Alvaro Lopez Ortega / 2026-09-08 Briefing

Created Tue, 08 Sep 2026 19:23:57 +0000 Modified Tue, 08 Sep 2026 19:24:47 +0000
7927 Words

OpenAI released ChatGPT Images 2.5, and denied claims it rushed an unreleased Navier-Stokes proof after seeing another researcher’s work. AI coding startup Cognition raised $2B at a $48B valuation, with about $900M annualized revenue. The US backed NextEra’s $1.9B loan to restart an Iowa nuclear plant for Google’s AI data centers. The UK announced plans to require Apple and Google to stop kids accessing nude images.

💼 Business & Finance

Ramp in early talks to raise ~$1B at ~$60B valuation, up from $44B; total raised $3B

Ramp is in early talks to raise about $1 billion in primary funding at a roughly $60 billion valuation, up from $44 billion in June. The negotiations are ongoing and could change, and Ramp declined to comment.

AI coding startup Cognition raises $2B at $48B (was $26B); run-rate revenue ~$900M vs $492M

AI coding startup Cognition raised $2 billion in a funding round led by Andreessen Horowitz and others, boosting its valuation to $48 billion from $26 billion in May. The company also reported that its annualized run-rate revenue grew to roughly $900 million, up from $492 million in May.

Circle to acquire Tazapay for $400M, among its largest disclosed deals

Circle has agreed to acquire Singapore-based cross-border B2B payments firm Tazapay for $400 million in stock, marking one of its largest deals since buying Poloniex in 2018. The acquisition would give Circle regulated “last-mile” infrastructure to connect stablecoins with local banking systems, as Tazapay processes over $25 billion in annual volume across 100 markets, with about 60% already involving stablecoins. The deal is expected to close in 2027, pending regulatory approvals.

Celero raises $275M at $3B+ valuation for long-distance AI data-center chip

Celero, a startup developing chips to enable long-distance connections between AI data centers, raised $275 million at a valuation exceeding $3 billion. The funding will support its work on improving inter-data-center links for AI workloads.

Runway co-CEO: annual recurring revenue hit $200M in Sept; expects $350M by end of 2026

Runway AI Inc. reached $200 million in annual recurring revenue in September, doubling over the past five months as businesses adopt its image- and video-generating software for marketing and ads. New customers include Dolce & Gabbana and Palo Alto Networks, and the company expects ARR to surpass $350 million by the end of 2026, according to a source.

Poseidon Aerospace raises $60M Series A from TQ Ventures; year-end pilotless cargo plane test flight

Poseidon Aerospace raised a $60 million Series A led by TQ Ventures, following an $11 million seed round, to advance its uncrewed cargo aircraft. The company plans the first test flight of its fixed-wing cargo plane, Egret, by the end of this year, with a focus on cutting cargo transport costs using combustion engines and remote piloting rather than emerging tech like eVTOL or hydrogen. Poseidon aims to serve defense and regional commercial cargo markets by operating its own aircraft rather than selling them.

Forus raises $150M at $3B valuation for AI that speeds drug prescriptions by navigating insurance

Forus, a healthcare automation startup using AI to expedite patients’ access to prescribed medications, raised $150 million in a Series C round at a $3 billion valuation. The round was led by Bain Capital Ventures, with participation from Accel, Thrive Capital, General Catalyst, Redpoint Ventures, BoxGroup, and Pear VC. Founded in 2023, the company’s platform manages steps like insurance coverage and financial assistance after a prescription is written.

Floating nuclear startup Bluecore lands $50M funding before setting sail

Floating nuclear power startup Bluecore has secured $50 million in funding to advance its plan of placing atomic reactors on barges. The company’s concept is to deliver mobile nuclear power generation directly to locations that need it.

Mistral raises €3B for European AI

Mistral raised €3 billion in a Samsung-led Series D round, valuing the French AI company at over €21 billion in the largest equity raise ever by a European tech firm. The funds will support expanded compute capacity, commercial growth, and international presence, reinforcing Mistral’s goal of becoming Europe’s full-stack sovereign AI champion with open-weight models and data-control infrastructure.

Multi-Agents LLM Financial Trading Framework

TradingAgents is an open-source multi-agent LLM framework for financial trading that coordinates specialized agents, including research analysts, a trader, and a portfolio manager. It supports multiple LLM providers, data sources, backtesting, and checkpointing. Recent releases have focused on correctness fixes, expanded model and vendor support, and improved system stability.

🤖 AI & Research

OpenAI Rolls Out ChatGPT Images 2.5

OpenAI has launched ChatGPT Images 2.5, a new image model offering up to 50% faster generation and lower latency than Images 2.0, along with sharper details and more natural lighting. It introduces features like Sketch drawing, Templates, image commenting, and prompt sharing, and is available across ChatGPT, ChatGPT Work, and Codex on all platforms. Two new API models, GPT-Image-2.5 Flare and Sunburst, were also released.

OpenAI’s Bubeck says he never asked to remove Alpöge from Navier-Stokes paper with Buckmaster

Sébastien Bubeck clarified that his message to Levent Alpöge was meant to coordinate their releases with good intentions, and he never asked for Alpöge to be removed from authorship of his work.

OpenAI Navier-Stokes breakthrough sparks controversy

OpenAI announced that an internal AI system, significantly more capable than GPT-6 Astra, produced a Lean-formalized proof solving the Navier–Stokes Millennium Prize Problem by demonstrating that a smooth 3D fluid can develop a finite-time singularity. Mathematician Tristan Buckmaster alleged OpenAI learned of his unpublished work with Anthropic’s Levent Alpöge and rushed to solve the problem, but OpenAI denied the claims and said it spent millions after hearing of a breakthrough. The proof has not been publicly released or independently verified, sparking controversy in the mathematics community.

DeepMind Maps Human DNA Variants

Google DeepMind has released AlphaGenome Atlas, a free, publicly accessible 1-petabyte dataset predicting the regulatory and molecular impact of all ~9 billion possible single-letter DNA variants in the human genome. It introduces an AlphaGenome Variant Impact (AVI) score to help prioritize variants in both coding and non-coding regions, and is already aiding researchers in identifying rare disease variants and uncovering new genetic associations for complex traits.

DeepSeek’s unprecedented hiring spree: ~150 senior engineers to fix demand/AI-agent-strained backend

Chinese AI startup DeepSeek is launching an “unprecedented” hiring push for about 150 senior engineers to overhaul its backend infrastructure, which has been strained by surging user demand and compute-heavy AI agents. The company cited rapid growth in data volumes, machine numbers, training workloads, and active users as key drivers behind the move.

Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses

A 4-bit quantization (Q4_K_M, 17GB) of Qwen3.8 27B matches the full BF16 model on benchmarks like Terminal-Bench 2.1 and GPQA Diamond, fitting on a 24GB GPU with room for context. Performance holds down to 2-bit, but 1-bit falls to near-random, and reasoning effort settings affect results more than quantization down to 4-bit.

DeepSeek v4.1 Flash is now available for internal beta testing

DeepSeek V4.1 Flash has entered internal beta testing, featuring a new architecture with native multimodal support, stronger performance, faster speed, and lower cost. To use it, keep the same base URL and set the model name to “deepseek-v4.1-flash-expires-on-0910”; pricing matches deepseek-v4-flash with a limit of 20 concurrent requests per account.

UAE-based Falcon AI NSFW classifier among top global open-source models (2025)

A UAE-based company, Falcons AI, has seen its NSFW image classification model downloaded over 50.8 million times on Hugging Face in the past month, ranking sixth among more than 1,400 models. This far exceeds DeepSeek’s downloads and highlights growing global demand for content moderation tools, as well as the UAE’s rising influence in open-source AI.

Navier-Stokes – Tristan Buckmaster [pdf]

Tristan Buckmaster offers a detailed examination of the Navier-Stokes equations, a foundational system in fluid dynamics. The PDF likely covers advanced topics such as the regularity and uniqueness of solutions, including his notable work on non-unique weak solutions. This is an important contribution to mathematical physics and partial differential equations.

The VMs Powering Mobile Agents (Instinct, Claude Code)

Mobile agents such as Claude Code and Instinct now run in provider-managed Firecracker microVMs instead of local computers, allowing phone access. Claude Code uses a custom Rust-based PID 1 and host-driven lifecycle, while Instinct rents E2B sandboxes running standard Ubuntu/systemd on the same hypervisor. This shift gives customers cloud-hosted VMs with persistent writable user disks and separate read-only system files.

I tested 10 model/harness combinations on the same Three.js task

Testing multiple model and harness combinations on a Three.js sci-fi hangar prompt showed success rates from roughly 78% to 97%, with notable trade-offs between speed and quality. The fastest run was Qwen 3.8 on OpenCode at 8m 48s with 95.64% success, while top-scoring GLM 5.3 Flash Max on OpenCode reached 96.89% success in 20m 29s. SOL 5.6 Max on Codex also performed well at 94.60%, highlighting that OpenCode and Codex harnesses generally delivered strong results.

AI cancer cures slowed by chip shortage, says Arm boss

Arm Holdings CEO Rene Haas predicted that AI will find a cure for cancer in our lifetime and that humanoid robots will become widespread within five years, but said AI growth is currently limited by a chip shortage. He dismissed fears of mass job losses and was skeptical about future UK chip manufacturing.

🔌 Hardware & Infrastructure

NextEra gets $1.9B DOE loan for Iowa nuclear plant revival with Google

The U.S. Department of Energy granted NextEra Energy a $1.9 billion loan to refurbish Iowa’s Duane Arnold nuclear plant, with Google planning to use the power for nearby AI data centers. This marks the second such loan for a nuclear restart, following a $1 billion award for Three Mile Island, as shuttered plants become a key clean power source for surging AI electricity demand. The facility is expected to restart in 2029, producing 615 megawatts.

Apple’s foldable iPhone: memory shortages push price past $1,999 goal; biggest iPhone revamp yet

Inspired by the growing popularity of foldable phones in Asia, Tim Cook pushed Apple to accelerate development of its own device around 2020. The resulting foldable iPhone, reportedly exceeding $2,000, became Apple’s biggest design change in a decade, with development challenges including memory shortages that pushed it past an initial $1,999 price target.

China AI compute up 177% YoY to 2,185 eflops in H1; targets 9,800 eflops by 2030 via $532B IT spend.

China’s intelligent computing capacity hit 2,185 eflops by the end of June, up 177% year on year, according to the MIIT. A new five-year plan targets 9,800 eflops by 2030, supported by 3.8 trillion yuan ($532 billion) in information infrastructure investment and deployment of large-scale AI computing clusters with up to 100,000 accelerator cards.

Chipmakers adopt High NA EUV lithography

ASML has secured commitments from TSMC, Samsung, and Intel to adopt its High NA EUV lithography machines for high-volume manufacturing, with TSMC targeting 2030 and Samsung 2028 to meet AI-driven chip demand. The companies are also shifting from 6-inch to larger 12-inch photomasks, which ASML and TSMC say will eliminate stitching constraints and enable more precise patterning for future advanced nodes. This mask-size push is part of an industry effort to prepare High NA EUV systems for high-volume production by 2033.

Arm debuts Neoverse CSS N4: semi-custom chip platform, 8-128 cores up to 3.8GHz on TSMC N3P

Arm debuted its Neoverse CSS N4 “Ranger” semi-custom chip platform, offering 8 to 128 Neoverse N4 cores per die at up to 3.8 GHz on TSMC’s N3P process, with support for DDR5/LPDDR6, PCIe 7/6, and CXL 4.0. Compared to the previous N2 platform, it doubles socket performance and delivers 1.75x memory bandwidth. Arm also announced new deployments of its Neoverse V3-based AGI chip at Oracle and ByteDance.

Huawei invests in Chinese lithography firms, helps them win deals with SMIC to cut foreign reliance

Huawei is investing in Chinese lithography companies and helping them secure partnerships with major fabs such as SMIC, aiming to reduce reliance on foreign suppliers. The tech giant is playing a “project management” role to build components needed for DUV machines, part of efforts to avoid export controls.

ZX Spectrum: Experimenting with 1-Bit Sound

The author experimented with ZX Spectrum 1-bit beeper sound, failing to replicate advanced PC speaker techniques but succeeding in multichannel chord playback. They also tested 1-bit PCM, which produced distorted 16kHz output, and PWM, which showed potential for better quality through variable pulse lengths. The week involved many failures alongside these successes.

Copperhead – Hardware as Fast as Software

Copperhead is an open-source AI platform for hardware teams to design, verify, and ship circuit boards through an eight-stage agent-run pipeline, from brief to dev plan, with gates ensuring artifacts exist and KiCad checks pass. It produces a complete design package including gerbers, firmware, and docs, committing once per stage. The platform is open core: the free CLI performs full design work locally, while paid cloud, team, and enterprise tiers add hosted runs, collaboration, CI, and governance.

Among European Companies That Use a CDN, Nearly 9 in 10 Use Cloudflare

A survey of 44,143 European companies using a CDN found that 89.6% rely on Cloudflare, far ahead of Amazon, Fastly, and Akamai. Cloudflare is the majority provider in every country examined, from 78.8% in Spain and Ireland to 95.6% in the Netherlands. This heavy concentration raises resilience concerns, since Cloudflare outages in 2025–2026 caused simultaneous disruptions across many sites.

Please consider running a NTP server in the NTP Pool

The NTP Pool urgently needs more servers due to constant high demand from billions of devices, and even a low-powered VPS with minimal bandwidth is sufficient to participate. Setup is easy, and timekeeping accuracy within 100 ms is acceptable. However, it is a long-term commitment: after removing a server from the pool, it may continue receiving UDP 123 traffic for years due to hardcoded client configurations.

Arm Mali G2-Ultra NX GPU: desktop-class mobile gameplay with AI-native graphics

Arm’s Mali G2-Ultra NX is the first AI-native Mali GPU, integrating neural accelerators into shader cores to enable neural super sampling and frame rate upscaling for desktop-class visuals within mobile power limits. It also delivers up to 24% higher benchmark performance and 14% higher non-AI gaming performance over the previous generation, supporting up to 120 FPS gameplay.

Europe’s largest carbon capture facility opens in the Netherlands

Europe’s largest industrial carbon capture facility has been inaugurated at Yara Sluiskil in the Netherlands, creating the first complete cross-border CCS value chain. The plant will capture up to 800,000 tons of CO₂ annually from ammonia production, ship it to Norway, and store it permanently beneath the seabed via Northern Lights. The project is expected to capture and store about 12 million tons of CO₂ over 15 years, supporting European industrial decarbonization.

🛡️ Security & Privacy

Boston Scientific left nursing its bottom line after cyberattack

Medical device maker Boston Scientific said a cyberattack in August will hurt its third-quarter and full-year sales and earnings as recovery efforts continue. The company warned that the financial impact will extend beyond the initial disruption.

How to secure hybrid meeting rooms without sacrificing user experience

Security has become the top purchase criterion for videoconferencing products, as hybrid work expands attack surfaces and regulatory pressures grow. Meeting-room solutions must bake in security by design to prevent data breaches and compliance issues, but they must also remain usable so employees don’t resort to risky workarounds.

LG accused of ’egregious invasion of privacy’ over TV data collection

LG is facing accusations of an “egregious invasion of privacy” over data collection practices on its TVs. The claims come amid scrutiny of the company’s monitors reportedly installing adware without user consent.

BigBear phishing crew nets thousands of Microsoft 365 credentials

Researchers infiltrated the admin panel of the BigBear phishing operation and discovered 5,137 stolen Microsoft 365 credentials linked to 461 organizations. The findings expose the scale of the crew’s credential-theft campaign.

Extortion crews have their eyes on high-value AI data, Google warns

Google warns that extortion crews are increasingly targeting high-value AI data, exploiting companies’ fear of exposing intellectual property. Because firms don’t want their IP leaked, they are often willing to pay ransoms, making them attractive victims.

Grindr pays £26M to settle UK privacy class action

Grindr has agreed to pay £26 million to settle a UK privacy class action. The dating app admitted no liability over allegations that it shared sensitive user data, including HIV status, with third parties.

US police fear Meta smart glasses could be used to secretly record them

US law enforcement agencies are increasingly concerned that Ray-Ban Meta smart glasses could be used to covertly record inside police facilities, prompting the NYPD to issue a memo citing videos taken in detention centers. The glasses, which can pass as ordinary eyewear, have raised broader privacy concerns, with critics noting that police surveil the public but object to being recorded themselves.

DHS ‘Predictive Policing’ Unit Is Analyzing Americans’ Financial Habits

Border Patrol is running secretive predictive policing units that analyze Americans’ financial activity and other data. They share this intelligence with local police, who pull over people not suspected of a specific crime but deemed worth searching.

We have a year to fix security everywhere

An open-weight AI model, GLM 5.3-flash, is now cheap enough to run locally and can perform dangerous hacking once safety refusals are removed, posing an imminent security threat. Initiatives like Project Glasswing and Daybreak aim to use frontier LLMs to identify and fix industry-wide vulnerabilities faster than humans, though deploying fixes remains the key challenge.

A Screensaver for LG’s Spying Smart TVs

Ad tech companies like LG use Automated Content Recognition to analyze everything displayed on smart TV screens. The website weowntheglass.com protests this by running fullscreen, slowly-moving public domain artwork on TVs to remind these companies who owns the glass, without endorsing violence. It is an independent, educational project not affiliated with any TV manufacturers.

💻 Software & Development

Reverse engineering the 2008 LogiCola, a logic drill program

Malik Piara reverse-engineered the 2008 LogiCola program to inform the redesign of LogiCola 3, responding to user feedback by reintroducing mastery-learning translation drills and a more playful interface. The updated version assigns unique color palettes to each logic set, places drills directly on the front page, and adds contextual quiz windows—while keeping the original program’s vibrant aesthetic.

I-have-ADHD: A skill to stop coding agents from burying the answer

The “i-have-adhd” project offers a skill/plugin for coding assistants that enforces ADHD-friendly output: action-first, numbered steps, and no unnecessary preamble or closers. It provides ten rules, such as leading with the next action and capping lists at five items. The skill is installable from GitHub, customizable via its SKILL.md file, and MIT-licensed.

LibreOffice breaks download records after declaring it has no AI features

LibreOffice 26.8 became its most downloaded update ever, surpassing 1 million installs in a week, after the project emphasized it includes no generative AI features. The Document Foundation clarified it is not anti-AI but has not found AI that meets its principles for default inclusion, such as privacy and user control. The record indicates there is still demand for AI-free software alternatives.

DaVinci Resolve 21.1

The article provides no substantive content, containing only website navigation and country selection options. There is no actual information about DaVinci Resolve 21.1 to summarize.

DriveSync – fast Git-styled Google Drive sync CLI

DriveSync is a Rust-based command-line tool that syncs a local folder with Google Drive, offering push, pull, and diff operations. It requires users to set up their own OAuth credentials via the Google Cloud Console. It is tested on macOS, with Linux and Windows builds available but not fully exercised.

FreeBSD 14.5-Release

FreeBSD 14.5-RELEASE has been announced as the sixth release of the stable/14 branch, focusing mainly on maintenance with bug fixes, driver updates, and refreshed third-party software. The release is available for multiple architectures via bootable ISO, network install, or USB images.

Recursion into Madness

Generative AI models fail through recursive self-feedback loops: image editing degrades into chaos after repeated localized edits, while open-ended text rewrites drift far from the source before settling into a stable “peak LLMese” style. Simple, well-specified text edits remain deterministic, but recursive divergence is significant for creative tasks.

Run an ArchiveTeam Warrior on your computer

The ArchiveTeam Warrior is a virtual machine that helps archive websites by downloading and uploading them to the archive. It is safe to run, using only your bandwidth and some disk space, and supports Windows, OS X, and Linux via VirtualBox or VMware.

Vagrant is approaching EOL. No more “Docker for VMs”

HashiCorp is deprecating HCP Vagrant, ending new box creation on October 1, 2026, support on November 2, 2026, and all operations on December 31, 2026. Users must take over hosting costs and rehost their Vagrant boxes elsewhere, while the community edition and CLI remain available. Migration tools include exporting boxes locally, guidance for S3 hosting, and static archive options.

My business partner sent a 5K vibe-coded PR that he didn’t even test

The author’s business partner used AI to quickly build an entire payments backend without testing or collaborating, resulting in broken code and inadequate documentation. The author sees this as a sign of AI-induced cognitive decline and, feeling defeated, plans to quit programming after the project.

Super Smash Bros. Melee is 100.00% decompiled

Super Smash Bros. Melee has been 100% decompiled and fully linked for the GALE01 version, with code and data totaling about 5 MB. The final commit removed an outdated TODO and included cleanup and fixes.

Githack: A Persistent Object Store for Lisp Based on Git

Githack is a Lisp object store that leverages Git as its backend, representing atomic objects as blobs and composite objects as trees with transactional support through Git branches. It enables persistent objects to reside alongside source code, supports nested transactions for savepoints, and automatically escalates to a two-phase commit protocol when a transaction spans multiple branches or repositories.

🌍 Policy, Society & Culture

The White House removed a Tetris clone called “Build the Wall” from its arcade website after The Tetris Company publicly criticized it and noted its history of pursuing copyright infringement cases. The game, part of a set of racially themed web games launched on September 4, was taken down by September 8. This marks a rare instance where a video game company’s legal warning prompted removal of White House content.

UK wants Apple, Google to stop kids taking, viewing, sharing nudes; praises Apple age limits

The UK government announced plans to introduce legislation requiring tech companies like Apple and Google to prevent children from taking, sharing, or viewing nude images on their devices. Culture Secretary Lisa Nandy told the House of Commons the bill would be brought forward “as soon as we can,” though she did not provide a specific timeline.

Australia mandates social media algorithm choice.

Australia’s government has proposed a “My Feed, My Way” initiative requiring social media platforms like Facebook, Instagram, and TikTok to offer users the choice to opt out of algorithmic content feeds, under a new digital duty-of-care framework. Platforms face fines exceeding $100 million for non-compliance, with additional protections aimed at shielding under-18s from harmful content and addictive features. The move, modeled on UK and EU frameworks, has sparked political clashes and friction with the Trump administration.

UK energy operator sticks with Palantir in £21M direct award

The UK’s National Energy System Operator (NESO) has awarded Palantir a £21 million direct contract, keeping the incumbent in place for its proprietary platform. The award is made while NESO prepares for a future competitive tender process.

Zelda: Ocarina of Time Remake Release Date and Gameplay

Nintendo announced that The Legend of Zelda: Ocarina of Time remake will launch for Switch 2 on November 5. It features a detailed graphical overhaul, manual jumping, a new dash move, a progress tracker called Threads of Time, and audio detection for playing the ocarina.

Disappointed Optimists

The article argues that design is fundamentally about behavior change, often serving corporate profit motives, which creates disillusionment among designers. It cites a survey showing designers report the highest exhaustion, poor management, and reluctance to recommend the career, partly due to AI’s impact. The author concludes that designers are “disappointed optimists,” caught between idealistic goals and the realities of the job.

How Climate Resilient Are the Largest Cities?

A study of 72 major cities using AlphaGeo’s Climate Risk & Resilience Index finds that adaptation measures can dramatically lower climate risk, causing cities with similar physical exposure to diverge by up to 33 points in resilience-adjusted risk. South Asian megacities like Ahmedabad and Dhaka face compounding crises due to high exposure and weak adaptation, while European, African, Latin American capitals and Chinese megacities demonstrate stronger resilience or rapid improvement.

An overview of how our family is prepared

The author outlines their family’s practical preparedness for likely local disruptions like water and power outages, rather than societal collapse. Their strategy focuses on sheltering in place with supplies including stored drinking and non-potable water, a camp toilet, a portable power station, propane appliances, and a rotating deep pantry to last a month.

AI Giants Work Hand-in-Hand with The Pentagon, Contracts Reveal

Despite public clashes over warfare ethics, Anthropic’s split with the Pentagon stemmed from a hidden government contract. Documents obtained by The Intercept reveal that Anthropic, Google, OpenAI, and xAI each signed deals worth up to $200 million to build AI prototypes for military advantage. The contracts show close collaboration, including two-way data sharing and engineers working alongside Defense Department officials.

We Must Return to the Office to Use AI in Person

The narrator satirically defends his company’s mandatory return-to-office policy, arguing it aligns with Mondo Mayo’s absurd core principles, its surplus of leadership, and the unfinished office construction, while also enabling competition with rivals. He ultimately concludes that despite the ridiculousness—including an office affair and AI surveillance—in-person work is necessary.

Why AI-generated slopaganda works

AI-generated propaganda, such as Trump’s fake images, is dangerous because research shows it shifts beliefs even when labeled as fake, as the effect bypasses conscious reasoning. Studies indicate AI chatbots are highly persuasive on political opinions and can even change views on conspiracy theories. However, using this power for beneficial deprogramming lacks a viable business model.

Gen Z is the first generation less cognitively capable than their parents

Dr. Jared Cooney Horvath told Congress that Gen Z is the first generation to be less cognitively capable than their parents, attributing the decline to widespread classroom screen use since 2010. In response, Denmark removed digital devices from schools and reintroduced physical books and writing, leading to improved concentration and student wellbeing.

PISA 2025 Students’ reading and mathematics performance declined across the OECD

PISA 2025 results show student performance across OECD countries fell to the lowest recorded levels in mathematics and reading, with one in five 15-year-olds now a low performer in core subjects. Science performance was stable overall, but socioeconomic background remains a strong predictor, and excessive or distracting use of digital devices is linked to lower achievement. A few systems, including Türkiye and the Slovak Republic, improved science outcomes, while Singapore and Estonia remain global leaders.

Why getting your hands dirty is good for you

Contact with soil and natural materials boosts the skin’s microbial diversity, helping defend against harmful bacteria and supporting immune and skin health. Research in Finland found that even brief exposure to soil, peat, or moss increased skin microbes, while children who played regularly in forest settings had greater microbial diversity and fewer pathogens. These benefits require frequent or prolonged contact, as the effects can quickly fade.

Do tech billionaires dream of shared abundance? We should probably just tax them

Elon Musk predicts that by 2036, AI and robots will create post-scarcity abundance, making money irrelevant, a view echoed by other tech billionaires like Mark Zuckerberg. The article argues these visions ignore how wealth will actually be distributed, citing Gabriel Zucman’s argument that taxing billionaires is necessary because wealth does not distribute itself.

Canada’s counter-tariffs take effect on various US goods

Canada’s counter-tariffs on billions of dollars in US goods took effect, targeting steel, aluminum, and dairy products in response to earlier US tariffs. The trade dispute intensified as President Trump threatened to block Bombardier aircraft sales in the US, while negotiations remain suspended.

Canada holds first defense talks with China in more than 8 years

China and Canada held their first Defence Coordination Dialogue talks in over eight years, with both sides expressing willingness to strengthen practical military exchanges and mutual trust. The resumption comes as Ottawa and Beijing work to reset relations amid growing tensions with the United States under Donald Trump, following broader efforts including high-level exchanges and eased trade tensions.

There’s a new “Google Jail” for independent wikis

Google’s March 2024 algorithm change places new domains in a “Google Jail,” indexing only their main page and severely limiting traffic for independent wikis, regardless of content quality or origin. To work around this, the author is launching new wikis on subdomains of established domains, which are indexed normally and perform well immediately.

📰 Other News

Whitehall’s latest plan to get things done: Less Whitehall

The UK government is introducing plans to cut bureaucratic delays by reducing consultations, legal caution, and judicial reviews. Ministers describe this as a war on an “industry of dither and delay.” The goal is to streamline decision-making across Whitehall.

Britain reboots its space strategy with £7.8B already on the launchpad

Britain has unveiled a renewed cross-government space strategy, backed by £7.8 billion in existing funding. The plan pairs civil ambitions with a growing focus on the military importance of orbit.

Arm pushes agentic AI and desktop-quality graphics in next-gen phone platform

Arm’s next-generation phone platform introduces new CPU and GPU designs aimed at enabling agentic AI and desktop-quality graphics. The focus is on achieving lower latency and richer gaming experiences within a tight 1 W power budget.

HPE makes its “unified storage” claim real as B10000 R6 hits GA

HPE’s B10000 R6 storage appliance is now generally available, delivering on the company’s unified storage vision by pairing block and adjacent file workloads. It offers independent scaling of performance and capacity.

Huawei proudly shows off an entirely un-American chip

Huawei showcased a chip containing no American components, highlighting its continued progress despite US export controls. The presentation carried the message “Resistance is futile,” suggesting that American restrictions have failed to halt Huawei’s technological advancement.

Serving Web from volatile Dial-up IP

A homelab with a volatile IP runs a cron job every 15 minutes to check if its URL is reachable. If unreachable, it SSHes to a reverse proxy, updates the configuration with the current IP using envsubst, and reloads lighttpd. This polling approach avoids relying on dynamic DNS notifications or extra services.

The Depths of JavaScript: Minesweeper in 247 Bytes

A JavaScript code golf version of Minesweeper was compressed to just 247 characters, featuring an 8x8 board, left-click reveals, right-click flags, recursive zero-opening, and win detection. The author spent months optimizing it with help from DNEK, using tricks like recursive HTML generation and browser auto-closing of tags to shrink the code.

Switching Password Managers in 2026

Switching password managers is made easier on iPhone and iPad through a secure iOS system interface, as demonstrated by exporting items from 1Password directly into Apple Passwords. This interoperability, supported by apps like Bitwarden and Dashlane, was enabled by passkey community standards such as the open Credential Exchange format, ensuring credentials transfer securely without vendor lock-in.

Replacing startplasma with shell scripts or dinit (Part 1)

The author explores recreating a minimal Plasma Wayland session using a simple shell script instead of the standard startup stack, setting key environment variables and launching core processes like kwin_wayland, ksmserver, and plasmashell. While the approach mostly works, they caution that manual startup is difficult due to environment variable propagation and subtle session dependencies.

Extreme Server Side Rendering

Extreme server-side rendering (XSSR) streams HTML slowly over an open HTTP connection so pages update dynamically without client-side JavaScript. User input is handled via hidden iframes and form submissions rather than JS. The article demonstrates the technique by building a Flappy Bird clone.

x86 evolution for segmentation and paging

Anubis is a proof-of-work system that protects websites from AI scraping by imposing a computational cost negligible for individual users but prohibitive for mass scrapers. It serves as a temporary compromise while better methods for detecting headless browsers are developed.

How to get a free .arpa domain

You can obtain a free .arpa domain by signing up for a Hurricane Electric IPv6 tunnel, converting its routed IPv6 prefix into an ip6.arpa DNS name, and delegating that name to a DNS provider like deSEC. This lets you host a website on a domain under .arpa, which is normally reserved for internet infrastructure, by using reverse DNS records for nonstandard purposes.

Statement on finite-time blowup for forced 3D Euler, and OpenAI’s conduct

The article shares a statement from Tristan Buckmaster, linking to his Mastodon post containing proofs related to finite-time blowup for the forced 3D Euler equations. It also addresses concerns regarding OpenAI’s conduct in relation to the matter.

Antiquated HTML Snippets and Artefacts

The articles discuss obsolete HTML snippets and artefacts created for browser compatibility, vendor extensions, and platform-specific hacks—such as the X-UA-Compatible meta tag, conditional comments for Internet Explorer, Chrome Frame, and ICBM location tags. These elements were once necessary for functionality or targeting specific browsers but are now irrelevant and persist only in legacy website markup as relics of early web development.

Finite-time blowup with smooth forcing for 3D incompressible Euler, Boussinesq, and IPM

A new mathematical result establishes finite-time blowup for solutions of the 3D incompressible Euler equations, as well as the Boussinesq and IPM models, when a smooth forcing term is present. The work demonstrates that singularities can form in finite time under these conditions, marking a significant advance in understanding fluid equation behavior.

The state of European cloud providers in 2026

The author describes their 2026 search for a European cloud provider for personal use, using a practical evaluation method focused on €10 and €20 price tiers, actual CPU allocation, benchmark results, hourly billing, IPv4 support, and traffic allowances. They compare these EU providers against American ones like Vultr for context while assessing ease of scaling and management console reliability.

Bitap: my favorite string matching algorithm

The article explains the bitap (shift-and) string-matching algorithm, which is particularly efficient for short patterns that fit within a machine word. It shows how the algorithm can be derived from a naive approach by processing text as a stream and using a bitset to represent in-progress matches, enabling simple and elegant bit-parallel updates.

Type inference has usability problems (2019)

The article argues that type inference, despite being popular in many programming languages, may harm code comprehension by increasing cognitive load, since readers must infer or look up variable types. It contends that editor tooltips are not an adequate substitute and that explicit type declarations serve as better documentation for code.

This Month in Ladybird - August 2026

In August, Ladybird added video playback on Twitch and additional YouTube formats, along with CSS scroll snap, JavaScript debugging in DevTools, pausable downloads, and session restore. The engine also saw a sustained performance focus, including a new style engine, layout caching, and off-main-thread CSS animations, with CSS parsing and painting moved to Rust. New sponsors were welcomed.

Sex, drones, and OnlyFans – Ukraine’s unlikely war fund

Ukraine is considering legalizing pornography to tax adult content creators and generate additional war revenue. Nearly 8,000 Ukrainian OnlyFans creators earned over $130 million in 2023 despite the practice being illegal, and officials estimate legalization could raise up to $25 million annually—enough for about 30,000 drones.

Nothing wants you to test its future devices

Nothing is accepting applications until September 10 for its Community Review Program, which lets selected users test unreleased hardware and software. Participants must sign an NDA, keep product details secret, and share feedback and creative content. The program is not available to users in India.

OUI-1: world’s first model for Generative UI

OUI-1 is a finetuned DiffusionGemma model that generates user interfaces in OpenUI Lang and runs on consumer hardware such as an RTX 5090. It was developed to overcome the base model’s low reliability on the Generative UI Benchmark by combining supervised fine-tuning with self-distillation across 27 component libraries, improving interface correctness while preserving speed.

LLM Attention Visualization

This article presents an interactive visualization that reveals how large language models selectively attend to previous tokens during text generation, using attention weights to show which source words influence each generated token. It highlights how this mechanism enables accurate copying of information and synthesis of ideas across phrases, such as reproducing addresses or merging concepts from multiple sentences. The tool is built with React and Transformers.js, using modified ONNX models to expose internal attention values.

DOJ Blocked ICE Agent Shooting Charge over Federal Prosecutor’s Objections

The Justice Department overruled a federal prosecutor’s strong objections and declined to charge ICE agent Christian Castro with civil rights violations for shooting Venezuelan immigrant Julio Cesar Sosa-Celis in Minneapolis, instead directing that he face only lesser false-statement charges. Castro allegedly fired through a door, wounding Sosa-Celis, and then lied about the incident. He still faces state felony charges in Minnesota, but Texas has refused to extradite him.

Connecting the Machines

Herdr’s 0.9 update adds multi-machine support, letting users connect remote machines over SSH and manage their workspaces and agents alongside local ones in a single TUI. The client now renders the outer UI while servers retain their own sessions, enabling agents to run on 24/7 machines without keeping a laptop open. Future plans include cross-machine agent collaboration and easier account-based connections, but the current CLI still operates within one server only.

Don’t ban children from social media, Estonia tells EU

Estonia opposes the EU’s plan to restrict children’s social media access, urging that platforms adapt rather than impose a ban. Its digital minister calls for making child-safety recommendations under the Digital Services Act legally binding. The EU is expected to announce an age limit on September 16.

10x More Efficient Pretraining

Magic’s pretraining recipe is now over 10x more compute-efficient than leading open-weight base models, matching DeepSeek V4 Pro Base with roughly 50x fewer FLOPs and outperforming all public open base models at larger scale. They validated results using heldout evals designed to avoid contamination, including code, reasoning, and knowledge benchmarks across 167 domains. The work focuses on scaling laws and data mixing to improve efficiency for coding and autonomous AI R&D.

Young Americans have never witnessed a functioning foreign policy

Younger Americans have only experienced dysfunctional U.S. foreign policy, leaving them without a model of effective international engagement. This generational gap reportedly shapes their expectations and criticisms of American diplomacy.

Why I’m Not Excited About the Graphene OS and Motorola Partnership

GrapheneOS has announced a partnership with Motorola to produce smartphones running its custom Android-based OS. The author, however, expresses skepticism and raises multiple questions and concerns about the collaboration, explaining why they are not optimistic about it.

C*: Unifying Programming and Verification in C

C* is a proof-integrated extension of C that embeds verification capabilities, including a symbolic execution engine and LCF-style proof kernel, allowing programmers to interleave proof-code with implementation and update proof state in real time. It uses C as the common language for both programming and proof development, supporting reusable proof libraries and programmable automation. Evaluations on C programs and the pKVM buddy allocator’s attach function show it verifies a broad subset of C and handles real-world reasoning tasks.