Alvaro Lopez Ortega / 2026-08-27 Briefing

Created Thu, 27 Aug 2026 19:11:47 +0000 Modified Thu, 27 Aug 2026 19:14:13 +0000
7817 Words

US prosecutors are investigating a major cyberattack on the ATF after a ransomware gang claimed responsibility. Manchester Airports Group said cybercriminals stole data affecting up to 8.7 million customers. Microsoft rebranded its 365 roadmap as “AI at Work,” while Salesforce credited AI-fueled Flex Credit consumption for half its bookings. UK teachers’ AI use has doubled but working hours remain unchanged. A call-blocker seller was fined £190,000 for making nuisance calls.

🤖 AI & Machine Learning

Anthropic’s Model Hardware Standard lets AI agents control microscopes, quantum hardware, robot arms

Anthropic introduced the Model Hardware Standard, a framework of rules specifying how AI agents should safely interact with physical systems such as microscopes, quantum computing hardware, and robot arms. The company aims to accelerate scientific research and manufacturing while managing risks, and will work with trusted partners on safety before making it widely available.

Meta memo: Hatch AI has its own computer, works when app closed, connects to email, IG, OpenTable

Meta is developing an internal AI agent called “Project Hatch” that can perform tasks such as ordering food, booking restaurant tables, and accessing apps like Instagram, Spotify, and OpenTable, all while working even when the app is closed. The agent is described as having its own computer to complete online tasks, and Meta may launch it within weeks at a potential price of up to $200 per month.

OpenAI tests Codex ‘Persistent mode’: AI agents work until put to sleep, generate follow-up tasks.

OpenAI is testing a new “Persistent mode” for its Codex AI agent that would allow it to keep working until “put to sleep” and proactively create follow-up tasks for itself. The feature, spotted in Codex’s public codebase and confirmed by OpenAI, is designed to make agents more autonomous, though the company says there are no immediate plans to launch it.

Google launches Gemini Omni 1.1 Flash: studio-quality video, scene extension, 4K upscaling, more

Gemini Omni 1.1 Flash introduces new generative video capabilities for developers, including scene extension, keyframe control, faster 360p previews, and upscaling to 4K. It now supports up to 10 seconds of prior context for consistent video continuation and up to three seconds of video reference input. The model is production-ready via the Gemini API in Google AI Studio.

Google pilots double-blind AI evals using crypto ‘box’ to curb benchmark leakage and protect IP

Google has launched the world’s first double-blind evaluation of a proprietary frontier AI model, using cryptographic safeguards to keep external test prompts hidden from the model and protect the model’s weights from evaluators. This approach aims to prevent benchmark contamination and ensure evaluation results accurately reflect model capabilities. The pilot tests a Gemini Flash Lite model with partners including the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons.

First fully AI-guided Russian drone kills 3 people in Ukraine

A Russian drone attack on a petrol station in Zaporizhzhia, Ukraine, killed three civilians and marked the first use of a fully autonomous Russian drone that independently selected its target. Investigators found an Nvidia Jetson Orin module in the wreckage, though Nvidia says Russia cannot buy such components directly. Experts warn the strike signals a future where autonomous weapons may kill civilians without direct human control.

Man, AI is killing my brain

A programmer describes how workplace pressure to match AI-accelerated colleagues led him to increasingly rely on AI coding tools, eventually losing his ability to understand code, research, or even write without assistance. He now feels his thinking abilities have eroded and fears for his career, as interviews also demand heavy AI usage. The piece is a personal lament about the cognitive toll of AI dependence.

Gemini Omni 1.1 Flash

Gemini Omni 1.1 Flash introduces new generative video controls for developers via the Gemini API, including scene extension up to 40 seconds, first/last frame specification, faster 360p previews, upscaling to 4K, and video references for consistency. The updates aim to make generative video more controllable, iterative, and production-ready.

My Claude quota ran out in 10 minutes, so I made a tool to find out why

Tare is a tool that installs into Claude Code and lets users ask in plain English about their token usage, such as why they hit limits or where their quota went, by reading Claude Code’s own logs. It requires no dashboards or commands, just a one-time install, and provides answers with causes rather than raw numbers. It can also generate reports, compare weeks, and suggest usage-saving changes.

Small Models Have Arrived

Small, fast, low-cost models like gpt-5.6-luna are now capable enough to make consumer AI apps viable, cutting per-task costs from roughly $1 to $0.10. While frontier models still drive novel breakthroughs, most business work is routine “token spewer” work—responsive, broad, and repetitive—creating rising demand for fast, cheap, and good-enough models.

Needle: The benchmark your search engine can’t memorize

NEEDLE is a live, open-source search benchmark designed to combat overfitting and data leakage that plague static benchmarks like BrowseComp. It uses queries from real agent search logs and production-inspired intents, running continuously in public so results are reproducible and contributions are welcome. The article argues AI agents increasingly need fresh or niche search information, but current engines remain inadequate for agentic traffic.

The Teaser Period: Why the AI Boom Is Hitting a Reset Wall

The article compares the 2008 financial crisis—caused by a scheduled wave of subprime mortgage rate resets—to today’s AI boom, arguing that massive take-or-pay compute contracts contain a similar “reset wall” of delayed payment obligations. These contractually inevitable payment shocks, triggered after data centers are built, are currently underpriced by the market.

Six months of writing code exclusively with agents

After six months of using AI agents to write all code, the author stopped hand-typing code in February, focusing instead on diagnosing and fixing agent shortcomings. This approach freed up time to run multiple agents concurrently, but introduced coordination problems such as file conflicts and resource contention.

MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training

MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training released a report calling for action on generative AI’s role in education. It found AI is already pervasive at MIT, with mixed student and instructor views, and is disrupting core aspects of the undergraduate experience, including assessment, collaboration, and mastery. The report proposes eight principles and recommends immediate and long-term actions for instructors and administration to guide MIT’s response.

Pollen Robotics (Hugging Face) Microduck

Microduck is a 25 cm open-source biped robot trained with reinforcement learning in simulation and deployable to the real robot, featuring seven pre-trained behaviors such as walking, sitting, kicking, and roller skating. It is fully hackable, allowing users to retrain policies on their own machines, publish them, and access the entire software stack on GitHub. Pre-orders are open in four colorways, shipping before Christmas 2026 at an introductory price of $399.

Getting video models to learn better, faster

Recent gains in image and video models stem primarily from data improvements—filtering and rebalancing, better annotations, and synthetic data—rather than architectural changes. The article details their 2024 budget-friendly approach using CPU-based traditional computer vision for scene detection, along with manual data review, to filter and rebalance massive datasets.

Laion Big Video Dataset

LAION-BVD is a large-scale open video dataset containing 1.3 billion video URLs and 80 million downloaded videos totaling 10 million hours, with synthetic captions for multimodal pre-training. Models trained on it achieve competitive performance on video-text, audio-text, and image-text benchmarks, improving with scale. The dataset is released exclusively for research purposes to broaden access to multimodal training data.

CEO fired developers to make room for AI. Developers create open source AI CEO

Open Executive is an open-source AI system that acts as a virtual executive team, providing a unified “executive voice” backed by eight specialist agents (e.g., CFO, CMO, General Counsel) with Harvard MBA-level knowledge. It uses RAG to combine built-in business knowledge with company documents, maintains episodic memory of past decisions in SQLite, and includes a scheduler for proactive follow-ups. Built with a Python/FastAPI backend, ChromaDB, and a Next.js frontend, it offers a consistent interface without exposing the underlying agent architecture.

🔒 Security & Privacy

ATF responds to ‘major’ cybersecurity incident after ransomware gang’s claims

The ATF has responded to a “major” cybersecurity incident following claims made by a ransomware gang. The US Justice Department is investigating the breach.

Cybercrooks jet off with Manchester Airports Group customer data

Manchester Airports Group, the UK’s largest airport operator, was hit by cybercriminals who stole customer data. The company believes approximately 8.7 million customers may be affected.

Nuisance-call blocker fined £190k for being a nuisance caller

Elderly Aids, a company selling devices meant to block nuisance calls, was fined £190,000 for itself making 758,000 unwanted calls per year. The firm was found to be bombarding people with these calls while marketing its call-blocking products.

FBI seizes China-linked hacking tools used against NASA, DOE, US Senate, other critical networks

The FBI seized hacking tools allegedly used by Chinese state-sponsored actors to target critical networks, including NASA, the Department of Energy, and the US Senate. The operation disrupted botnets linked to Beijing’s cyber activities.

OpenAI, Anthropic, AWS, Microsoft, 100+ firms warn: limited window for AI attacks, act collectively

Major AI and tech companies, including OpenAI, Anthropic, AWS, and Microsoft, along with over 100 other firms, are warning that there is only a limited window to prepare for AI-enabled cyberattacks. They are calling for collective action to address this emerging threat.

Meta’s AG settlement: age verification to use reliable signals from Apple and Google app stores

Meta’s settlement with state attorneys general over kids’ safety includes an age-verification framework that relies on “reliable age signals” from Apple and Google’s operating systems and app stores. The deal is also structured to pressure rivals like TikTok, YouTube, and Snap into adopting similar protections, with part of Meta’s payment contingent on their compliance. Meta has publicly framed the changes as an emerging industry standard while using the settlement to position itself against competitors.

Australian police arrest two men, 21 and 23, for alleged TeamPCP cybercrime group membership

Australian Federal Police arrested two men, aged 21 and 23, from Western Australia for allegedly being members of TeamPCP, a cybercrime group blamed for a prolonged series of open-source software supply chain attacks that extorted thousands of businesses worldwide. The group embedded malicious code in code libraries, used a self-propagating worm called Shai-Hulud, and even ran a recruitment contest to expand operations.

Meta: new teen safeguards on Facebook, Instagram are US-only; foreign officials will watch closely

Meta’s new teen safety safeguards on Instagram and Facebook will initially apply only to U.S. users. The changes, part of settlements costing up to $18 billion, include time limits for younger users and restrictions on disabling safety settings. Officials outside the U.S. are expected to closely monitor the rollout.

8.7 million customers hit as Manchester Airports Group breach exposes airport WiFi sign-ups

A cyber attack on Manchester Airports Group exposed data of 8.7 million customers at Manchester, Stansted, and East Midlands airports, including WiFi registrations, email addresses, phone numbers, and parking and booking details. Payment and bank information was not compromised, but experts warn the stolen data could be used for convincing phishing scams and fraudulent travel-related messages.

The Server Called Paranoia: Defend Autistici/Inventati Before September 25

Autistici/Inventati, a privacy-focused tech collective, has been designated as a global terrorist organization by the US administration. The group is reposting a Sabot Media article that connects the designation to its history. They thanked Sabot Media for the summary and said they will use the saved time to fight the designation.

Flock CEO gets his house blurred out on Google Maps and Zillow

XCancel has been discontinued after its operators received a cease-and-desist letter from X Corp. on Monday. The service is paused until further notice, with the team seeking legal advice and declining to share additional details.

Chinese Implants in the Supply Chain

Security researchers discovered two new implants, SPEAKINGSTONE and DARKLANTERN, in a ZBT router purchased from a US supplier. DARKLANTERN is a critical unauthenticated backdoor that listens on UDP port 9992 and allows arbitrary command execution from the internet. Both implants are written in Nim, communicate over UDP, and are launched by the same binary.

Apple threat notifications and spyware: what everyone should know

Apple’s threat notifications alert users that they may have been targeted by mercenary spyware, but do not confirm whether an attack succeeded or identify the attacker. Recipients should seek expert digital forensic help to investigate and preserve evidence, though finding no traces does not guarantee safety since sophisticated spyware can leave few signs.

US Government designates host of noblogs.org a “global terrorist”

The US State Department designated Italian collective Autistici/Inventati, which runs the noblogs.org platform, as a “Specially Designated Global Terrorist” on August 26, 2026, alongside Palestine Action and Masar Badil. The designation cites far-left politics and alleges the collective supports violent extremist groups, but the article highlights apparent errors and false conflations, including confusing Rose City Antifa with another project. Critics view the move as part of a broader effort to criminalize anti-fascists and domestic protesters.

Xcancel and Nitter have been taken down

Nitter.net and Xcancel.com have been taken down following a cease and desist letter from X Corp. Both services are no longer operational as a result.

PayPal Blocks GrapheneOS

PayPal’s app now refuses to run on GrapheneOS, crashing with a root detection security exception. The report notes uncertainty about whether enabling contactless NFC payments triggered the block.

🛠️ Software & Open Source

Nitter might be no more, but Twitter is back (sort of)

Nitter, the alternative Twitter front-end, may be shutting down, but a new service called Twitter.now aims to fill that role. Unlike Nitter, Twitter.now will require a fee to use.

Microsoft slaps a fresh coat of AI paint on the Microsoft 365 Roadmap

Microsoft has renamed its Microsoft 365 Roadmap to the “AI at Work” roadmap, signaling a shift toward highlighting upcoming AI capabilities in its productivity suite. The rebrand underscores the company’s focus on AI-driven features for workplace users.

AROS, the FOSS recreation of AmigaOS, comes to Raspberry Pi

AROS, the open-source reimplementation of AmigaOS, now has nightly builds that run natively on Arm-based Raspberry Pi hardware, including the Pi 3 and older ARMv6 devices. The initial build boots to the Workbench desktop in about five seconds, and the narrow Pi hardware range makes it easier for the small OS to support fully.

Check out the width of that Debian display

Debian is polling its developers with an eight-way ballot on how to handle AI-generated code, ranging from banning bots to letting them loose. The vote highlights the distribution’s ongoing internal debate over the role of AI in its open-source ecosystem.

Google sets Android app memory limits, citing AI hardware constraints

Google is introducing new Android app performance thresholds, including memory-use and code optimization limits, due to industry-wide memory chip shortages worsened by the AI data center boom. Developers have until February 2027 to comply, and Play Store apps must also support automatic sign-in restoration across devices by April 2027.

Rockstar execs talk GTA VI: player habits like diet affect character looks and in-game social media

Rockstar executives say Grand Theft Auto VI faces “unprecedented pressure” as the follow-up to GTA V, though they emphasize their own internal expectations are even higher. The game will feature interactive systems where player habits, such as diet, affect characters’ appearances, alongside in-game social media elements.

Float Bloat: vector serialization gone wrong

“Float Bloat” is a widespread issue in vector search where clients cast and serialize float32 embeddings as float64, doubling storage and network costs without adding accuracy. This stems from default serialization behavior in popular languages and appears across SDKs, documentation, and clusters, with an estimated global overhead of over 20 petabytes of unnecessary disk storage.

UNIX V4 workshop at Low Resource Computing

At a Low Resource Computing 2026 workshop, the author led a hands-on session using UNIX V4, recovered from a 1974 magnetic tape, teaching software development with the ed editor on a teletype. They modified the kernel to support 32 simultaneous terminal connections, and participants logged in via telnet and physical terminals. The repository includes the disk image, terminal logs, and configuration notes.

Symmetries in boolean scans

APL relies on optimized primitive operations like scans and folds, which execute much faster than user-defined equivalents. A key example is using a not-equal scan (≠\) to efficiently track which positions in a string are inside quotes, replacing an explicit loop.

png2jxl: Convert PNG to lossless JPEG XL with byte-for-byte reconstruction of the original PNG

png2jxl converts supported PNG images to pixel-lossless JPEG XL files that can reconstruct the original PNG byte-for-byte. It offers a Python API, command-line interface, and Pillow integration, with automatic verification and options like only_if_smaller. The tool handles common 8-bit PNG formats and uses preflate-rs to preserve the original DEFLATE stream exactly.

Please stop flooding our projects with AI slop to furnish your CV

Open source maintainers are seeing a rise in AI-generated contributions, such as trivial pull requests and low-severity vulnerability reports, that appear aimed at gaming GitHub metrics for career gain. These maintainers are responding by closing pointless PRs and becoming more selective, warning that such activity erodes the trust that open source relies on.

tailcat: like netcat, but over Tailscale’s data plane, without Tailscale’s control plane

Tailcat is an open-source tool that combines Tailscale’s encrypted data plane with netcat-like functionality, using out-of-band connection metadata instead of Tailscale’s control plane. It supports piping stdin/stdout, exposing local ports, running auth-free SSH, and testing connectivity via ping, all secured with end-to-end WireGuard encryption. The project includes a CLI, a Go library, and an experimental browser-based WebAssembly demo.

Announcing our first Maintainers in Residence

The Rust Project announced its first round of funded maintainers, including Gen Li, Chris Denton, Alejandra González, and León Liehr as Maintainers in Residence, with Jason Newcomb and Jonas Böttiger receiving Maintainer Grants. The funding supports their rust-lang maintenance work for at least 12 months and comes from the Rust Foundation Maintainers Fund, backed by Google, AWS, OpenAI, and other donors.

Changes to SourceHut’s terms of service regarding LLMs

SourceHut is updating its terms of service to restrict or prohibit the use of LLMs and generative AI on its platform, following community discussion and internal review. The decision is driven by concerns about excessive resource usage, outages caused by AI crawlers, and ethical objections tied to how AI exploits open source software. The new policy will include enforcement examples, a timeline, and options for affected users.

We found a division by zero bug in FFmpeg with a vibecoded fuzzer

Anubis is a Proof-of-Work system that protects websites from AI scrapers by adding a computational cost that is negligible for individual users but expensive for mass scraping. It is a temporary solution until better headless browser fingerprinting is developed, and it requires modern JavaScript features.

Grok Bot for Linux: Unofficial port of the official app (open source)

An unofficial Linux port of the Grok Bot desktop app repackages the official Windows Electron app for Linux, replacing native addons. It provides the full Grok Bot UI and Cursor account sign-in, but is community-supported and not endorsed by xAI or Cursor. Installation is via building from source or using provided .deb, AppImage, or tarball packages for x86_64 Linux.

Dwarf Fortress is getting the mother of all magic updates

Dwarf Fortress is getting a Myth and Magic update that introduces procedural magic systems tied to each world’s fundamental cosmology, enabling magical research, materials, enchantments, rituals, and spells that feel like discovery. The update goes beyond combining spell effects to create unique magical situations, and is scheduled for later this year.

Salem Robotics (YC S26) – Software for industrial inspection robots

Salem Robotics, a YC S26 startup founded by researchers with nuclear industry experience, develops software that gives existing mobile robots task-specific intelligence for inspections in hazardous industrial facilities. The company combines AI with classical robotics to handle precise physical tasks like nuclear contamination surveys and leak detection, focusing on joint-level control and verifying the inspection results themselves. They sell directly to industrial facilities, starting with paid technical validations and then moving to ongoing deployments.

Decompiling a Nintendo 64 game in 84 days

The article details using AI agents to decompile a Nintendo 64 game, finding they excelled at matching standard library code but required strong prompting to reuse existing sources instead of rewriting them. Key efficiency gains came from automated exact-match scripts, a shared learnings file for compiler quirks, the N64 Decomp Workbench for debugging mismatches, and explicit deadlines to manage agent timeouts.

Software engineering is about managing complexity

Writing code is not the same as building software; the real challenge lies in deciding which components should exist, how they interact, and which tradeoffs are acceptable. AI excels at generating code, but software engineering requires selecting an architecture suited to specific constraints like team size, user scale, and business needs. Ultimately, AI offers solutions, but engineers must own the context-driven decisions and tradeoffs.

Restoredrill – proves your Postgres backups restore

restoredrill is a CI-native tool that verifies PostgreSQL backups by restoring them into a throwaway container, running user-defined checks, and generating a timestamped JSON report for audit evidence. It is Postgres-only and early-stage, designed to make scheduled restore testing loud and fail-closed, with features like RPO freshness checks and SQL assertions. Unlike full backup platforms, it focuses solely on producing auditor-ready proof that restores actually work.

Harness Engineering

Harness engineering uses deterministic tooling, agent-based review, and periodic entropy checks to keep AI-generated code correct and coherent over time. The concept originated from Birgitta Boeckeler’s analogy to test harnesses, with context engineering and architectural constraints as key components.

CoMaps integration with the wider FLOSS ecosystem

CoMaps, a fully free and open-source app, now integrates with other FLOSS apps, including Gadgetbridge for turn-by-turn navigation on smartwatches and OpenVitals for recording activities with navigation context. These integrations are privacy-respecting and require users to grant special permissions for data sharing.

Emacs 31: An unofficial guide to Markdown-ts-mode

Emacs 31 introduces a new experimental but feature-rich markdown-ts-mode that supports CommonMark, most GFM specs, and extra tools. It is built-in, so it must be enabled manually through configuration rather than installed from MELPA. When opening Markdown files for the first time, you may need to install the Markdown tree-sitter grammar.

RealDiff – runtime behavior diffing for pull requests (six languages)

RealDiff is a .NET tool that detects runtime behavior changes in pull requests by running the same test suite on both branches, recording method arguments and return values, and diffing the execution traces. It catches behavioral differences that source review misses, such as changes in sort stability, and reduces false positives by subtracting run-to-run variation and reporting the origin of the change rather than its downstream effects.

“No way to prevent this” say users of only language where this regularly happens

A vulnerability in GNU gzip (CVE-2026-41992) caused out-of-bounds reads in its LZH decoder due to a shared global buffer not being reset, prompting urgent patching. The article satirically notes this occurs because gzip is written in C, the language where memory safety issues regularly happen, yet programmers claim they are helpless to prevent them.

Changes to Sourcehut’s terms of service regarding LLMs

SourceHut announced it will update its terms of service to restrict or prohibit the use of large language models and generative AI on its platform, following community discussion and internal review. The decision is driven by practical concerns such as excessive resource usage and outages from AI-generated projects and crawlers, as well as ethical objections that LLMs exploit open-source software without respecting licenses and conflict with the organization’s mission.

Pnpm 12.0

pnpm 12 is a stable Rust rewrite that preserves pnpm 11’s commands, flags, settings, and lockfile format, while introducing a few deliberate behavior changes. These include canonical resolution of git dependencies, reporting unrecognized workspace settings, deterministic lockfiles for cyclic dependency graphs, and hardlink-first package imports on Linux. It is installed from the next-12 tag, as npm’s latest still points to pnpm 11.

Technical Writing Style Guide

The article outlines guidelines for technical writing, emphasizing clarity, active voice, and US English. It advises placing important information first, using short sentences, and employing transitions to connect ideas. Procedures should be written with numbered steps, imperative verbs, and parallel structure.

🖥️ Hardware & Infrastructure

Schrödinger’s backup: not actually recovered until you try to restore IT

Backups can appear complete yet remain unverified, so organizations must actively test restoration to ensure data can be recovered during a ransomware attack or outage. A survey found 53% of IT professionals were only somewhat confident in their recovery ability, with limited testing. Verification is essential to prove backups are complete, uncorrupted, and fully restorable before a real disaster occurs.

HP has a solution to expensive AI tokens: Buy a more expensive PC

HP suggests that as cloud AI token costs rise, users should consider running workloads on pricier, higher-end PCs. This shift benefits HP by improving its margins on more expensive hardware.

You could own AI.slop, if applicant for new global top-level domains get their way

Applicants for new global top-level domains have proposed quirky options including .slop, .borg, .itiswhatitis, .therapy, .hype, and .tea. If approved, these could become available for domain registration.

Nutanix built $20m AI cluster to reduce use of Copilot and Claude, expects ROI in a year

Nutanix has built a $20 million AI cluster to reduce its reliance on external AI tools like Copilot and Claude, expecting to recoup the investment within a year. The company also sees Arm architecture as a potential advantage for edge computing amid memory shortages.

Samsung’s Galaxy S26 FE: Exynos chip, 12GB RAM, reduced cameras, $700+ (up $50)

Samsung unveiled the Galaxy S26 FE at $700, a $50 price increase over its predecessor, due to industry-wide memory chip shortages. The phone features an Exynos chipset, 12GB of RAM, and a reduced camera system compared to the S26. It launches with carriers on Sept. 3 and at other retailers on Sept. 4.

Hugging Face unveils Microduck: $400, 10-inch bipedal robot with Pollen Robotics, Seeed Studio

Hugging Face unveiled Microduck, a $400 singing, skating bipedal robot inspired by a one-eyed duck, set to go on sale later this year. It is the company’s second robot, allowing developers to teach it new skills using reinforcement learning, simulation, and open-source languages like Python and JavaScript. The robot was developed in part by Pollen Robotics, which Hugging Face acquired last year.

What happens when a GPU reads memory

The article traces how a GPU global-load instruction (LDG.E) travels through an RTX 4090, from register read and warp issue through the load/store unit and coalescer to L1 cache. It details the full hardware path for a vector-add kernel, including sectors, cache lines, address translation, L2 slices, and DRAM access on misses, all reverse-engineered via timing experiments since NVIDIA documents little of this process.

Saving 100 terabytes of memory by optimizing 1.1.1.1’s DNS cache

Cloudflare reduced the per-entry memory footprint of its 1.1.1.1 DNS cache by over 50% through five storage optimizations, freeing roughly 100 terabytes of memory across its fleet. The changes also improved performance, boosting insert throughput by 43% and cutting lookup latency by 19%.

Suica, Japan’s First IC Transit Card

Suica, Japan’s first IC transit card, relies on Sony’s FeliCa contactless technology, which uses the reader’s electromagnetic field to power the card and store balance and ID locally, enabling fast transactions without a live server connection. This design prevented network delays and congestion at Tokyo’s crowded station gates, replacing older manual and magnetic ticket systems. The article details the technology’s development, initial resistance, performance challenges, and eventual success.

Launching Route 53 Files

Route 53 Files is a new AWS file system that exposes Route 53 hosted zones as standard NFS file systems, allowing DNS records to be edited with familiar UNIX tools and automatically synced bidirectionally with Route 53. It works across EC2, ECS, EKS, and Lambda, supports concurrent shared access from multiple compute resources, and presents record sets as files and record names as directories.

507 Mechanical Movements

The article highlights the 1868 book “507 Mechanical Movements,” providing a link to its digitized version on the Internet Archive. It serves as a historical reference documenting various mechanical mechanisms and movements.

Dissecting the Apple M1 GPU, the End

Open-source developers reverse-engineered Apple’s M1/M2 GPU, eventually delivering full graphics acceleration in Asahi Linux for Macs running Linux. The project culminated in conformant OpenGL 4.6, OpenGL ES 3.2, Vulkan 1.4, and OpenCL 3.0 support, enabling Proton/Direct3D games on the platform.

💼 Business & Markets

Salesforce: 50% of bookings from customers refilling the tank, consuming Flex Credits, wanting more

Salesforce reported that half of its bookings came from customers refilling their Flex Credits, signaling strong consumption of its AI offerings. The SaaS giant’s boost, partly attributed to “Claudeforce,” raises questions among users about how their AI usage will be monetized.

Anthropic plans secondary stock sales in IPO, mulls lockup over 180 days - The Information

Anthropic is considering an IPO that would allow existing shareholders to sell their stakes, a departure from SpaceX’s approach of limiting such sales. This move could provide early investors and employees with liquidity while still pursuing a public listing.

Nvidia’s supplier commitments for AI chips and systems hit $279B in Q2, up from $119B in Q1

Nvidia’s commitments to suppliers for AI chips and systems reached $279 billion in its latest quarter, more than doubling from $119 billion the prior quarter. The strategy aims to secure supply and support less-creditworthy customers, but carries financial risks despite helping the company pursue aggressive growth targets.

CivilGrid raises $26M Series A for ‘Google Maps for the underground’

CivilGrid, founded by former PG&E engineer Josh Mackanic, aggregates data on underground utility assets, property ownership, and regulations to create a “Google Maps for what’s underground,” helping prevent costly utility strikes. The company raised a $26 million Series A led by Spark Capital, and a PG&E case study identified $60 million in potential savings across 1,600 planned gas projects.

Cybersecurity stocks surge: Okta +20%, CrowdStrike +15% as AI drives attacks, spending

Cybersecurity stocks surged Thursday after CrowdStrike and Okta reported earnings that beat estimates and raised forecasts, citing AI-driven cyberattacks boosting demand for security tools. CrowdStrike rose 18% and Okta jumped 27%, with the broader sector rallying as Palo Alto Networks, SailPoint, Zscaler, and Rubrik each gained at least 10%.

Kalshi partners with Weather Co. for verification; weather trading volume to hit $1.1B by 2026

Kalshi has partnered with the Weather Co. to verify outcomes in its growing climate and weather prediction markets. The platform aims to expand into new areas like sporting events and expects its weather trading volume to reach $1.1 billion in 2026.

OpenAI launches ads on ChatGPT Free and Go tiers in India, 100M+ weekly users; ad manager next month

OpenAI will begin showing ads on its ChatGPT Free and Go subscription tiers in India, where it has over 100 million weekly active users, starting with 50 brands and partnerships with WPP and Omnicom. The company plans to launch an ad manager next month with a minimum daily campaign budget of ₹725 (~$7.60). This follows its ad rollout in the U.S. and Europe as OpenAI expands revenue streams ahead of a potential IPO.

X launches NFL Gametime hub with live scores, stats, creator content through 2027 Super Bowl.

X has launched “NFL Gametime,” a dedicated NFL feed and hub on its platform featuring real-time scores, statistics, news, and creator content, available from season kickoff through the 2027 Super Bowl. The move aims to attract both sports fans and advertisers following X’s July settlement with the World Federation of Advertisers over an alleged boycott.

Amazon now hides all customer reviews on products

Amazon is now hiding nearly all product reviews, showing only the top 5-10 and disabling filtering and ordering features. Users can opt out, but only by submitting a request via email that takes five days to process.

Nvidia projects $673B in sales as AI demand widens

Nvidia forecasts 70% revenue growth for fiscal 2028, implying about $673 billion in sales and surpassing Apple and Alphabet in projected revenue. The outlook is driven by widening AI demand beyond hyperscalers to regional providers, startups, and enterprises, though supply constraints currently cap growth. Nvidia is also expanding beyond chips into broader data-center infrastructure.

Trade (and Tariffs)

The article, titled “Trade (and Tariffs),” contains only xkcd.com’s standard boilerplate joke, listing absurd and obsolete requirements for viewing the site. It provides no actual information about trade or tariffs, instead humorously instructing readers on ad blockers and “Boat Mode.”

Pause – weekly curated 1:1 coffee matches for London professionals

A London-first physical network offers curated, in-person matches for focused 45-minute conversations, skill exchanges, and shared workspaces at independent venues. It includes weekly peer check-ins and verified, high-intent introductions every Tuesday, aiming to foster meaningful human connection.

Meta’s self-inflicted resignation-wave

Meta is offering large retainer equity grants ($400K–$1M+) to senior engineers (IC6+) who resign, especially those with offers from rivals like Anthropic or OpenAI, in an attempt to retain talent. This marks a reversal from Meta’s previous practice of not making counteroffers when engineers resigned, following earlier layoffs and forced reassignments that prompted many to seek jobs elsewhere.

Nvidia agrees to acquire Hugging Face for $13B

Nvidia has agreed to acquire Hugging Face, the popular AI model hosting platform, in a deal valued at $13 billion. The acquisition marks a major consolidation in the AI industry, giving Nvidia control over a key hub for open-source AI models and developers.

🌍 Science & Society

Want to lead Whitehall’s AI strategy? AI experience is not essential

The UK government is hiring a lead for its AI strategy, but prior AI experience is only “desirable,” not essential. The successful candidate will be responsible for shaping the skills blueprint for 550,000 civil servants.

AI use among UK teachers doubles, but working hours still don’t come down

AI use among UK teachers has doubled, but a new assessment finds it has failed to reduce their working hours. The technology has not alleviated the burden of teacher workload, earning an “F” for this purpose.

Trump officials draft EO for AI self-regulatory body, needs buy-in

A draft executive order to establish a self-regulatory organization for artificial intelligence has been circulated among some Trump administration officials. However, the proposal still requires buy-in from former President Trump and other key figures to move forward.

A Million Kakapos

As AI makes coding nearly free, open source software should prioritize adaptability and robustness over narrow optimization, much like evolution favors resilient generalists. The article contrasts fragile, hyper-specialized species like the kakapo with hardy survivors such as cockroaches and sharks, arguing that long-term thriving depends on flexibility and redundancy rather than peak performance.

Two German airport workers die of malaria after ‘mosquito arrives on plane’

Two Frankfurt Airport workers have died from malaria after six employees were infected in a rare outbreak likely caused by mosquitoes that arrived on a plane. Health officials say the risk to the public is very low because the disease is only transmitted by mosquitoes, not person to person. An investigation is underway, including analyzing captured mosquitoes and blood samples to determine the source.

Yayoi Kusama has died

Japanese avant-garde artist Yayoi Kusama, famed for her polka-dot sculptures and installations, died at a Tokyo hospital on 14 August at age 97. Known for her vibrant works that drew global crowds and fetched record prices for a female artist, she continued painting until her final days.

US designates Palestine Action a terrorist group

The US Treasury has designated UK-based group Palestine Action as a terrorist organization, freezing any of its US-held assets over alleged acts of terrorism since July 2020. The group was already banned in the UK in 2025, and co-founder Huda Ammori condemned the US move as hypocritical.

Engineered yeast for converting plastic and biomass compounds into food

Scientists have engineered yeast capable of converting compounds from plastics and biomass into edible food. This approach offers a potential method to recycle waste materials into a sustainable food source.

Uefa is preparing criminal legal action against Fifa president Gianni Infantino, accusing him of “criminal mismanagement” over a scrapped plan to sell stakes in World Cup and other competitions to private investors. Uefa has applied for access to documents in Florida to pursue proceedings in Switzerland, alleging the plan was secretly structured to benefit Infantino and undervalued Fifa’s assets. The proposal was abandoned after backlash, and Uefa may also pursue action against other Fifa officials.

Peter Cullen, Voice of Optimus Prime in ‘Transformers,’ Dies at 85

Peter Cullen, the Canadian voice actor best known as Optimus Prime in Transformers and Eeyore in Winnie the Pooh, died Wednesday at his home in Los Angeles at age 85. His agent confirmed the news, and his family honored his legacy of compassion and integrity. Cullen also voiced King Kong in the 1976 remake and created the Predator’s clicking sounds.

Hollywood’s Video Game Era Is Here. Will It Be Any Good?

Hollywood has pivoted from the infamous 1993 Super Mario Bros. flop to the record-breaking 2023 animated Mario film, reflecting a broader shift toward video game adaptations as superhero films decline. With major filmmakers attached to projects like Resident Evil, Elden Ring, and Call of Duty, the industry is betting on gaming’s massive audience—but whether these films will offer better storytelling than the MCU remains uncertain.

Aphantasia Beginner’s Guide

Aphantasia is the inability to visualize, meaning those affected think in concepts rather than mental images. The guide provides a simple “red apple” test to help people determine if they have the condition and explains how to distinguish visualizers from conceptualizers. It also includes resources, scientific insights, and community support for those newly discovering they have aphantasia.

Landsat 9 pass captured the effects of yesterday’s catastrophic glacier collapse

XCancel has been shut down after receiving a cease and desist letter from X Corp. on Monday, August 24th at 8PM EST. The service will remain suspended until further notice as the team seeks legal advice.

Air Conditioning Is Not a Luxury, It Is a Necessity

Air conditioning is a life-saving technology that has drastically reduced heat-related deaths, yet European governments often restrict it through permits, regulations, and cultural resistance, contributing to over 175,000 heat-related deaths annually. Modern AC systems are increasingly energy-efficient, and fears about their health effects are largely unfounded, making these restrictions a costly and deadly mistake.

Alzheimer’s Damage in Mice Reversed with Just Two Injections

Researchers developed a Nano-ERASER therapy that degrades the protein PTBP1, causing brain astrocytes to transform into functional neurons. In Alzheimer’s mouse models, two injections improved cognition, memory, and nest-building while reducing disease markers. The results are promising but early, requiring further testing in primates before human use.

The Dutch are filling gas wells with cement

The Dutch government is permanently sealing Groningen’s inactive gas wells with cement to prevent future governments from resuming production, even though the field holds billions of euros in reserves and official advisers recommended keeping it mothballed. The decision follows local protests over minor earthquakes and reflects how concentrated local opposition can override broader economic interests, similar to Germany’s nuclear phase-out.

Kusama Yayoi has died

Japanese artist Yayoi Kusama has died, according to an announcement from her official website. She was renowned worldwide for her avant-garde works featuring polka dots, pumpkins, and immersive mirror installations.