Alvaro Lopez Ortega / 2026-08-22 Briefing

Created Sat, 22 Aug 2026 21:08:55 +0000 Modified Sat, 22 Aug 2026 21:09:31 +0000
5070 Words

The Dutch data protection authority fined Uber €825 million ($966 million) for using automated systems to suspend drivers without proper notice or human review, violating GDPR. Uber plans to appeal. Meanwhile, security experts warn that AI agents are increasingly weaponized by attackers, urging organizations to adopt AI-driven red teaming and treat every agent as a privileged identity, as adversaries will attack regardless.

🤖 AI & Automation

Uber hit with nearly $1B fine over algorithmic driver suspensions

The Dutch Data Protection Authority fined Uber €825 million ($966 million) for using automated systems to suspend driver accounts without proper notice or human review, violating the GDPR. The case originated from complaints by 171 French drivers and affected drivers’ income. Uber plans to appeal, arguing that few drivers were affected and that the policies included human review.

AI slop is good for business if you know what you’re doing

AI slop—low-quality, mass-produced AI content—can be a business opportunity for those who understand how to leverage it, according to the article. The piece suggests that one person’s irresponsible use of AI creates openings for others to capitalize on.

Why your local LLM feels dumber than it is

Local LLMs often underperform reference benchmarks because of differences in hardware, software, and sampler settings, not just model quality. To assess your setup accurately, run representative benchmarks and examine logits and KL divergence to quantify deviation from the baseline. Be wary of low KLD claims that lack full disclosure of the evaluation environment.

Anthropic appears to be A/B testing reduced effort levels in Claude Code

Anthropic is A/B testing a server-side change in Claude Code 2.1.236+ that reduces the “high” effort setting to the previous “low” value (10/100), with no mention in the changelog. The change affects only certain users in the test group, and the author discovered it by inspecting requests after suspecting their own tools were broken.

ElevenLabs, TwelveLabs, ThirteenLabs

The author investigates the prevalence of AI startups named with a number plus “Labs,” finding examples from 0 to 99. They highlight the oddity of seventyonelab.com, a retro web-design portfolio, and ponder why this naming scheme is so widespread.

80% of developers find AI coding more addictive than helpful

A survey of 305 developers found that 80% feel their use of AI coding tools is more like dependence than an advantage, with many working after hours and losing sleep. While AI can boost productivity and promotion odds, it also fuels workaholism, burnout, and frustration from debugging flawed code. Trust in AI accuracy has dropped to 29%, down from 40% last year.

Digging the grave of my skills: Hollywood creatives training AI to do their jobs

Hollywood writers, directors, and producers are earning $12 to $200 an hour to train AI models for companies like Anthropic and OpenAI, teaching them to replicate tasks such as screenwriting and production scheduling. This work comes amid a severe industry slump, with LA shoot days down 48% since 2021 and US film/TV jobs down 28% since 2022, as AI begins replacing roles. Many creatives take the jobs reluctantly, with one likening it to being “handed a shovel and asked to dig the grave of my profession.”

Munder Difflin – Agent harness to run an office of your clones

Munder Difflin is a free, open-source multi-agent harness that runs locally on your laptop, using your existing keys and subscriptions to create autonomous clones of individuals. These clones collaborate through end-to-end encrypted messaging and a shared knowledge base to handle tasks like code reviews, bug fixes, and documentation, escalating only key decisions to humans. Optional paid plans add 24/7 cloud VMs and encrypted clone-to-clone messaging across an organization.

Embedded AI

“Embedded AI” is a new book by David Such that teaches engineers and makers how to add machine learning to embedded systems, covering hardware selection, data preparation, deployment, and integration. It features over 25 hands-on projects, including wake-word detection and AI noise suppression, using boards like the Arduino UNO and Raspberry Pi Pico.

The Instant team joins OpenAI

The Instant team is joining OpenAI, with new signups for Instant Cloud closed and existing users required to migrate within 12 months. All cloud apps will shut down on August 31, 2027, backups remain available for a year after, and subscriptions started after July 31, 2026 will be fully refunded. Instant remains open source, and a guide is available for self-hosting and migrating apps.

Linus Torvalds uses AI to debug an Intel GPU driver bug

Linus Torvalds fixed an Intel GPU driver bug where the flat CCS storage offset was incorrectly rounded up, causing reserved compression hardware memory to be handed out as usable VRAM and overwritten, leading to black screens on Battlemage G21. The one-line fix changes the rounding to round down, excluding the CCS storage from the memory allocator. Torvalds noted that AI-assisted debugging helped narrow down the issue, though the AI repeatedly doubted it was solvable.

Robot comment classifier

A developer created a classifier to detect robot-generated code comments after encountering AI-written comments that included hallucinated or misleading claims about code behavior. The classifier correctly identified several AI-generated comments with high confidence, but also misclassified some, showing imperfect accuracy on both robot and human examples.

Inherent (DeepMind alumni, $50M) says Faraday agent beats GPT-5.5 at reproducing research findings

Inherent, a London AI lab founded by DeepMind alumni, says its Faraday agent outperformed frontier models like OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8 at independently reproducing published scientific findings. The system runs on a much smaller 27-billion-parameter model and uses reinforcement learning to develop “research taste,” a skill the startup aims to apply toward its broader goal of building AI scientists.

Zalando, Zara, ASOS bet on AI virtual fitting rooms to improve online shopping and cut returns

Apparel retailers including Zalando, Zara, and ASOS are investing in AI-powered virtual fitting rooms to improve online shopping experiences and reduce costly returns. However, as the article notes, the technology still has limitations, with a test of Zalando’s tool producing awkward fits and unclear results.

Narrowing US-China AI gap as low-cost Chinese models lure businesses

Chinese AI models are rapidly closing the gap with US rivals, now leading in usage and cost while matching top-tier performance at a fraction of the price. Budget-friendly releases like Moonshot’s Kimi K3 have driven widespread adoption, with Chinese models dominating downloads and being used by major US companies despite Washington’s chip export curbs.

Ox Alpha, stealth AI model with 1M-token multimodal context and 100T tokens/day, viral on OpenRouter

An anonymous AI model called Ox Alpha was released for free on August 20, boasting a 1-million-token multimodal context window, a 131,072-token output limit, and the capacity to process 100 trillion tokens per day. The AI community is speculating about its origin, with the strongest evidence pointing to Chinese lab Zhipu as the likely developer, though DeepSeek, SpaceXAI, and Microsoft’s unreleased MAI model have also been suggested.

Greg Brockman expands role at OpenAI, controls product and scaling teams after exec departures

Despite a turbulent year of lawsuits and executive departures, Greg Brockman has quietly consolidated power at OpenAI, becoming the effective second-in-command and day-to-day boss. His authority expanded to include product strategy and the company’s entire “scaling” arm, particularly after other leaders left, solidifying his role as OpenAI prepares for its IPO.

🔒 Security & Hacking

If you’re not using AI to attack your own systems, your adversaries will

AI agents are increasingly used by attackers to hack organizations, while also creating new vulnerabilities through data-integration channels and unmanageable machine identities that bypass traditional security. Experts urge defenders to treat every agent as a privileged identity and to adopt AI-enabled automated red teaming, since adversaries will inevitably attack—whether or not organizations pay for their own security testing.

A journalist was banned from a Flock Safety convention, so they hacked it

A journalist who was banned from a Flock Safety convention circumvented the ban by hacking into the event. The incident highlights tensions between the surveillance company and press access. Details of the hack were not fully disclosed.

InjectionBunny, a NTFS3 SUID injection for privilege escalation

Anubis is a proof-of-work system designed to protect websites from AI scrapers by adding a challenge that is trivial for individual users but costly for mass scraping. It is a temporary measure while better headless-browser fingerprinting is developed. Users with plugins like JShelter must disable them for the site to work.

Stalking the Wily Hacker: 40 years later

A retrospective marks the 40th anniversary of Clifford Stoll’s pioneering hunt for a hacker, reflecting on how that early case shaped modern cybersecurity practices. The piece revisits the challenges of tracking digital intruders and the lasting lessons for today’s security landscape.

💻 Software & Dev Tools

NetBSD and My Life (2005)

Gary Rolland, a UK network administrator, uses NetBSD 2.0.2 on 29 high-end servers to support over 4,800 users, handling MySQL, Apache, Postfix, and Samba traffic. He contrasts NetBSD’s reliability with his previous Windows servers, which frequently failed and disrupted his personal life. NetBSD has proven dependable and significantly improved both his work and home life.

hdiutil is deprecated in macOS 27 Golden Gate

macOS 27 Golden Gate deprecates hdiutil in favor of diskutil image, though some options like -puppetstrings and -srcfolder flags are missing. In a test, diskutil was faster and produced smaller DMGs than hdiutil, but it failed on a root-owned file without prompting for authentication, while hdiutil succeeded after an admin prompt.

Make your logo extra bright on HDR screens

Certain logos on LinkedIn appear brighter or whiter on HDR screens because a gain-map has been added to the JPEG image. LinkedIn is the only social network found that doesn’t strip out these gain-maps, though they can also be used on personal websites. The author created a browser-based utility with Claude Code to help others add this effect.

Scrap

The author arrived in Pittsburgh with romanticized winter expectations but struggled with freezing conditions while renovating an unheated house. After a cast-iron bathtub was stolen from their yard, they learned about the city’s scrappers and later willingly gave away another tub, even helping scrappers haul an old furnace.

Learning about “The Unix Time-Sharing System”

Ken Thompson began developing Unix after AT&T exited the Multics project, continuing his disk throughput research on a spare PDP-7. In three weeks, he completed the operating system, including a compiler, editor, assembler, loader, and user protection. The article also credits Unix’s file descriptor system as a major innovation over Multics’s single-level store.

A Friendly Introduction to Racket

Racket is a modern descendant of Lisp, a language family known for innovations like garbage collection, first-class functions, and the REPL, as well as treating code as data. The article traces Lisp’s evolution from its 1958 invention through Scheme to Racket, which now focuses on language-oriented programming, and notes its current use in education, research, and industry.

The New MCP Roadmap

The updated MCP roadmap outlines five priority areas: agentic messaging primitives, HTTP-native transport unification, agent identity and security, improved primitives, and SDK developer experience. Key initiatives include server-initiated events, maturing the Tasks extension, adopting DPoP, and enabling progressive discovery for tools.

A formal degree and algorithmic problem-solving is the answer. Always has been

A formal computer science degree and strong algorithmic problem-solving skills are essential for anyone using AI to program, just as surgical training is required to use a scalpel. Without this foundation, even well-meaning individuals make serious mistakes, leading to poor engineering priorities, fragile systems, and career failure. The author’s anecdotes show that lacking these fundamentals results in misplaced focus on trivial details and ultimately unsuccessful outcomes.

There’s no reason for software to be slow anymore

LLMs are drastically reducing the cost of performance engineering, making it practical to build highly specialized, workload-specific software that was previously too expensive to develop. Experiments with the FRE regex engine show that AOT-compiled code can improve longer query performance by 2–4x, though it may hurt short query performance due to compilation overhead.

hdiutil is deprecated in macOS 27 Golden Gate

In macOS 27 Golden Gate, hdiutil is deprecated in favor of diskutil image, though some options like -puppetstrings and certain create -srcfolder flags are missing. Unlike hdiutil, diskutil image does not prompt for authentication when encountering root-owned files, causing failures unless those files are removed. However, diskutil was significantly faster, produced a smaller disk image, and excluded the ~/.Trash/ folder by default.

NetBSD and my life (2005)

Gary Rolland, a UK network administrator, describes how his company runs 29 mission-critical NetBSD 2.0.2 servers supporting over 4,800 users. The servers handle MySQL, Apache, Postfix, and Samba, pushing over 870GB of data daily. He credits NetBSD’s stability with reducing emergency callouts and improving his personal life, unlike his previous Windows servers that frequently crashed.

An arcade fighting game over SSH

An arcade fighting game can now be played entirely over SSH at sshfighter.com, requiring no installation or download. It features hand-drawn pixel sprites, ranked matches, replays, and a bot API, with 18 fighters and active online competition.

A 2026 Survey of Rust GUI Libraries

A 2026 survey of Rust GUI libraries tests frameworks by building a QR code generator. Azul fails due to inability to read system fonts and poor editor support, while Blinc has unfinished components and non-standard reactive behavior. The article concludes with usability ratings and a summary table.

nss-userhosts - hosts files for unprivileged users

This Rust-based glibc NSS library lets unprivileged users define personal host entries in $HOME/.config/userhosts/*.conf, supporting both IPv4 and IPv6 lookups. To use it, build and install libnss_userhosts.so.2, then add userhosts to the hosts line in /etc/nsswitch.conf. Reverse DNS resolution and full XDG directory support are not yet implemented.

There continue to be reasons for software to be slow

The author argues that LLMs can reduce the cost of performance optimizations, but the claim that software will no longer be slow is flawed because key assumptions often fail in practice. Tolerance for slowness tends to increase—such as with monorepo git performance and LLM autocomplete latency—meaning expectations shift rather than demand speed. Despite supporting performance work, the author remains skeptical of broad predictions that LLMs will eliminate slow software.

OTel Isn’t Going Well (And I Made A Spreadsheet About It)

OpenTelemetry’s perceived slow progress is caused by a “three-way crash” between a strict binary stability gate, a small maintainer bench, and an enormous scope of supported languages and frameworks, which discourages marking features as stable. To manage this complexity, the project splits stable core from community-maintained contrib integrations, but the result is uneven language support and lengthy semantic-convention debates.

Stop Making TUIs

The author argues that developers should move away from terminal and command-line interfaces, since AI tools now make building native user interfaces much easier. They describe creating several apps—including a Markdown viewer, a SageMath calculator frontend, a music player, and a wiki tool—without hand-writing the UI code. The piece concludes that AI-assisted native UI development is a practical and enjoyable alternative to traditional TUI/CLI work.

Compile-Time Improvements in LLVM 23

LLVM 23 achieved a -6.75% compile-time improvement in -O3 builds, driven by optimizations to hash tables (linear probing, compact occupancy, xxh3 hashing), SmallVector’s grow path, and dominator tree representation/construction (child-sibling storage, edge lists, avoiding successor materialization). These changes reduce allocations, improve cache behavior, and enable better inlining.

The cool things of Gleam

The article praises Gleam’s simplicity and readability, highlighting features like built-in @deprecated attributes, first-class todo, simple generics, and powerful case expressions for pattern matching. It also appreciates Gleam’s choice to avoid exceptions in favor of types like Option.

Roomy is generally available

Roomy, an open-source Discord alternative for the decentralized web built on Bluesky’s AT Protocol, is now generally available. The release follows significant protocol work, including the proposed “Arbiter” standard for group and role management, and the client is being prepared for private spaces and a thin-client interface.

📱 Gadgets & Hardware

Casio decides it’s about time the simple digital watch got a little smarter

Casio’s new F-B100W digital watch adds Bluetooth connectivity and step tracking while retaining a two-year battery life. It modernizes the classic simple digital watch with smart features without sacrificing longevity.

Chinese robot runs 100M sprint quicker than Usain Bolt’s world record

A Chinese-developed humanoid robot named Lightning ran the 100m in 9.32 seconds, surpassing Usain Bolt’s human world record at the World Humanoid Robot Games in Beijing. The event featured over 2,000 robots competing across 51 events to showcase China’s robotics progress amid intensifying competition with the US. Other achievements included a robot high jump of 2.88 meters, exceeding the human record.

Initial focus for our partnership with Motorola is a regular non-folding device

The initial focus of the partnership with Motorola is a regular, non-folding device. The article content provides no further details, as it only displays a placeholder notice.

PiKVM project has gained support for web-camera forwarding

PiKVM has added an early-access feature for official V3/V4 devices that forwards a local web camera to the remote host as a standard USB device, supporting up to 1280x720 resolution. Enabling it requires an OS update, installing the ucamera package, and configuration changes, and it may cause device failures; V3 users are advised to add heatsinks to prevent overheating.

Turning My CASIO F-91W Into a Contactless Payment Device

The author describes hacking a classic Casio F-91W watch to enable contactless payments, avoiding the need to unlock a phone or pull out a card. They explain how NFC payment cards work, using a microchip and antenna powered by electromagnetic induction from payment terminals. The project blends nostalgia with innovation by embedding this modern payment capability into the iconic watch.

Nvidia to raise prices 15%+ on Vera Rubin, Grace Blackwell systems in early 2027

Nvidia has notified major customers that server prices will rise by more than 15% in many cases, driven by soaring memory chip costs. The increases apply to systems shipping in early 2027, including those with Vera Rubin and Grace Blackwell chips, and vary by chip generation and memory configuration.

Amazon hikes prices on Echo, Fire TV, Kindle, eero due to memory and storage costs

Amazon raised prices across several devices, including Echo speakers, Fire TV models, Kindles, and eero routers, citing significant increases in memory and storage component costs. The move follows similar price hikes by other tech companies amid a memory chip shortage driven by AI demand. Amazon said it absorbed the costs as long as possible and will offer promotions throughout the year.

📊 Economy & Trade

Anthropic IPO filing will show AI backlash as a risk factor, sources say

Anthropic is preparing for a massive IPO, potentially valuing the company around $2 trillion, but will cite public and political backlash against AI data centers as a key risk factor. A Gallup poll shows 70% of Americans oppose such construction, and politicians are enacting restrictions that could slow the compute capacity essential to Anthropic’s revenue growth.

Hook, hold, harvest and hide: Meta’s alleged strategy laid out in first week

California and 28 other US states are suing Meta in a trial alleging its “hook, hold, harvest, hide” business model intentionally addicts children and collects their data without parental consent. Meta denies the claims and faces potential damages of up to $200bn, plus court-ordered product design changes. The trial is expected to last six to eight weeks.

Canada suspends trade negotiations with USA and match tariffs dollar for dollar

Canada has suspended trade negotiations with the U.S., citing last-minute changes to proposed terms as unfair and unreliable. In response to the U.S. imposing 50% tariffs on roughly $28 billion of Canadian goods, Canada will match those tariffs dollar for dollar. Canada reaffirmed its strategy of building domestic strength and diversifying trade partnerships.

$40T Debt, 6.7% Mortgages and $5 Diesel

President Trump is confronting significant economic headwinds, including a $40 trillion national debt, mortgage rates at 6.7%, and diesel prices reaching $5. These factors represent major challenges for his administration’s economic agenda.

Canada vows to match Trump’s 50% tariffs after trade deal talks fail

The US and Canada failed to reach a last-minute deal to lower tariffs, with each side blaming the other for the collapse. Canadian PM Carney accused the US of adding unfair demands, while the US said Canada sought extra concessions. The US imposed tariffs on about $20 billion of Canadian goods, and Canada vowed to retaliate, escalating the trade conflict.

Canada will match US tariffs ‘dollar for dollar’ as trade talks break down

Canada will impose reciprocal tariffs “dollar for dollar” on US goods after trade talks broke down, with Prime Minister Mark Carney suspending negotiations over what he called unfair and uneconomic last-minute changes to US terms. The collapse leaves Canada facing new 50% US tariffs on about 5% of its exports, including wine, dairy, and clothing.

New U.S. tariffs to take effect after midnight, Canada plans to match levies

The U.S. imposed 50% tariffs on $28 billion in Canadian goods after trade talks collapsed, prompting Prime Minister Mark Carney to suspend negotiations and plan retaliatory tariffs after Labour Day. Carney said the U.S. “asked for too much and offered too little,” and most premiers backed his decision, though Alberta’s Danielle Smith urged resuming talks.

S&P 500 CEO median pay hits $17.3M, widening CEO-worker ratio to 312-to-1

In 2025, S&P 500 CEO median pay hit a record $17.3 million, with the average CEO-to-worker pay ratio widening to 312-to-1 when excluding Elon Musk. Average CEO compensation rose 21% to $22.8 million, the highest on record, driven largely by stock awards. Including Musk’s reported $158.3 billion package, the average jumps to $340.1 million and a 5,387-to-1 ratio.

US Debt-to-GDP Ratio

The US Debt Clock tracks the national debt using official U.S. Treasury data, projecting estimated changes between updates as of March 19, 2026. It reports the federal debt and debt per person, noting the debt grows by roughly $83,000 per second. The article also highlights that New York City alone carries over $103 billion in debt.

Cheap energy, land, proximity to Beijing made Ulanqab, China a data center hub: ~100 centers (Wired)

Ulanqab, Inner Mongolia, has become a major AI data center hub due to cheap energy, abundant land, and proximity to Beijing, with nearly 100 centers built or under construction since 2016. Chinese firms like DeepSeek, ByteDance, and Alibaba are investing heavily, but the city faces a water scarcity challenge that could strain local resources.

🎭 Culture & Oddities

Skylab completists, clear some shelf space for 26 DVDs

A new 26-DVD collection captures recovered Skylab footage, including downlinks and silent reels that were once thought lost. The release is the result of decades of archive hunting.

Belgian car salesman becomes prince after DNA test proves royal parentage

A 26-year-old Belgian car salesman, Clément Vandenkerckhove, has become a prince after a DNA test confirmed he is the son of Prince Laurent, the king’s brother. He has inheritance rights to Laurent’s private estate but receives no royal allowance, has no official duties, and is not in line to the Belgian throne. He may keep his mother’s surname rather than adopt the royal family name.

Wildfires across Europe detonate forgotten World War bombs

Wildfires across Europe, intensified by climate change, have detonated forgotten World War I and II bombs and mines in France, Belgium, Germany, and the Netherlands, endangering emergency crews. The blazes have exposed hidden munitions in former battlefields, including Belgium’s High Fens and near Bordeaux, with experts warning that climate change is reactivating previously stable weapons. This lingering lethal legacy from past wars is now posing new risks as fires sweep across scorched landscapes.

A Kantian Critique of “Sorry” by Justin Bieber

The article argues that Justin Bieber’s “Sorry” fails as a genuine moral apology under Kantian ethics because it treats the apology as a strategic tool to regain a relationship rather than as a duty owed to the wronged person. The song’s focus on whether forgiveness is still possible and on the wrongdoer’s own losses makes it a hypothetical imperative, not a categorical one, so the apology is not made in good faith as moral acknowledgment.

Never Be Angry at Work

Anger at work is toxic because it alienates colleagues and leads to being excluded from decision-making, making you less effective. While anger often comes from caring deeply about work, it must be managed and channeled productively rather than expressed openly.

4 in 10 travellers visit tourist sites to capture social media content

A new study shows that 4 in 10 travellers visit tourist sites specifically to create social media content, with Gen Z now citing TikTok as their top travel inspiration source, ahead of personal recommendations. Travel influencers influence 43% of respondents, and 37% feel pressure to make their trips look exciting online. Overall, the findings confirm that social media has fundamentally reshaped how people discover, plan, and share travel experiences.

Jason Arday Did a Lot of Harm

Jason Arday’s death is mourned, but the author argues he was not merely a victim of racism or the press, instead calling him a fraud who fabricated his autism story and plagiarized extensively. The piece criticizes institutions and advocates for failing to challenge implausible claims, saying his case exposed a broader failure to uphold the reality of autism.

You should never be angry at work

Anger at work is toxic and counterproductive, as it intimidates colleagues and leads to being excluded from decision-making, creating a self-reinforcing cycle of isolation. Although anger often stems from caring about your work, which can make you competent, the advice is to move past it rather than let it overflow.

The Schrödinger Email

The author attempted to forward emails from Outlook to Gmail but encountered a hard rejection because Gmail flagged the redirected message as suspicious due to the new sending domain’s low reputation. This illustrates that email authentication (SPF, DKIM, DMARC) from the original hop does not guarantee acceptance after forwarding, as the second SMTP delivery changes the sender’s infrastructure. The post explains how each protocol behaves under forwarding, highlighting why DKIM is especially important for preserving message authenticity.