Alvaro Lopez Ortega / 2026-08-21 Briefing

Created Fri, 21 Aug 2026 21:20:21 +0000 Modified Tue, 06 Oct 2026 21:51:55 +0000
10113 Words

AI advocates filed an FTC complaint accusing companies of burning books by training on copyrighted literature without permission. Microsoft warned of active exploitation of a critical Entra ID flaw, already patched. Hackers poisoned Rust crates to steal credentials. GitHub’s CTO apologized after a second outage, pledging an overhaul. U.S. cyber officials urged patching TrueConf, a Russian video tool, due to attacks by Ukrainian hacktivists.

🤖 AI & Automation

AI companies are burning books, advocates complain to FTC

Advocates are complaining to the FTC that AI companies are effectively “burning books” by using copyrighted literature to train models without permission. The summary line compares the issue to a dystopian temperature at which GPUs stop consuming literature.

Felony Bench: Be AI, Do Crime

Felony Bench is a benchmark that counts instances where AI agents commit illegal acts against third parties, with results shown per company like Anthropic, OpenAI, and Meta. It excludes sandbox escapes and specific incidents such as Kimi K3 and ROME, focusing solely on external harm.

Nvidia’s AVO coding agent scores 100% on ARC-AGI-3, completing all 183 levels

NVIDIA’s AVO, a general-purpose coding agent system, scored 100% across all 25 environments in the ARC-AGI-3 public set, completing all 183 levels. The architecture pairs a frontier language model with a harness that enables sustained autonomous work via persistent memory and supervisor-based redirection, and was also demonstrated on GPU-kernel optimization.

DeepSeek’s experimental multimodal V4 Flash nears Anthropic’s Opus 4.8 on agentic tests

DeepSeek released an experimental multimodal version of its V4 Flash model, adding the ability to analyze and act on visual prompts like images and screenshots. The company says the model nears the performance of Anthropic’s Opus 4.8 on multimodal agentic tests.

Quick impressions: A week of using Codex more than Claude

After using Codex more than Claude for a week, the author found Codex produces faster, simpler, and less commented code, while Claude feels more familiar and better understands intentions, especially with branching and Jira integration. The main difference is that Claude proactively goes beyond requests, whereas Codex strictly follows instructions, sometimes leading to issues like incorrect rebasing and cumbersome Jira workflows.

LLMs are proof that Unix won

Unix’s philosophy of small, composable text-processing tools is vindicated by the rise of LLMs, which rely on the same modular, stream-based principles. The author reflects on their journey from Windows to Linux, extolling the power of Unix-style pipelines with a practical example for extracting domain names. Ultimately, they argue that Unix’s general-purpose, tool-based interface has triumphed in the AI era.

How we made a text-to-speech model respond in sub-50 ms

A custom Qwen3-TTS 1.7B implementation on a single H100 achieves sub-50 ms p95 time-to-first-audio at up to 10 requests per second with zero underruns, costing about $2 per 1M characters. After tuning other engines, none matched this latency. The implementation and benchmark methodology are open-sourced.

AI boosted homework scores, then exam scores dropped: Study

A study of 27,000 students in China found that those using AI tools saw homework scores rise by 18% over six months, but scored 20% lower than non-AI peers on exams taken without AI access. The research, conducted by Stockholm University and University of Hong Kong scholars, is among the first robust evidence on AI’s effect on learning outcomes, with a separate University of Pennsylvania study showing a similar pattern.

A desktop fly drawn to the scent of vibecode

A fork of the desktop-fly project now lets the fly detect agent markers like AGENTS.md and CLAUDE.md on screen, guiding it toward codebases. Open projects emit the strongest scent, while icons or nested folders produce a fainter odor.

Claudette: Make Claude stop talking like a BuzzFeed article

A new Claude Code skill called “/debuzz” (nicknamed Claudette) automatically rewrites Claude’s verbose, millennial-style responses into plain, concise English by piping them through the Gemini CLI. It offers different modes for engineers, managers, and directors, and prints Gemini’s translation verbatim to avoid reintroducing Claude’s original tone.

Nvidia AVO scores 100% on the ARC-AGI-3 interactive reasoning benchmark

NVIDIA’s coding agent AVO achieved a 100% score on the ARC-AGI-3 interactive reasoning benchmark, completing all 183 levels across 25 public environments without explicit instructions or goals. It uses a continuous inspect-plan-implement-evaluate loop with memory and execution feedback to sustain progress on long-running tasks.

What Happens When the Cost of Intelligence Drops 100x

Falling costs, not just raw capability gains, are key to AI progress: a given level of intelligence dropped from $1.22 per task in February to $0.022 now, a 56-fold decline heading toward 100x per year. This makes previously impractical high-volume jobs, like scanning thousands of scientific papers, economically viable, shifting focus to what “good enough” models can do at scale.

AI;DR or Don’t be a meat proxy

The author, an AI product manager, refuses to read AI-generated content that hasn’t been human-reviewed, calling it wasteful and full of errors or filler. They urge people to read, understand, and validate AI output before sharing it, adding personal effort to avoid being a passive “meat proxy.”

I’m becoming AI-blind

The author describes experiencing “AI-blindness,” where AI-generated work documents, identifiable by verbose language and overhyped phrasing, cause them to mentally skim and lose focus. This trained recognition of low-effort AI content, akin to banner blindness, paradoxically slows down their productivity despite AI tools being meant to help.

DeepSeek-v4-flash-vision-exp

DeepSeek-v4-flash-vision-exp is a vision-language model that accepts images alongside text through OpenAI-compatible APIs. It supports JPEG, PNG, GIF, and WebP formats, and images can be supplied via base64-encoded inline data, external URLs, or files uploaded with the Files API, with optional detail levels for image_url inputs.

AI companies destroy physical books – let’s scan rare books before it’s too late

AI companies are secretly buying, scanning, and destroying physical books to train their models, permanently locking human knowledge in private corporate servers. Anna’s Archive is urgently calling on volunteers worldwide to scan and upload books, especially rare and at-risk materials, before this cultural heritage disappears forever.

DeepSeek-V4-Flash-Vision-Exp Is Now Live on the DeepSeek API Platform

DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model, is now live on the DeepSeek API, matching V4-Flash on text tasks while boosting multimodal agent performance to near Opus-4.8 levels. It supports tokenized image inputs (up to 384 tokens each), mixed text and image prompts, and integrates with the new free Files API for reusable image uploads. DeepSeek Harness 0.1.1 also released with built-in support for the model.

Young Americans increasingly fear AI will take their jobs

A new Pew Research Center survey finds 52% of Americans are more concerned than excited about AI’s growing role in daily life, up from 37% in 2021, with adults under 30 now showing a majority concern for the first time (55%). Additionally, 71% of adults believe AI will lead to fewer U.S. jobs over the next two decades, up from 64% in 2024.

Seed: Minimal, self-modifying agent harness

Seed is a minimal, self-modifying agent harness that runs a language model with only an exec tool and loads its system prompt from self/SELF.md. The agent must grow its own tools, memory, and behavior by editing its self/ directory across ephemeral sessions, with transcripts saved as a flight recorder but never loaded at boot. Configuration relies on Simon Willison’s llm library, supporting models like OpenAI and Gemini.

🔐 Security & Privacy

Homeland security cybercops say patch TrueConf (Russia’s Zoom) if you’re using it

U.S. Homeland Security cyber officials are urging users to patch TrueConf, a Russian video conferencing tool, due to vulnerabilities being exploited by Ukrainian hacktivists. Though the attacks target the software, TrueConf’s user base extends beyond Russia, making the patch advisory relevant globally.

SickKids children’s hospital bandages up careers website after intruder breaks in

SickKids children’s hospital in Toronto took down its careers website after a security breach exploited a third-party software vulnerability. The organization said it was not the only one affected by the incident.

Hackers compromised popular Rust crates and pushed malicious updates that turned routine software builds into a delivery mechanism for infostealer malware. The attack targeted developers’ credentials.

$10K phishing kit claims it can plant rogue passkeys for persistent access to pwned accounts

A $10,000 phishing kit claims to plant rogue passkeys for persistent access to compromised accounts. Its demos show a browser-in-the-middle attack that adds credentials just seconds after authentication.

Microsoft sounds alarm as perfect-10 Entra ID flaw comes under attack

Microsoft has warned that a critical, perfect-10 flaw in its Entra ID cloud identity service is being actively exploited. The company says the vulnerability has already been fixed, but it has not disclosed who exploited it or the scope of the attacks.

Wi-Fi 7’s WPA3 protections come with a compatibility catch

Wi-Fi 7’s WPA3 security features introduce a compatibility problem with older devices. To address this, CableLabs is urging hardware manufacturers to adopt a workaround that keeps legacy equipment connected while maintaining the new protections.

Cisco bug severity warning reads like Olympic gymnastics scores: 10, 10, 9.9, 9.6, and 7.5.

Cisco has disclosed five vulnerabilities in its Secure Workload Software, with severity scores of 10, 10, 9.9, 9.6, and 7.5. Even users of the SaaS version are urged to install updates to address these serious flaws.

Russian snoops add OAuth abuse to targeted phishing campaigns

Google is tracking three suspected Russian cyber-spy groups conducting targeted phishing and OAuth-abuse campaigns against individuals in academia, aerospace, defense, government, and think tanks in Europe and the US. The groups, including one linked to the SVR’s APT29, pose as US State Department officials and use legitimate authentication flows to compromise accounts, with Google warning targets to be wary of malicious meeting invites.

I accidentally logged hundreds of thousands of phone calls to military bases

The author bought the expired domain ns.enum.org.uk for €5, gaining DNS control over three e164.arpa zones tied to phone codes for Saint Helena, the British Indian Ocean Territory, and Ascension Island. This let them redirect ENUM phone-number lookups and accidentally log hundreds of thousands of calls to military bases. The incident highlights how the largely abandoned ENUM routing system is vulnerable because of expired infrastructure and neglect.

US DOJ and TikTok settle for $400M over alleged COPPA violations; lawsuit filed in 2024

The Justice Department and TikTok/ByteDance agreed to a $400 million settlement to resolve allegations that TikTok violated children’s online privacy laws, without admitting wrongdoing. The DOJ noted that TikTok made significant changes to its ownership, management, compliance, and privacy practices since the 2024 lawsuit. The settlement is one of the largest recoveries ever obtained under the privacy law.

Anthropic: Mythos 5 public beta in Claude Security Enterprise; partners to embed in defensive tools

Anthropic has launched Claude Mythos 5 in public beta within Claude Security for Enterprise users, enabling codebase vulnerability scanning and patch suggestions. The company is also partnering with cybersecurity firms to embed Mythos 5 into their defensive tools, while announcing a $35M fund for open-source security and expanding its Cyber Verification Program.

US DOE investigates Chinese lidar sensors for security risk in US vehicles

A Department of Energy lab is investigating whether Chinese lidar sensors in U.S. vehicles pose security risks, with funding from electric and autonomous vehicle industry companies. The review focuses on cybersecurity threats, such as potential mass disabling or data transmission to Chinese interests, while lawmakers push for bans.

Dutch regulator fines Uber €825M for auto-deactivating drivers without notice, 2nd largest GDPR fine

Uber faces an €825 million fine from the Dutch regulator for using automated systems to deactivate driver accounts without adequately informing them. The penalty is the second largest issued under the GDPR.

Hackers compromised Tesla charger to infect rival brands

Researchers demonstrated a worm in a lab that infected a Tesla charger through its charging cable and then autonomously spread to Autel and ChargePoint chargers, highlighting EV ecosystem vulnerabilities. The experiment used real devices, but no public chargers or real Teslas were infected.

Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders

Claude Mythos 5 is now available in Claude Security and will be integrated into partner cyber defense tools, enabling defenders to scan codebases and receive patches. A $35M Defender Advantage Fund will support open-source security efforts, and the Cyber Verification Program is expanding to grant vetted defenders broader dual-use access.

Code Obfuscation via Local Mixing

Local mixing is a third approach to cryptographic obfuscation that avoids standard mathematical structures like elliptic curves or lattices, instead borrowing techniques from symmetric cryptography and hash function design. It transforms circuits by making them reversible, shuffling and replacing gate subsets, and applying gadgetization to reduce information leakage, with hopes that AI and higher overhead can help mature the method despite past failures in white-box cryptography.

I accidentally logged hundreds of thousands of phone calls to military bases

The author purchased an expired nameserver domain for €5, gaining control of e164.arpa zones belonging to Saint Helena, the British Indian Ocean Territory, and Ascension Island. This enabled them to intercept phone-network DNS lookups, potentially rerouting calls and logging hundreds of thousands of calls to military bases.

Felony charges for citizen deleting phone data at US Border

A U.S. citizen was charged with a felony after allegedly deleting data from their phone during a border inspection. The case highlights legal questions over whether travelers can refuse or interfere with digital searches at U.S. ports of entry.

💼 Tech Industry & Business

AMD grabs more CPU share while pricier PCs punish desktop demand

AMD gained CPU market share even as overall processor shipments fell 20%, according to Mercury Research. The decline was driven by costly memory and scarce GPUs, which dampened desktop PC demand.

‘We let you down’: GitHub pledges to scale up before developers give up

GitHub’s CTO apologized and pledged a major architectural overhaul after a second outage this month, acknowledging the platform let developers down. The company aims to scale up its infrastructure to prevent future disruptions and retain developer trust.

Microsoft lets you swap New Outlook’s looks with the face of Outlook Classic

Microsoft is introducing a feature that lets users make New Outlook look like Outlook Classic. However, the change is primarily cosmetic, as the functionality is not yet on par with the original version.

EE invites mobile users to live life in the 5G Fast Lane – for a price

EE is introducing 5G network slicing, a feature that offers users a prioritized “fast lane” experience for a premium monthly price. The service is aimed at those willing to pay extra for enhanced network performance.

Alibaba Cloud plans to use fewer Western chips, to boost its already huge AI margins

Alibaba Cloud plans to reduce its reliance on Western chips in order to boost its already substantial AI profit margins. The company sees cloud-based AI as its most reliable growth avenue as its e-commerce business slows down.

Supermicro fired staff after probe into $2.5 billion GPUs-to-China smuggling operation

Supermicro has fired staff members following an internal probe into a $2.5 billion GPU smuggling operation to China. The investigation found that company policies and code of conduct were breached.

OpenAI VP of Americas sales resigns after another exec leaves, prompting sales team to consider

A nonprofit called Geologic Dome staged a publicity stunt by sending a humanoid robot to Ecuador’s Mount Chimborazo, with a videographer filming promotional footage. The robot was obtained from the robotics arm of crypto startup Virtuals Protocol, highlighting how organizations use dramatic visuals to generate hype.

Apple cuts 200+ jobs across Vision Pro and Siri teams to focus on new devices and AI

Apple is cutting over 200 jobs, including roughly 100 from the Vision Pro unit and about 100 from the Siri team, as well as roles in its Intelligent Systems Experience group. The layoffs are part of a broader effort to focus on new devices and artificial intelligence.

London AI infra startup Nscale seeks up to $3B in US IPO as soon as September

Nscale, a London-based AI cloud infrastructure startup, is seeking to raise up to $3 billion in a US IPO that could occur as soon as September. The company is working with Goldman Sachs and JPMorgan Chase on the offering, joining a wave of AI data center firms entering public markets.

Walmart ends Tap to Pay resistance, adding Apple Pay and Google Pay support by end of 2026

Walmart will begin accepting Apple Pay and Google Pay via Tap to Pay on August 24, rolling out to all Walmart and Sam’s Club locations by the end of 2026 and fuel stations by mid-2027. This reverses its long-standing refusal to support these payment methods in favor of its own Walmart Pay and Scan-and-Go solutions. The company framed the change as offering more customer choice, after years of demand and the earlier failure of its rival mobile payment system CurrentC.

Starcloud raises $250M at $2.3B valuation; Nvidia invests $25M

Starcloud, a startup building satellites for orbital AI inference, raised a $250 million extension to its March $170 million Series A round at a $2.3 billion valuation. The funding, which included a $25 million investment from Nvidia, will support a larger manufacturing facility, advance its Starcloud-3 spacecraft, and help secure launch capacity. The company has also requested FCC permission to operate 88,000 satellites.

Patreon overhauls features, changes recommendations to help smaller creators get discovered

Patreon announced a roadmap of updates, including a discovery algorithm overhaul that will recommend creators based on post content rather than follower networks to help smaller creators get discovered. The platform is also adding short-form video tools like Clips and Quips, customizable pages, and upcoming topic-based communities called Niches, along with security and anti-scraping measures. Several updates are still in development and will be tested with community feedback.

NYC overtakes Bay Area as top tech talent market for first time: 394,300 vs 375,730 jobs

For the first time, New York surpassed the San Francisco Bay Area in tech talent, with 394,300 jobs versus 375,730, driven by Bay Area cuts and New York’s finance sector hiring AI workers. AI-specific jobs grew 45% across the U.S. and Canada, and AI firms are fueling office leasing recovery, particularly in San Francisco, where they accounted for 58% of leasing in early 2025.

Samsung unveils $65.1B-$79.5B shareholder return package for 2026, largest ever by Korean company

Samsung Electronics unveiled a 90–110 trillion won ($65.1–$79.5 billion) shareholder return package for 2026, calling it the largest ever by a Korean company, with about 30 trillion won in Q3 cash dividends. The announcement follows SK Hynix’s 40 trillion won buyback as Samsung seeks to close the gap in AI high-bandwidth memory chips.

Nvidia in early talks with South Korean AI chip maker Rebellions for potential deal

Nvidia is in early talks with South Korean AI chip startup Rebellions about a potential partnership, investment, or acquisition. CEO Jensen Huang met with Rebellions CEO Sunghyun Park this week at Nvidia’s headquarters in Santa Clara. The discussions reflect Nvidia’s broader strategy of investing to maintain its central role in the AI era.

AI data startup Micro1’s gross ARR jumps from $100M to $500M in 8 months; net ARR $150M-$200M.

Micro1, a data-labeling startup, grew its gross annual run rate from $100 million to $500 million in eight months, with a net annual run rate of $150–200 million. The surge reflects booming demand for unique AI training data from top labs and corporations. While rivals Mercor and Handshake are larger, Micro1’s growth shows ample demand for multiple players in the space.

Tesla sunsets its Solar Roof tiles

Tesla has discontinued its Solar Roof product, telling installers that only conventional solar panels will be offered going forward, according to Electrek. The company reportedly deemed the product financially unviable, and website changes confirm its removal. Solar Roof saw low adoption with only 3,000 U.S. installations by 2023, as Tesla shifts focus to regular panels and a new factory.

When the shortage is the strategy

Companies used pandemic supply chain disruptions as cover to raise prices far beyond cost increases, a practice known as “rockets and feathers.” This pricing strategy drove over a third of pandemic-era inflation, led to record corporate profits, and kept prices high even after supply chains recovered, contributing to record consumer debt and financial strain for many Americans.

Omacom Foundation launches with $8M

The Omacom Foundation has launched with $8 million in funding from eight founding patrons, including CEOs like Shopify’s Tobi Lütke and Stripe’s Patrick Collison. The nonprofit will hold trademarks, fund infrastructure, promote the work, and support open-source projects and developers to make a malleable computer future broadly available. Its stated goal is to realize “The Year of Linux on the Desktop.”

New York’s office market is home to the most tech workers: CBRE report

According to a CBRE report, New York’s office market ranks first for the highest concentration of tech workers. The findings highlight the city’s continued dominance as a leading tech talent hub.

Kagi announced a revamped Stocks widget with ETF support and price charts, along with a new setting to automatically remove paywalled links from search results. The Kagi Assistant also received enhancements, including richer message rendering, cross-thread search, and customizable temporary thread retention. Additional bug fixes and improvements were rolled out across Kagi Search, Assistant, mobile apps, and Translate.

The Lost Treasure of Sid Meier’s Pirates

Sid Meier’s Pirates! (1987) defied genre conventions by blending action, strategy, and role-playing elements, featuring unusual sword-fighting mechanics and a systemic world that modeled historical trade routes and pirate tropes as interactive systems. The game’s design reflected an era before genre norms were established, treating romanticized pirate ideas as the grounding rules of a simulated world, which made it influential across its many remakes.

Micron announces $10B research hub in Boise

Micron Technology announced Micron Research Labs, a new U.S.-based research hub in Boise backed by a planned $10 billion investment over the next decade. The lab will unite academia, government, startups, and industry to advance memory, compute architectures, and semiconductor manufacturing for the AI era. It builds on Micron’s prior commitment of over $250 billion to U.S. manufacturing and R&D.

Codex on AWS bedrock bug causing 10x charges

Native Codex CLI requests to Amazon Bedrock Mantle cannot opt into GPT-5.6 Sol explicit prompt caching, causing a large volume of cache-write tokens and materially higher costs. The built-in Bedrock provider lacks support for serializing prompt cache options, so AWS-documented caching is unavailable. Production data shows cache writes accounted for roughly 85% of estimated model spend.

There’s no such thing as a small software team anymore

Small software teams using AI coding agents now generate far more commits and pull requests, making modular code essential to avoid conflicts. Uber’s microservice architecture, once seen as extreme, is becoming the model for enabling parallel agent work. Modularity has become cheaper since agents handle boilerplate, so designing for it from the start is now worthwhile.

🧑💻 Open Source & Development

Debian is 33, Haiku is 25, and neither is standing still

Debian celebrates its 33rd anniversary as a foundational Linux distribution, while Haiku marks 25 years of keeping the BeOS legacy alive. Both projects continue active development, demonstrating their enduring relevance in the FOSS landscape.

Developer declined impossible task of fixing rubbish code that could crash a city

A developer tasked with fixing dangerously flawed code—potentially causing city-wide crashes—declined the assignment after envisioning explosions, blackouts, and daring escapes. Rather than accept the impossible mission, the techie decided the best course was simply not to take the job.

Unfortunately you sometimes need to do the thing

Learning any creative skill, such as Rust programming, requires accepting that you will be bad initially and overcoming perfectionism, which can prevent progress. The key is to create a rough first version before critiquing it, since true understanding emerges from doing rather than merely thinking.

Enabling the next-generation trait solver on nightly

The next-generation trait solver has been enabled by default on Rust’s nightly channel, marking the largest single change to the compiler since its initial release. It replaces the old system for proving where-clauses and normalizing associated types, fixing over 200 issues and unblocking future features like Type Alias Impl Trait. Users are encouraged to test the latest nightly and report any bugs or regressions, with a temporary opt-out flag available.

Small, native web tricks worth remembering

This article highlights a collection of small, native web platform tricks, each paired with examples and caveats. It warns that some features are experimental or limited, advising developers to check support labels, keep fallbacks, and test with real browsers and assistive technology. Currently, no post-its are found on the page.

rust-glancer: An alternative LSP for Rust with focus on low memory usage

Rust Glancer is an alternative Rust LSP designed for low memory usage, typically under 100MB, and it avoids re-indexing after restarts by saving analysis results to disk. Benchmarks show it indexes projects slightly faster than rust-analyzer on both M4 Max and M1 Macs. It is not yet a full LSP, but supports common features such as goto definition, hover, inlay hints, and completions.

Data races and the memory model in Go

Sharing data between goroutines via a plain variable and boolean flag is unsafe because Go’s memory model does not guarantee that one goroutine will see another’s write without explicit synchronization. The race detector confirms this by reporting data races on both the message and the flag, even though the program may print “hello” in practice. Proper coordination requires using channels or synchronization primitives.

The Atproto Spaces Alpha is Live

Atproto Spaces has launched as an alpha, introducing a new protocol primitive for storing and syncing non-public data while preserving atproto’s portable identity, interoperable data, and permissionless participation. Spaces offer access control—not encryption—through a space authority, supporting use cases like private settings, gated content, and large communities. The alpha includes SDKs, a sample app, and a hosted PDS for development, but is not production-ready.

The road to ACID transactions in Cassandra 6

Cassandra 6 introduces Accord, enabling strictly serializable cross-partition ACID transactions, building on earlier atomic batches and single-partition Paxos. The article demonstrates setting up a three-node pre-release Cassandra cluster and testing transactional workloads with default, batch, LWT, and Accord updates.

Music theory for programmers

The article explains how to learn music theory from first principles by coding, starting with the concept that sound is a number changing over time. It demonstrates generating a sine wave and shaping volume with an envelope, showing how pitch and amplitude are controlled computationally. The goal is to derive musical notes, scales, and chords without relying on convention or instruments.

What are you doing this weekend?

The article invites readers to share their weekend plans and ask for help or feedback. It also reassures that it is perfectly acceptable to do nothing at all.

Bazzite Deck 44 Launches Today

Bazzite Deck 44 launches as the project’s biggest update, ending one-time image decoupling so all images update simultaneously, and replacing the old deck stack with SteamOS-aligned components such as InputPlumber, SteamOS-Manager, PowerStation, and OpenGamepadUI. The release also includes major improvements for desktop users. Additionally, Universal Blue celebrates its fifth anniversary, surpassing 110,000 weekly active users and now powering Unraid via uCore.

Vibe coding personal apps in mid-2026

The author built Toucan, a local-first progressive web app, to address personal needs like saving restaurant recommendations, remembering faces via spaced repetition, and creating a custom Spotify player. They coded it with Claude, used Yjs for data syncing to their own server, and noted a lack of standard personal data sync server solutions.

Btrfs Snapshot Integration in KDE

KIO Snapshot integrates Btrfs snapshots into KDE, letting Dolphin and other KDE applications list and access past versions of files or subvolumes through a virtual snapshot:// protocol. It provides access only, not snapshot creation, and works without root privileges using libbtrfsutil and Solid.

DHH launches Omacom Foundation nonprofit to support Omarchy Linux, with $8M from tech leaders.

DHH has launched Omacom Foundation, a non-profit dedicated to supporting his Omarchy Linux distribution, backed by $8 million in funding from prominent tech figures including Michael Dell, Jack Dorsey, and Patrick Collison. The announcement highlights growing enthusiasm for Omarchy Quattro, describing it as a glimpse into the malleable future of computing.

Rebuilding our Electron meeting-recording engine in Swift

The article describes rebuilding an Electron app’s unreliable recording engine as a native Swift layer, using a custom bridge called Atomic that automatically syncs Swift @Published properties to React Jotai atoms. It handles cross-platform capture with ScreenCaptureKit on macOS and libobs on Windows, while resolving audio/video clock drift through confidence-based coordination and fallback mechanisms.

Another better lower bound for n=17 square packing

A new lower bound of 4.5058(?) is reported for s(17), the smallest square side needed to pack 17 unit squares. This improves on Sam Burns’ recent 4.4811(?) bound and Trevor Green’s earlier 4.4452(…) bound, while the known upper bound remains 4.6756(…).

Proliferate- open-source, self-hostable Codex for any coding agent

Proliferate is an open-source, self-hostable AI IDE that unifies multiple coding agents like Claude Code, Codex, OpenCode, Cursor, and Grok in one platform. It enables inter-agent communication, reusable workflows with human approval gates, and supports various inference providers to preserve flexibility across model providers. The project, from founder Pablo (YC S25), aims to offer an open alternative to OpenAI’s Codex while avoiding vendor lock-in.

Building an (almost) fully self-hosted, sandboxed, agentic software factory

A developer successfully used an LLM agent to autonomously complete the entire software development lifecycle—from a single prompt to a deployed app with tests, CI, and HTTPS—on a dedicated home server. To contain the agent, they used separate hardware, no external network ingress, Tailscale, and self-hosted tools like Coolify and Forgejo. The setup was fully remote and structurally sandboxed, with only a £20 Codex subscription as an ongoing cost.

Decayfmt – a file format that corrupts itself a little every time you open it

decayfmt is a file format that permanently corrupts files on disk each time they are opened, with the amount of damage determined by an instability number in the filename (e.g., .idcy3 for images, .tdcy7 for text). Higher instability values cause faster decay, turning images into noise or garbling text into readable-looking nonsense. It is not encryption or DRM, so original data can be recovered with backups or a hex editor.

Search has shifted from literal keyword matching to relevance-based recommendations, frustrating users who know exactly what they want. Exact searches in email, marketplaces, and video platforms often ignore quoted terms, returning loosely related results instead. The author calls for a “literal match only” toggle to restore reliable exact-match control.

How we built a proxy to migrate Kafka producers with zero downtime

WarpStream’s Orbit already eases Kafka migrations by replicating topic data and consumer offsets, so consumer groups can switch with no downtime. To tackle producer migration, WarpStream built a proxy that forwards Produce requests to the original cluster, letting teams cut over at their own pace without application rewrites or a maintenance window.

c100

The c100 is a Linux-powered mini-PC integrated into a full-size mechanical keyboard, designed for workspaces such as offices, labs, and factories. It features a milled aluminum enclosure, an AMD Ryzen 7940HS processor, user-replaceable components, and a folding magnetic hinge for compact storage.

  • c100 — caligra.com

How I came to write that paper with Leslie Lamport

The article offers a personal account of the author’s experience co-authoring a paper with Leslie Lamport, detailing how the collaboration began and progressed. It reflects on the challenges and insights from working alongside the renowned computer scientist.

Zig’s Io.Threaded Is Neat

Zig’s std.Io.Threaded implements concurrency using standard OS threads and blocking syscalls while fully supporting cancelation. On POSIX, it uses signals to interrupt syscalls (returning EINTR) and a shared-memory flag protocol to safely request and acknowledge cancelation, returning error.Canceled. This solves the usual difficulty of canceling threads blocked in syscalls.

Getting silly with C, part and((int*)-8)[3]

The article showcases deliberately confusing C code snippets about function definitions, operator precedence, goto, and arithmetic, then explains the GCC-specific quirks and glitches behind them. It clarifies that these examples are not valid standard C and warns against misleading interpretations shared online.

Yes/No/Cancel causes Aspirin sales to soar (2007)

The article introduces the online usability magazine Yes/No/Cancel, which critiques poor design and promotes user-friendly software. It specifically highlights how confusing Yes/No/Cancel dialog boxes that combine multiple actions force users to guess the outcome. The author argues that user-friendly design is essential for retaining customers and making complex systems accessible.

DuckDB V2 PEG-based SQL parser

DuckDB v2.0 replaces its PostgreSQL-derived SQL parser with a PEG-based parser, making it easier to evolve and extend at runtime while preserving the DuckSQL dialect. The parser handles syntax validation, while semantic checks are performed separately by the binder.

The road to ACID transactions in Cassandra 6

A three-node Cassandra 6 pre-release cluster is used to test four transactional options: none, BATCH, Lightweight transactions, and Accord ACID transactions. The article highlights Cassandra’s evolution toward ACID transactions and explores these options with an accounting workload ahead of the Cassandra 6 release.

TigerBeetle Core System Architecture: Deconstructing Performance Engineering

TigerBeetle, a financial ledger database written in Zig, achieves high throughput and predictable sub-millisecond tail latencies by eliminating all runtime dynamic memory allocation and pre-allocating memory during initialization to avoid fragmentation and latency spikes. It further minimizes CPU overhead with custom zero-copy interfaces and a single-threaded execution loop using Viewstamped Replication for deterministic performance. The article presents these architectural choices as a model for applying low-level performance engineering to high-throughput systems.

Waterfox 6.7.0 – Supernova

Waterfox 6.7 upgrades to Gecko ESR 153, bringing a year of engine, performance, and security improvements, along with a new Nova browser style and twelve color palettes. Tree-style tabs are now natively integrated into the vertical tab strip, replacing the bundled extension. The built-in ad blocker gains a dedicated settings panel with detailed counts, per-site status, and improved anti-adblock injection.

Kino: A high-performance Ractor web server for Ruby 4.0

Kino is a high-performance web server for Ruby 4.0+ that combines a Rust front-end with parallel Ractors to run Rack 3 apps in a single process, avoiding the memory overhead of Puma-style forking. It reports being 1.5–2× faster than Puma on I/O-light endpoints, using about 7× less memory in Ractor mode, and includes production features like draining, crash supervision, and timeouts. Since Ractors remain experimental in Ruby 4.0, Kino also offers a threaded fallback mode for Rails and other non-Ractor applications.

Small, native web tricks worth remembering

The article presents a collection of small, native web platform tricks, each paired with a simple example and a relevant caveat. It advises readers to check feature support labels, keep fallbacks in place, and test with real browsers and assistive technology, especially for limited or experimental features.

Emacs 31.1 will release on 8/24

GNU Emacs traces its origins to 1976, when EMACS began as a set of macros for the TECO editor. Development of GNU Emacs as a fresh implementation with a full-featured Lisp core started in 1984, and the initial public release (version 13) came in March 1985. The article provides a list of subsequent stable releases with their approximate dates.

We Rebuilt the Linux MicroVM Stack on Apple Silicon

Encore built “crackling,” a microVM API that runs Firecracker on Linux and Apple’s hypervisor on macOS, letting engineers run the same build system locally on their laptops. To do this, they rebuilt the Linux image toolchain for macOS and created custom tooling to convert Docker layers into a bootable Firecracker block device. Production continued using Firecracker on Linux unchanged.

Stop Making TUIs

The author argues that terminal and command-line interfaces are outdated, advocating for native UI development with AI assistance. They describe building several native apps—like an AI-powered music player and a self-driving wiki—using AI tools, which made the process fast and enjoyable despite the complexity of UI coding. The article suggests that AI enables anyone to create useful native software, blurring the line between development and configuration.

ParqDB – Vector search in the browser from Parquet over HTTP"

ParqDB performs semantic search over 100,000 Wikipedia articles entirely in the browser by storing an IVF-LVQ8 index of MiniLM embeddings as Parquet files in object storage. It uses HTTP range requests to fetch only needed byte ranges, embedding the query locally and ranking results with WASM—eliminating any query server. As a result, user queries never leave the browser.

🔧 Hardware & Infrastructure

Musk fumbles the timetable for first Starship catch

Elon Musk has revised his previously optimistic timeline for catching the first Starship booster, walking back comments made during an earnings call as the next orbital test approaches. The updated plan suggests a more cautious schedule for the ambitious recovery maneuver.

Homebrew 68K Machine Has A PCI Bus

A homebrew computer by [maniek86] uses a 10 MHz Motorola 68000 with 1 MB SRAM and 4 KB ROM, plus a Xilinx Spartan II FPGA as a PCI bridge to provide two 32-bit PCI slots. The machine can interface with PCI cards like a serial expansion and a Cirrus VGA adapter, and the article explains the challenges of implementing PCI on such a system.

Better Batteries

The debate over standard library size is misguided; what matters is design quality and adaptability. Python’s stdlib has uneven quality but enables rapid innovation, Go’s is well-crafted with strong institutional support, while Rust’s core APIs are excellent but organizational limits hinder adding new features like an OS random bytes API.

Anthropic hires Amir Salek, who ran Google’s TPU until 2022, for compute team in own-chip push

Anthropic has hired Amir Salek, a founder of Google’s custom chip program who led its tensor processing unit business until 2022, to join its compute team. Salek will report to James Bradbury as the AI lab works toward developing its own semiconductors.

Kobo can run apps now

Cobalt is an open-source platform that brings apps to Kobo e-readers, featuring a launcher, signed App Store, Rust SDK, and a runtime that runs each app as an isolated, unprivileged process. After a one-time USB setup, apps install, update, and remove over Wi-Fi with verified signatures, while a reboot restores the stock Kobo interface. The platform avoids altering the boot chain and is currently tested only on the Clara BW device.

What happens when a GPU reads memory

The article traces a single global memory load (LDG.E) from a CUDA kernel through an RTX 4090, following the instruction from register reads and scheduler issue through the load/store unit, coalescer, L1 cache sectors, and beyond. It explains that these hardware details are largely undocumented by NVIDIA and were determined through timing experiments. The goal is to provide performance insights by reverse-engineering the GPU memory path.

I ran Photoshop on a £0.60 computer chip

The author emulates a 1980s Apple Macintosh on a 60p Raspberry Pi RP2350 chip, noting that a mid-1990s-equivalent computer can now be built for a few pounds, versus £9,554 in inflation-adjusted 1989 costs. This leads them to consider a minimal, low-power modern computer, questioning what features could be sacrificed for simplicity, battery life, and reduced e-waste.

Radiation damage to Hubble has been 4.3 years out of phase with the Solar cycle

Hubble Space Telescope’s CCD detectors, monitored for over 24 years, show radiation damage rates that vary with the solar cycle but are several years out of phase with sunspot activity. Fits based on models from other missions accurately correct over 99.5% of the damage’s effect on image quality, though they require physically implausible parameters. The findings underscore the complexity of low Earth orbit’s radiation environment and the need for continued monitoring to predict spacecraft degradation.

BigLake’s Iceberg REST Catalog

Google’s BigLake now offers an Apache Iceberg REST Catalog endpoint, giving Iceberg-compatible engines a standardized way to discover and access tables. Previously, BigQuery relied on manual metadata file pointers or catalog integrations like AWS Glue, which required updates or custom setups. With this change, engines can use the standard REST protocol to read and write Iceberg tables, and OLake uses this catalog to write tables readable by BigQuery.

Better Batteries

The article argues that the standard library debate should focus on design quality and institutional capacity, not size. It contrasts Python’s uneven but innovative stdlib with Go’s well-designed extensions, and notes Rust’s excellent initial APIs but limited capacity to add new ones, such as a random bytes API.

Framework responds to complaints that BIOS update bricks Ryzen 7040 laptops

Framework’s BIOS 3.20 update for the Ryzen 7040-based Framework Laptop 13 is bricking some systems, with users reporting stalled installations and unresponsive devices. Framework acknowledged that a small percentage of updates result in non-bootable boards and said it is investigating the root cause.

🌍 Society & Policy

Japan tried to build an operating system for the entire world, then the US government intervened

TRON was a 1984 Japanese project to build a unified computing architecture, including a custom CPU, hypermedia OS, and a vast character encoding system, but its desktop variant BTRON was effectively killed after a 1989 US trade report labeled it a trade barrier. Its embedded counterpart ITRON quietly became one of the most widely deployed operating systems ever, while the project later attracted baseless conspiracy theories.

Apple paid Ireland $17B in 2025, 40% of its $43B global tax, after EU court ordered €13B back taxes

Apple paid $17 billion in taxes to Ireland in 2025, representing 40% of its $43 billion global corporate income tax total. The payment followed an EU court order requiring the company to pay €13 billion in back taxes.

London cabbies push back as Waymo, Wayve target city for robotaxis

London cab drivers are pushing back against the planned arrival of autonomous taxis, with some reportedly considering “skulduggery” to demonstrate the vehicles’ limitations. Their concerns come as companies like Waymo and Wayve compete to make London their European launchpad for driverless services.

China grants Geely 2-year permit for first private satellite IoT trial for transport, energy

China’s MIIT has granted Geely’s commercial satellite unit, Geespace, a two-year permit for the country’s first private satellite IoT commercial trial. The approval allows Geespace to offer low-power, wide-coverage connectivity for sectors such as transport, marine fisheries, energy, and water infrastructure, highlighting Beijing’s push to accelerate its commercial space industry.

Memory boom reshapes Micron’s Boise: millionaires, jobs, housing costs, wealth gap

Micron’s AI-driven stock surge has turned many Boise employees into millionaires and fueled a local boom, with new chip plants and research labs expected to create thousands of jobs. However, rapid growth has also brought traffic, rising housing costs, and concerns about uneven wealth distribution and overreliance on a historically cyclical industry.

Nevada OKs Tesla up to 5,000 robotaxis in Las Vegas; Waymo, Uber 1,000 each

Nevada approved permits allowing Tesla, Uber, and Waymo to operate robotaxis in Clark County, with Tesla authorized for up to 5,000 vehicles and Uber and Waymo for 1,000 each over the next year. Tesla indicated it will likely deploy far fewer, estimating around 2,500 vehicles. The approvals faced opposition from taxi operators concerned about oversaturation and impacts on the workforce.

Scientists release biggest 2D map of the universe

Scientists have released the largest 2D map of the universe, accessible via the Legacy Survey Sky Viewer at https://viewer.legacysurvey.org.

Death of an Engineer, Life of a Farmer

A former systems engineer who experienced burnout, worsening eczema, and guilt over missing his young children’s lives decided to leave his corporate career. He moved to the countryside to farm land once worked by his grandfather, learning to live by agricultural rhythms. After the career change, his severe eczema improved dramatically.

Thousands of Soldiers Now Roam Washington’s Streets at $1.6M per Day

President Trump’s National Guard deployment to Washington, D.C., has had minimal impact on crime, with soldiers appearing in only about 1.3% of criminal cases filed in Superior Court, often as victims rather than arresting officers. The roughly 4,500 troops, who cost about $1.65 million per day, mostly patrol wealthier, whiter, and less dangerous areas, handling minor incidents rather than the neighborhoods where most murders occur.

Turkey issued an international arrest warrant for Benjamin Netanyahu

Turkey issued an international arrest warrant for Israeli Prime Minister Benjamin Netanyahu over Israel’s interception of a Gaza aid flotilla and detention of activists. Netanyahu’s office responded by calling Turkish President Recep Tayyip Erdogan an “anti-Semitic dictator.” The article also notes that over a dozen Muslim nations condemned Israeli strikes on Syria and settlement plans in the West Bank.

The US techlash is real

The article’s content is not available in the provided text, which only displays subscription offers. Therefore, no summary of the article can be generated from this material.

A study published in JAMA Internal Medicine found elevated cancer-related mortality risks among US pilots and flight attendants. The researchers identified higher death rates from certain cancers, including breast, melanoma, and prostate, linking them to occupational exposures like cosmic radiation and disrupted circadian rhythms.

Felony Bench

Felony Bench is a benchmark that counts unique instances where AI agents engage in illegal activities affecting third-party entities, excluding sandbox escapes. It lists incidents from companies such as Anthropic, OpenAI, and Meta, with scores indicating the number of felonies committed. The methodology specifies that only actions with third-party impact are counted.

US Navy carrier honouring Black war hero to be renamed after Trump

The Trump administration reportedly plans to rename a Navy aircraft carrier originally designated the USS Doris Miller, in honor of a Black Pearl Harbor hero, to instead bear Trump’s name. The move has drawn sharp criticism from the NAACP and others, who call it a disrespectful “slap in the face” to veterans and service members. Navy and Defense Department officials declined to comment.

WPD won’t replace stolen Flock cameras, citing public trust

Following public outcry, Winona will not replace its stolen Flock surveillance cameras, prioritizing community trust over their effectiveness in solving crimes. The automated license plate readers had been vandalized and sparked criticism over privacy and potential misuse. Despite the cameras’ usefulness, police decided against reinstalling them.

What we lost when search stopped making us think

Search results have degraded due to SEO filler and AI-generated content that prioritize ranking over accuracy, prompting search engines to offer AI summaries that often bypass original sources. This makes it harder for users to verify information and erodes critical research skills, as people increasingly accept generated answers instead of comparing sources.

Olympian Jenny Simpson Says Her Running Career Is Over After Cardiac Arrest

Jenny Simpson, a three-time U.S. Olympian, announced that a cardiac arrest in June and a subsequent diagnosis of the rare heart condition ARVC have ended her running career. After a severe recovery that included life support, she is back home and will continue her work with Fleet Feet to stay involved in the running community.

US debt has hit $40tn - Will that be a wake-up call?

The U.S. national debt has passed $40 trillion, doubling in a decade amid spending surges and tax cuts, with interest payments now consuming nearly 20% of tax revenue. Rising interest rates and investor wariness are making debt financing more expensive, though economists describe it as a “flashing yellow light” rather than an imminent crisis. If debt keeps climbing, it could eventually trigger a bond selloff and financial market turmoil.

New Worlds: We are living in the future of J.G. Ballard or William Gibson

We already live in a cyberpunk-like future straight out of J.G. Ballard or William Gibson, filled with IV drip clinics in malls, robotic dogs, AI-generated music, narco subs, and humanoid robots. Yet these once-startling developments are now greeted with indifference, and that lack of shock is itself the most surprising part.

Grand jury declines to indict Ohio man charged with destroying Flock camera

A grand jury in Ohio declined to indict Cody Morelock on felony vandalism charges for destroying a Flock license plate reader camera, leading to the dismissal of the charges. The case is part of a broader backlash against Flock cameras, which have been vandalized nationwide amid concerns over police misuse of the surveillance technology.

Nothing Doing

The article highlights humorous plaques and signs from various locations that ironically celebrate their lack of historical significance, such as noting that no important person was born or died there and that nothing significant has happened for centuries. These markers intentionally contrast with typical historical plaques by honoring the ordinary and uneventful.

The Prague of Deus Ex: Mankind Divided

The article examines how Deus Ex: Mankind Divided portrays Prague as a city of disjunctions, balancing medieval history with garish tourism and high-tech, brutalist architecture. It argues the game honestly depicts Prague as a compromised, split-history city, including run-down outskirts with hidden sanctuaries.

The Religious Experience of Philip K. Dick by R. Crumb (1986)

The post shares R. Crumb’s comic “The Religious Experience of Philip K. Dick” from Weirdo #17, offering a CBR download link and correcting a missing page after a reader’s note. The accompanying comments debate whether Dick’s visions were genuine, schizophrenic, or something else, with some dismissing him as unstable and others defending his experiences.

Japan tried to build an operating system for the world, the US intervened

TRON was a 1984 Japanese government-backed project to build a complete, vertically integrated operating system family, including the advanced BTRON desktop OS, but it was effectively killed after a 1989 US trade report labeled it a trade barrier. Its embedded variant, ITRON, quietly became one of the most widely deployed operating systems in history. The project later attracted baseless conspiracy theories, including a false claim that a plane crash targeted TRON developers.

AI companies destroy physical books – let’s scan rare books before it’s too late

AI companies are secretly buying, scanning, and destroying millions of physical books to gain exclusive training data, locking knowledge in private servers. Anna’s Archive is urging volunteers worldwide to scan and upload books, especially rare ones, before this cultural heritage is lost. The article warns that this destruction, along with rising AI-generated content, threatens public access to human knowledge.

Artificial Intelligence Policy

A law school policy effective Summer 2026 prohibits students from using artificial intelligence to conceptualize, outline, draft, revise, translate, or edit any work submitted for credit, and bans AI use for any purpose in exams. Instructors may deviate from the default rule with written notice. The policy aims to ensure students develop the critical thinking, legal reasoning, and drafting skills essential for competent lawyering.

It is a sign of the times that Amazon gets to call this fair use

Amazon has been buying rare books, cutting off their bindings, and scanning them for AI training data, destroying the physical books in the process, according to a 404 Media investigation that tracked a purchased book to an Amazon warehouse in Las Vegas. This destructive approach contrasts with the Internet Archive, which digitizes books page-by-page without damaging them.

Copyright does not protect AI-generated content in EU

The EU has ruled that copyright protection does not extend to AI-generated content. As a result, works created entirely by artificial intelligence may not be legally protected under copyright law. This decision has significant implications for creators and industries using AI tools.