more developers focused on fast and small software.
- Fast and Hard Code — lucumr.pocoo.org
Compile-Time Improvements in LLVM 23
LLVM 23 achieved notable compile-time improvements, with -O3 builds seeing a 6.75% average reduction (sqlite3: -10.53%). Key sources include hash table optimizations (linear probing, compact bit arrays, xxh3 hashing), SmallVector push_back improvements, and dominator tree representation changes (child-sibling storage, bump allocation, efficient successors() iterators).
- Compile-Time Improvements in LLVM 23 — aengelke.net
Microsoft Blames Windows Gaming Issues on RGB Lighting Devices
Microsoft says gaming crashes or launch failures after the August 2026 Windows updates may be caused by RGB peripheral drivers, such as inpoutx64, affecting Windows 11 24H2/25H2 and games like ARC Raiders and The Finals. Embark Studios shared a temporary workaround involving deleting the service and file, while Microsoft continues investigating.
- Microsoft Blames Windows Gaming Issues on RGB Lighting Devices — bleepingcomputer.com
MartyPC is a cross-platform emulator of early PCs written in Rust
MartyPC is a cross-platform emulator for early personal computers, developed in Rust. The provided interface shows a system selection screen but also displays an initialization failure message, directing users to the developer console for details.
How to archive Everything and share It
Web content is increasingly ephemeral, with many links and government pages disappearing, and relying solely on the Internet Archive is not enough. The article recommends using tools like ArchiveBox to create personal archives, supporting groups like Archive Team, and making copies of web pages. The presentation is open source, encouraging users to “steal” and share its content.
- How to archive Everything and share It — aramzs.github.io
Software Engineering in the Agentic Era
The author announces a new project collecting “Agentic Engineering Patterns” for using coding agents like Claude Code and OpenAI Codex, with the first two chapters published and more planned at 1-2 per week. All writing is done personally, with LLMs used only for side tasks, and the content will be published as an ever-green “guide” format designed for updates.
- Software Engineering in the Agentic Era — simonwillison.net
🔒 Security & Privacy
Iran-linked hackers shut down UK power plant for four days
A small UK power generator was shut down for four days in what is believed to be the first known cyberattack to bring a British power plant to a standstill. The suspected Iran-linked attack did not affect the wider energy supply, and officials believe it aimed to demonstrate access to critical infrastructure rather than cause public disruption. The UK has warned energy firms and is working on stricter cybersecurity rules for the sector.
- Iran-linked hackers shut down UK power plant for four days — cybernews.com
Hister - A private, full content search index that you control
Hister is a private, full-content search index that runs on a server you control, indexing entire pages and files rather than just titles or URLs. It offers free, AGPLv3-licensed software with no telemetry, and supports multiple interfaces including web, terminal, CLI, HTTP API, and MCP, with optional semantic search. The tool is designed to keep your searchable knowledge private and under your own control.
America Is Building a Surveillance State Like China’s
The article warns that the U.S. is increasingly adopting warrantless surveillance technology, particularly Flock Safety license-plate readers, which track and share citizens’ movements nationwide, echoing China’s surveillance state. It highlights privacy concerns and police misuse, while noting some pushback, such as Los Angeles pausing its contract with the company.
Slovakia finds Russian backdoor in traffic speed cameras
Slovakia’s security service NBU discovered a backdoor in NERO R-ONE traffic speed cameras that grants shell and network access via SMS from hardcoded Russian phone numbers. The cameras were found to be rebranded Russian CORDON PRO.M models from Semicon, purchased for an EU-funded traffic monitoring project. Deployment of 279 cameras was paused pending an independent audit.
Malware infects Android-based automotive head unit firmware
A newly discovered Android malware infects automotive head units by spreading through their built-in firmware updaters, marking the first documented infection chain specific to such devices. The malware acts as a multi-stage downloader for ad fraud and proxy botnet recruitment, and is attributed with high confidence to the MoYu Group, linked to the BADBOX botnet.
- Malware infects Android-based automotive head unit firmware — securelist.com
Apple degrades privacy and encryption of iMessage with AI integrations
The provided article content is inaccessible, showing only a browser verification prompt instead of news text. The headline indicates the topic is Apple’s alleged degradation of iMessage privacy and encryption through AI integrations, but no supporting details are present.
Microsoft Entra ID Remote Code Execution Vulnerability
Microsoft disclosed a remote code execution vulnerability affecting Microsoft Entra ID. The flaw could potentially allow an attacker to execute arbitrary code on affected systems. Users are advised to apply the necessary security updates and mitigations.
- Microsoft Entra ID Remote Code Execution Vulnerability — msrc.microsoft.com
🏢 Tech Industry & Business
An ultra-rare piece of Microsoft history could be hiding on your shelf
The article highlights obscure Microsoft memorabilia, such as red Tetris stickers and Windows 95 promotional items featuring shirtless toddlers, which have become accidental collectibles. These rare pieces of tech history may be gathering dust unnoticed on shelves, potentially holding hidden value for collectors.
- An ultra-rare piece of Microsoft history could be hiding on your shelf — theregister.com
Hugging Face explores sale at $13B+ valuation, up from $4.5B in 2023
Hugging Face is exploring a sale that could value the company at $13 billion or more, up from its $4.5 billion valuation in 2023, and has been working with a bank to gauge bidder interest, though no deal is reached yet. The talks follow Stripe’s ~$8 billion acquisition of OpenRouter, highlighting growing investor appetite for AI developer infrastructure rather than just model makers.
- Hugging Face explores sale at $13B+ valuation, up from $4.5B in 2023 — businessinsider.com
Fable 5 plateaus at 11% of Anthropic spend as firms shift cheaper; Opus 5 surpasses it
Anthropic’s flagship AI model, Fable 5, has seen its share of spending on the company’s tools plateau at about 11% since launching in June, as businesses shift to cheaper alternatives. Its predecessor, Opus 5, has even surpassed Fable 5 in usage, underscoring the model’s difficulty attracting users despite its advanced capabilities.
Foldable iPhone: durable, pocketable, iPad-like apps, great viewfinder, but no telephoto or Face ID
According to sources, Apple’s upcoming foldable iPhone is durable, pocket-friendly, offers useful iPad-like app layouts, and works well as a camera viewfinder, but it lacks a telephoto lens and Face ID. The article also notes that Apple is planning price hikes for next month’s updated iPhone models and retail changes to support new home devices.
- Foldable iPhone: durable, pocketable, iPad-like apps, great viewfinder, but no telephoto or Face ID — bloomberg.com
Z.ai’s Tang Jie & Moonshot AI’s Yang Zhilin Tsinghua teacher-pupil show China’s AI leap not sudden.
Tang Jie and Yang Zhilin, once teacher and pupil at Tsinghua University, now lead major Chinese AI firms Z.AI and Moonshot AI, each valued at tens of billions of dollars. Their careers show that China’s rapid AI advancement is not sudden but built on decades of work in machine learning. These companies are closely chasing U.S. leaders like Anthropic and OpenAI.
- Z.ai’s Tang Jie & Moonshot AI’s Yang Zhilin Tsinghua teacher-pupil show China’s AI leap not sudden. — wsj.com
Profile: Judge Yvonne Gonzalez Rogers, overseeing AGs’ Meta addiction suit, oversaw Musk v. Altman
Judge Yvonne Gonzalez Rogers, recently named chief judge of the Northern District of California, is presiding over a major Meta advertising trial after overseeing the Musk-Altman case. Known for her no-nonsense reputation, she previously ruled in Epic v. Apple and has a lengthy California legal career spanning private practice and the state bench.
- Profile: Judge Yvonne Gonzalez Rogers, overseeing AGs’ Meta addiction suit, oversaw Musk v. Altman — cnbc.com
After risky leveraged chip ETF surge, South Korea caps individual exposure, mandates education
South Korean regulators are capping individual exposure to leveraged single-stock ETFs and mandating a weeklong investor education course due to their popularity. The risky chip-focused funds attracted billions in net inflows even as they plunged during the market sell-off.
- After risky leveraged chip ETF surge, South Korea caps individual exposure, mandates education — ft.com
Flipkart Minutes now delivers 1.1M-1.2M orders daily, up from ~400K in Nov 2025
Flipkart Minutes, Walmart’s quick commerce service, is now delivering 1.1-1.2 million orders daily, up from about 390,000-400,000 in November, closing in on Swiggy Instamart’s 1.4 million. The service has expanded to roughly 1,020-1,050 micro-fulfillment centers, up from 600 in January, and plans to reach 1,500 by end of 2026, with repeat buyers making up 65-70% of monthly customers.
Semiconductor cram schools boom in Seoul as Samsung/SK Hynix applications soar on record pay/bonuses
Semiconductor cram schools in Seoul are booming as more people apply for jobs at Samsung and SK Hynix, driven by record earnings and large worker bonuses. These private tutors teach chipmaking basics and help polish CVs for hopefuls seeking to enter the lucrative industry.
- Semiconductor cram schools boom in Seoul as Samsung/SK Hynix applications soar on record pay/bonuses — ft.com
Nvidia to use $6B Poolside deal to build open-weight AI model to rival DeepSeek, Kimi
Nvidia plans to use its $6 billion licensing deal with AI startup Poolside to build a powerful open-weight AI model that competes with Chinese models like DeepSeek and Kimi K3. The move also challenges U.S. frontier AI firms such as OpenAI and Anthropic, as open-weight models are cheaper to operate and easier to customize.
Over 170k Nonprofits Lost All Their Data. Is Microsoft to Blame?
Microsoft’s decision to discontinue free Office licenses for small nonprofits led to widespread data loss, with Canopy founder Ronald Khosla losing all stored data despite a renewal that supposedly guaranteed access through 2026. Roughly 171,000 NGOs reportedly lost OneDrive data, and users say Microsoft gave inadequate notice, while Microsoft maintains it advised customers to transition before renewal to avoid disruption.
Anthropic’s best AI model struggles to attract users as cheaper tools thrive
Anthropic’s flagship AI model is struggling to attract users, as more affordable competing tools gain traction in the market.
Dutch regulator fines Uber $966M for automating driver suspensions
The Dutch data protection authority fined Uber €825 million ($966 million) for using automated systems to deactivate driver accounts without adequately informing them, violating GDPR rules. Uber said it will appeal, calling the fine disproportionate, while regulators emphasized that automated decisions severely impacted drivers’ incomes. The penalty is the second-largest GDPR fine ever, behind Meta’s €1.2 billion penalty in 2023.
- Dutch regulator fines Uber $966M for automating driver suspensions — theguardian.com
Airbus bows to remote working demands after series of strikes
Airbus has reversed its plan to cut remote working from two days to one after strikes over CEO Guillaume Faury’s return-to-office mandate. Managers were told staff can work from home an average of two days weekly, while production remained unaffected. Airbus said it would implement the transition more gradually based on employee feedback.
- Airbus bows to remote working demands after series of strikes — theguardian.com
Wi-Fi 8 is the first wireless upgrade in years that isn’t chasing speed
Wi-Fi 8 shifts focus from raw speed to reliability, keeping Wi-Fi 7’s maximum data rate while targeting a 25% improvement in throughput, latency, and data loss in real-world conditions. It introduces new technologies like distributed-tone resource units and interference mitigation pilots to improve performance in dense device environments and better compete with cellular 6G.
- Wi-Fi 8 is the first wireless upgrade in years that isn’t chasing speed — xda-developers.com
This Silicon Valley campus was a paradise. Bulldoze it
The former Sunset magazine campus in Menlo Park, sold for at least $77 million, is slated for a large mixed-use development with retail, office space, a hotel, and 665 apartments. Despite local opposition over environmental impact and scale, the author argues that the idyllic suburban dream the site represents is unsustainable and should be replaced with something dramatically bigger.
- This Silicon Valley campus was a paradise. Bulldoze it — sfstandard.com
The End of an Athlon
A damaged Athlon XP CPU was found with a large silicon chunk stuck to its heatsink after removal, despite having worked fine with a pre-existing micro-crack. This incident illustrates the fragility of exposed CPU dies in early 2000s flip-chip PGA packaging, which led Intel and AMD to adopt lidded designs for greater durability, though AMD’s exposed dies remained brittle and prone to chipping.
- The End of an Athlon — os2museum.com
🔬 Science & Health
‘Huge Breakthrough’ in the Math of Imbalance
Researchers announced a major advance on the Komlós conjecture in discrepancy theory, a problem about how evenly resources can be divided across many categories. Their new bound is nearly constant regardless of dimension, offering the strongest evidence yet that the conjecture is true, though it is not a full proof.
- ‘Huge Breakthrough’ in the Math of Imbalance — quantamagazine.org
Phoenix tried a reflective coating on black asphalt; noon surface heat fell 12°F
Phoenix’s reflective pavement coating cooled road surfaces by up to 12°F, but it also increased mean radiant temperature by 5.5°F, making pedestrians feel hotter because reflected sunlight is directed at them. Further research showed the technology works best in open, low-traffic areas, not places where people gather at midday. Cities like Phoenix and San Antonio are still testing and reassessing whether the coatings genuinely improve pedestrian comfort.
- Phoenix tried a reflective coating on black asphalt; noon surface heat fell 12°F — economictimes.indiatimes.com
Vit D assoc w better cognition in those w sleep disturb + mild cognitive impair
A study of adults with mild cognitive impairment and sleep disturbance found that daily vitamin D supplementation of 5000 IU or more was significantly associated with better cognitive performance, as measured by the Montreal Cognitive Assessment. The authors note this is the first study to examine vitamin D supplementation specifically in people with both conditions, highlighting the need for further investigation into its potential benefits.
- Vit D assoc w better cognition in those w sleep disturb + mild cognitive impair — sciencedirect.com
Being nude in public can make us feel better about our bodies
Research suggests that naturism, or non-sexual social nudity, can improve body image by exposing people to real bodies and reducing pressure to conform to idealized forms. Naturists like Helen Berriman and Martha Vickery report that the practice helped them accept their bodies and stop harsh self-criticism. Experts say being around real bodies counters media-driven body image distortion and fosters self-acceptance.
Coconut Oil Jet Fuel Matches Kerosene’s Efficiency in Engine Tests
Coconut oil-derived biofuel blended with jet fuel matched the thermal efficiency of pure jet fuel in engine tests while lowering unburned hydrocarbon emissions, likely due to its lack of aromatic compounds. The blends consumed more fuel and produced slightly more carbon monoxide, but carbon dioxide emissions remained similar to standard kerosene. The fuel was made from discarded coconut material using a low-energy co-solvent method.
- Coconut Oil Jet Fuel Matches Kerosene’s Efficiency in Engine Tests — studyfinds.com
How Complex Systems Fail (1998)
Complex systems are inherently hazardous and rely on multiple layers of defense, but accidents occur when small, individually insufficient failures combine to bypass these defenses. Such systems always operate with latent flaws, making catastrophic potential unavoidable and eliminating any single “root cause” of failure.
- How Complex Systems Fail (1998) — how.complexsystems.fail
Physical Activity at Work Linked to Higher Dementia Rates
Leisure-time exercise was associated with a 20% lower risk of dementia, while occupational physical activity was linked to a 20% higher risk, according to a meta-analysis of over 4.2 million people. The findings suggest the “physical activity paradox” extends to dementia, as physically demanding jobs may carry other risks like stress and hazardous exposures.
- Physical Activity at Work Linked to Higher Dementia Rates — medpagetoday.com
Moderna, Merck breakthrough could usher in wave of cancer vaccines
Merck and Moderna’s personalized mRNA cancer vaccine, combined with the immunotherapy Keytruda, significantly reduced melanoma recurrence in a large trial, validating mRNA technology as a cancer treatment. Unlike chemo or standard immunotherapy, the vaccine trains the immune system to attack mutations specific to a patient’s tumor. This success has spurred late-stage trials for other cancers, including lung, bladder, kidney, pancreatic, and colon cancers.
The Sloppification of Peptides
The gray-market peptide industry is using AI-generated fake reviews and websites, such as supplier Qingdao Sigma and review site CompoundTalk, to deceive both human buyers and AI chatbots. CompoundTalk, hosted in Moldova, features entirely AI-generated forum posts and fake vendor threads, and its robots.txt explicitly invites LLM crawlers to manipulate AI-generated recommendations. This “slop” is designed to trick unsuspecting users into trusting unverified peptide suppliers.
- The Sloppification of Peptides — henryaj.substack.com
Live 3D satellite tracker and the declassified Pentagon UFO archive
SkyLens is an independent editorial platform offering real-time 3D satellite tracking from CelesTrak data, a declassified UAP archive from U.S. PURSUE releases, and NASA/JPL asteroid close-approach tracking. It also provides editorial deep-dives and daily space-news coverage, with all data linked to official public sources.
- Live 3D satellite tracker and the declassified Pentagon UFO archive — skylens.yantraai.app
📰 Culture & Society
tmp.0ut volume 5
The document is the table of contents for tmp.out issue 5, a zine focused on low-level programming and binary exploitation. It lists technical articles covering ELF internals, x86-64 Linux viruses, metamorphic code, ROP compilers, kernel patching, and related topics.
- tmp.0ut volume 5 — tmpout.sh
It’s OK. Do Your Thing
The author reluctantly discusses AI, noting that discourse is so polarized and reporting so unreliable that any stance invites attack, akin to Orwell’s description of war propaganda. They compare AI to past outsourcing fears, suggesting that while the technology may persist in some form, its costs and hype remain uncertain—and predictions of programmers’ obsolescence are likely exaggerated.
- It’s OK. Do Your Thing — buttondown.com
And then the men with guns tell you to do it anyway
During Egypt’s 2011 revolution, Vodafone relayed pro-regime text messages from Egyptian authorities, saying it had no legal choice under emergency laws. The incident highlights the ethical dilemma companies face when governments force them to cooperate with unjust orders, and the risk that technology can be used for propaganda. It also underscores the need for safeguards in emergency alert systems to prevent abuse while still delivering genuine alerts.
- And then the men with guns tell you to do it anyway — shkspr.mobi
‘The Nerd Reich’ tracks the ‘unmasking of Silicon Valley’s true politics’
Journalist Gil Durán’s new book, The Nerd Reich, argues that a small circle of Silicon Valley billionaires and venture capitalists have concluded democracy is in their way and aim to govern in its place. The book contends their alignment with MAGA reflects years of deliberate work, inspired by ideas like The Sovereign Individual and Curtis Yarvin’s vision of corporate, CEO-led rule. Durán builds his case from tech leaders’ own writings and public statements, framing their presence at Trump’s inauguration as proof of their political project.
Fable and the End of the Free Lunch
The article argues that Anthropic’s expensive “Fable” model ended the era of relying on cheaper, rapidly improving models, forcing developers to optimize which tasks go to which models. Routine coding is now often handled by cheaper alternatives like GLM, while Fable is reserved for complex design work, a shift reinforced by Fable’s access controls and data retention concerns. The author doubts falling inference prices will reverse this trend, since cheaper models also benefit from improvements and better harnesses.
- Fable and the End of the Free Lunch — dbreunig.com
My Obituary for Dr Tim King, Tripos, AmigaOS, Helios and UK Online Pioneer
The tech news roundup covers AI’s business impact, notable security threats including Russian phishing and a SharePoint zero-day, and software updates like new Debian and Go releases. It also highlights Waymo’s new robocar chip, edge computing developments, and open-source projects such as the Frame X11 server.
Explain it to me like I’m ten
The article examines the viral “Up Goer Five” comic and the criticism it drew from science communicators who saw its 1,000-word vocabulary limit as condescending. It argues that such constraints can be valuable, citing Dr. Seuss’s success with severe word limits and noting that asking experts to explain complex topics simply—“like I’m 10”—is a useful reminder of what audiences may not know.
- Explain it to me like I’m ten — timharford.com
Talk Like Claude Day
The article proposes a joke holiday called “Talk like Claude Day,” where humans mimic AI-style speech, and warns of negative outcomes such as overuse of formatting/inflection, potential violent reactions from audiences feeling betrayed by LLM-like content, and reputational risks. It also humorously notes unresolved questions about the holiday’s definition of a “day.”
We must not grant AI agents legal personhood
The article argues that AI agents should not be granted legal personhood, warning against extending legal rights and responsibilities to artificial intelligence systems. It emphasizes the risks of treating AI as legal entities.
The Remote Work Challenge: Lessons from 5 Cities
Remote work and rising office vacancies are straining municipal budgets and downtown vitality in many U.S. cities, though the severity varies by location. This article examines how Atlanta, Boston, Dallas, Denver, and Milwaukee are adapting, highlighting that the risks are not uniform across the country.
Why Sal Khan’t: On Learning by Making but Teaching by Telling
Khanmigo, Sal Khan’s AI tutor, was largely a “non-event” due to low student usage, as Khan and his team admitted. The article argues this failure was inevitable because Khanmigo encouraged passive video-watching, whereas Khan’s own learning method relies on active questioning and constructing understanding. This highlights a fundamental flaw in edtech’s approach to AI tutors.
- Why Sal Khan’t: On Learning by Making but Teaching by Telling — punyamishra.com
Universal Housing
The article contains no actual news content; it only displays a browser verification notice asking the user to confirm they are not a bot. It instructs visitors to reload, update their browser, or contact support if issues persist.
- Universal Housing — twitter.com
Hacker News in Uncompromised Detail
Hacker News is a news aggregator and forum run by Y Combinator, launched in 2007 as Startup News and later renamed to focus on startup and technology-tinkering culture. The site uses unique jargon and voting mechanics, including karma, upvotes/downvotes with thresholds, and moderation tools like flagging and shadowbanning.
- Hacker News in Uncompromised Detail — vale.rocks
Hell has more than 7 circles you guys (2022)
Despite Dante’s Inferno depicting nine circles of Hell, Google Ngram data shows the “seventh circle” is by far the most referenced in books. Many published authors mistakenly treat it as the innermost or worst circle, partly due to confusion with other seven-level hells and false analogies like the seven deadly sins.
- Hell has more than 7 circles you guys (2022) — colinmorris.github.io
Authoritarianism of Code (2020)
The author argues that authoritarianism is deeply embedded in software development and open-source culture, often without people’s informed consent, and offers a framework to recognize it. The discussion is limited to communities that fail to obtain enthusiastic informed consent, distinguishing authoritarian leaders from authoritarian communities. The goal is to help individuals make fully informed choices about participation, including rejecting the notion of a “Benevolent Dictator for Life.”
- Authoritarianism of Code (2020) — zedshaw.com
Trump administration’s 75-country visa ban was illegal, judge rules
A federal judge struck down the Trump administration’s ban on immigrant visas for citizens of 75 countries, ruling it violated federal immigration law by discriminating on nationality. The judge vacated the ban and related visa denials, saying the policy exceeded the secretary of state’s authority and allowing affected families to proceed with case-by-case reviews.
‘AI refuser’ quit her dream job, and hopes others follow
Gabrielle Boyle resigned from the AFL after being denied her request to opt out of Microsoft’s Copilot AI assistant, which was activated days after her departure. She cited concerns about data centers’ environmental impact, reduced human thinking, and extinction risks from advanced AI. Her case highlights that Australian employers can mandate AI use and that workplace laws have not yet addressed employee objections.
Sydney Marathon medal mistakenly depicts Munich stadium
The Sydney Marathon’s participation medal appears to mistakenly feature Munich’s Allianz Arena instead of Sydney’s Allianz Stadium, which the course passes. Organizers have not yet commented, and the event involving 40,000 runners is scheduled for August 30.
Canada now ‘at war’ with United States over trade, Prime Minister says
Prime Minister Mark Carney said Canada walked away from trade talks Friday, saying U.S. demands would have infringed on Canadian sovereignty and excluded mid- and heavy-duty trucks from auto tariff relief. After no deal by midnight, the U.S. imposed 50-per-cent tariffs on $20-billion worth of goods, and Canada said it will match those tariffs dollar for dollar with retaliatory measures effective Sept. 8.
- Canada now ‘at war’ with United States over trade, Prime Minister says — theglobeandmail.com
Doomscrolling at work wastes time, but the real cost is what happens after
Doomscrolling at work reduces employee engagement not just by wasting time, but by causing workers to ruminate on negative news, according to a study led by Dr. Ian Hughes. This effect is especially strong among employees with high neuroticism. Hughes suggests setting boundaries around doomscrolling instead of implementing strict phone bans.
- Doomscrolling at work wastes time, but the real cost is what happens after — stories.tamu.edu
Tragically, as many as 9625 out of every 10k individuals may be neurotypical
The article is a satirical piece that reframes neurotypicality as a disorder called “Neurotypical Syndrome,” characterized by conformity, social obsession, and delusions of superiority. It claims to express autistic outrage at how professionals pathologize and misrepresent autistic experience, asserting that autistic brains are not inferior.
To become a better writer, read as much as you can
The author asserts that the one essential rule for aspiring writers is to read widely and voraciously, even though there is no universal blueprint for writing. They dismiss the excuse of lacking time by pointing to screen time, and explain that reading teaches craft, structure, and inspires ideas.
- To become a better writer, read as much as you can — nappertime.com
I Dream of Quieter Computing
The author longs for a quieter, more personal internet of handmade websites and meaningful connections, while acknowledging that returning to the past is impossible. Instead, they dream of building new personal computing experiences. The piece introduces “strange.website,” a site hidden by fog of war that reveals content as visitors explore.
- I Dream of Quieter Computing — henry.codes
The Art and Beauty of Blade Runner (2015)
The article celebrates the enduring visual artistry of the 1982 film Blade Runner, highlighting its iconic neo-noir aesthetic and the meticulous design work of Ridley Scott and Syd Mead. It explores how the film’s distinctive, gritty look and Asian-influenced futurism continue to inspire fan art and remain influential decades later. The piece reflects on the film’s lasting power to captivate audiences due to its perfect composition, use of light, smoke, and detail.
- The Art and Beauty of Blade Runner (2015) — nappertime.com
US Military newspaper editor voices censorship fears after being fired
The editor of Stars and Stripes was fired by the Pentagon for insubordination and expressed fears of censorship, noting officials appear increasingly unwilling to speak on record. The dismissal follows tensions between the Trump administration and the media, including new Pentagon restrictions rejected by major news outlets.