A federal lawsuit says Waukesha County deputies used Flock cameras over 100 times to track a Navy veteran who recorded a traffic stop. New reporting reveals 3M concealed PFAS in Americans’ blood for decades; a UCSF researcher tied tobacco giants’ food brands to today’s ultra-processed diets. OpenLake beat NVIDIA AIStore in MLPerf storage with 6.72 GiB/s writes and 11.55 GiB/s reads.
🤖 AI & Machine Learning
We Beat MLPerf: Modern Storage for KV Offload and LLM Training
OpenLake achieved the highest read and write bandwidth in the MLPerf Storage v3.0 Llama 3.1 8B checkpointing benchmark, delivering 6.72 GiB/s writes and 11.55 GiB/s reads through its Infinity Core I/O Engine. The result topped published Closed division S3 submissions using one client node and one data parallel instance, outperforming NVIDIA AIStore and Nebius Object Storage. OpenLake’s performance came from an NVMe-backed gateway using io_uring, pinned threads, and I/O coalescing.
- We Beat MLPerf: Modern Storage for KV Offload and LLM Training — theopenlake.com
Poetry book that Anthropic tried to censor
Stanley Kunitz’s 1930 debut poetry collection, Intellectual Things, explores mortality, the body, and the limits of reason. The article profiles Kunitz’s difficult early life—including his father’s suicide and his mother’s erasure of his memory—and his later career as a Pulitzer-winning, two-time U.S. poet laureate and anti-censorship advocate.
Claude’s new system prompt doesn’t want to reproduce song lyrics
Anthropic updated Claude’s system prompts to explicitly forbid reproducing song lyrics, poems, or excerpts, and to ban generating images of copyrighted characters, logos, and artwork—likely a response to music publishers’ lawsuits. The new rules include detailed refusal examples, such as declining to draw Sonic the Hedgehog for a birthday banner. Anthropic publishes these prompts for transparency, allowing easy comparison of changes across model versions.
- Claude’s new system prompt doesn’t want to reproduce song lyrics — simonwillison.net
AI Incident Handling Creates Engineer Blindness
AI-driven incident response tools are increasingly handling routine system issues, but this reduces engineers’ hands-on experience and leaves them less prepared for rare, complex failures—a challenge dubbed the “ironies of automation.” Drawing on aviation’s use of simulators for uncommon emergencies, the articles advocate for realistic incident simulations to help engineers preserve critical troubleshooting skills.
- AI handles incidents, engineers lose touch with their systems — sylvainkalache.com
Artificial Analysis Intelligence Index v4.2
Artificial Analysis released Intelligence Index v4.2 as an interim update, adding more realistic agentic and long-context evaluations (AA-Briefcase, GDP.pdf), removing the saturated GPQA Diamond, and increasing private test-set weighting to prevent gaming. Key results show Anthropic’s Claude Fable 5.1 leading, followed by OpenAI’s GPT-6 Astra, with Anthropic, OpenAI, Meta, and Z.AI on the cost frontier.
- Artificial Analysis Intelligence Index v4.2 — artificialanalysis.ai
Is it too much to ask devs to use AI to review their hand-crafted code?
Developers who pride themselves on hand-crafted code should still consider using AI tools to review their libraries for bugs, memory leaks, and security vulnerabilities. Even if they reject AI-generated code, AI can serve as an advanced linter to identify issues that they then fix manually.
OpenAI to create misalignment reporting framework after wiki incident in training, eval, deployment
XCancel received a cease and desist letter from X Corp on August 24 at 8PM EST, prompting it to halt operations until further notice. The team is seeking legal advice and will not share additional details. They thanked users for their trust over the past two years.
- OpenAI to create misalignment reporting framework after wiki incident in training, eval, deployment — x.com
Microsoft: publisher-hired expert found only ~60K of 8.2M Copilot chats had 16+ words matching news
In new legal filings, Microsoft claims its Copilot AI rarely reproduces significant excerpts from copyrighted works, citing an analysis of 8.2 million chat logs where fewer than 1% contained at least 16 words matching news content and only a few dozen instances of substantial overlap. Publishers, including The New York Times, dispute this, saying discovery evidence shows Microsoft and OpenAI stole journalism, while Microsoft argues the data supports fair use and the transformative purpose of AI training.
- Microsoft: publisher-hired expert found only ~60K of 8.2M Copilot chats had 16+ words matching news — theverge.com
🔒 Security & Privacy
Flock used >100 times to track veteran who recorded traffic stop
A federal lawsuit alleges Waukesha County Sheriff’s deputies used Flock license-plate cameras over 100 times to track Navy veteran Napoleon Jones in retaliation for lawfully recording a traffic stop. Jones was unlawfully stopped and arrested after the recording, released without charges, and later discovered the surveillance searches had no legitimate law enforcement purpose.
Trusting-Trust Attack on Linux Distributions
A trusting-trust attack can be implemented via the GNU strip utility rather than a compiler, by manipulating finished ELF binaries. In NixOS, a single tampered strip in the bootstrap seed propagated the payload across generations into the final system, backdooring nearly all binaries, including a complete graphical installer.
Git hosting that never leaves Europe
A European Git hosting platform keeps all repositories in the EU, shielding them from US legal reach like the CLOUD Act. It blocks low-effort “slop” contributions, never trains AI on user code, and prioritizes availability and developer experience over AI features. It also provides GitHub-compatible APIs and a CLI to simplify migration from GitHub.
- Git hosting that never leaves Europe — pushin.eu
New Zealand is determined not to keep China’s espionage a secret
New Zealand’s intelligence agency has explicitly named China as the only country conducting espionage against it at scale, despite Chinese pressure to soften its language. The latest threat assessment warns that China is aggressively targeting public and private sectors, as well as individuals with security clearances, for sensitive information. This makes New Zealand notably more outspoken than Australia, which does not name China in its own threat assessments.
- New Zealand is determined not to keep China’s espionage a secret — lowyinstitute.org
Nitter has more working instances than before the takedowns
Despite recent takedowns, Nitter now lists more working instances than before, including several clearnet and Tor options. The page also catalogs rate-limited and former instances, and provides guidance on running a public instance, such as obtaining session tokens and handling legal complaints.
- Nitter has more working instances than before the takedowns — codeberg.org
Reversing MikroTik’s Silent Patch: The RouterOS 7.23.4 Fix They Wouldn’t Explain
MikroTik’s coordinated RouterOS updates (7.23.4, 7.24.2, 6.49.21) silently fixed three bugs, disclosed only by a vague SSH changelog line. Reverse-engineering the NPK binaries exposed two exploitable chains: a low-exponent RSA signature forgery leading to an mtget overflow, and an SSH username of “-2” reaching a legacy login path that lets an authenticated read-only session grant full RouterOS command execution. The researcher explicitly did not reproduce a stock, credential-free method to make SSH accept the literal “-2” username.
Microsoft: Spammers adopt ASCII smuggling, an AI prompt injection tactic, to evade email filters
ASCII smuggling, a technique originally used to hide malicious prompts in AI attacks, is now being abused by spammers to evade email filters by encoding text in invisible Unicode tags. Microsoft detected a surge in daily spam detections using this method, rising from about 21,000 to 2.5 million before dropping off in mid-May. These characters are invisible to humans but readable by computers, helping spammers bypass keyword-based detection.
- Microsoft: Spammers adopt ASCII smuggling, an AI prompt injection tactic, to evade email filters — arstechnica.com
US, UK sign MOU to investigate, share info on syndicates behind SE Asia online scam centers
The US and UK signed a memorandum of understanding to coordinate investigations and share intelligence on organized crime syndicates behind online scam centers, many based in Southeast Asia. The partnership aims to disrupt the billions of dollars lost annually to investment and romance fraud schemes run by these networks.
- US, UK sign MOU to investigate, share info on syndicates behind SE Asia online scam centers — therecord.media
Google patches actively exploited Chrome zero-day allowing RCE in sandboxed renderer
Google has patched an actively exploited high-severity zero-day vulnerability in Chrome’s V8 engine, tracked as CVE-2026-85046, along with 11 other security flaws. The type confusion bug could allow remote code execution through malicious HTML pages and has been exploited in the wild. Users are advised to update to Chrome 152.0.7977.82/.83 as soon as possible.
- Google patches actively exploited Chrome zero-day allowing RCE in sandboxed renderer — bleepingcomputer.com
Claude, DeepSeek, and Qwen AI agents used to target governments across Asia
Chinese-speaking hackers used AI models including Claude, DeepSeek, and Qwen to automate cyberattacks against government and industrial targets across Asia, such as Taiwan’s Kuomintang archives, Indonesia’s foreign ministry, mainland Chinese government systems, and Vietnamese hosts. The attackers compromised an Office Automation system, stole 822 account records, and accessed sensitive government, health, and education data. The campaign is attributed to a Chinese-speaking operator with moderate confidence, but no specific hacking group has been identified.
💻 Programming & Developer Tools
Learn Programming with OCaml
An English translation of a French book on learning OCaml has been released, translated by Urmila Nair and funded by The OCaml Software Foundation. The book is openly licensed under CC BY SA 4.0, with a channel for reporting typos and errors.
- Learn Programming with OCaml — usr.lmf.cnrs.fr
There’s No Limit to How Bad Code Can Get
Technical debt has no clean reset, so comparing a bad codebase to a “sinking ship” falsely implies there is an endpoint or floor. At Amazon, a legacy order-processing system stayed operational only through constant pager-driven fixes while repeated re-architecture attempts failed and added more engineers and dysfunction. Ultimately, code quality can keep worsening indefinitely.
- There’s No Limit to How Bad Code Can Get — zachkehs.com
Understanding Rust’s dyn Trait Vtables
Rust’s runtime polymorphism via dyn Trait relies on vtables, which the author compares with C++ virtual functions and CRTP to highlight memory-level differences. The pieces contrast dynamic dispatch with Rust’s compile-time monomorphization and generics, warning that Rust should not be treated as C++ with different syntax. Both emphasize understanding Rust’s unique design, with code examples and experiments available on GitHub.
- Visualizing Rust’s Vtables: How dyn Trait Works In Memory — sofiabelen.github.io
.gitignore Everything by Default
The article suggests using a .gitignore that ignores all files by default and then explicitly whitelists only desired files (e.g., * followed by exceptions like !*.go) to prevent accidental commits of unwanted content. This approach is presented as an alternative for projects cluttered with local files, though not ideal for every repository, and it recommends git check-ignore -v to debug ignored paths.
- .gitignore Everything by Default — packagemain.tech
Twenty Years of jQuery: How a Little Library Rewired Web Development
jQuery’s first stable release in August 2006 made JavaScript approachable by simplifying DOM manipulation and browser compatibility, democratizing front-end development and influencing later frameworks. Although native APIs and modern frameworks like React have rendered its procedural style legacy, jQuery remains embedded in roughly 66% of websites today.
CSS for Printing to Paper (2025)
The article explains how to use CSS for printing web pages, focusing on the @page rule to control paper size, orientation, and margins, as well as @media print for print-specific styles. It demonstrates that @page settings operate outside the DOM and correspond to browser print dialog options, with examples showing how different size and margin values affect output. The author, who creates printable HTML forms for work, shares practical tips from this experience.
- CSS for Printing to Paper (2025) — voussoir.net
Pimacs – An Emacs UI for Pi
Pimacs is a custom Emacs interface for the Pi coding harness, created because the author wanted native Emacs integration rather than using terminal UIs or editor-specific LLM tools. Its design mirrors Magit’s collapsible section layout to support reviewing LLM-generated output, and includes a prompt input, customizable header and mode lines, and image support. The interface is built to be extensible while remaining provider-agnostic.
- Pimacs – An Emacs UI for Pi — ananthakumaran.in
Interview with Janet and Fennel Creator, Calvin Rose
In an interview, Janet creator Bakpakin describes the language as a non-commercial, Lisp-like scripting project where existing users matter more than growth. He considers the core mostly complete, prioritizes stability over added features to avoid bloat, and sees his role as moderating the community rather than strictly running it.
- Interview with Janet and Fennel Creator, Calvin Rose — alexalejandre.com
C++26: std::hive
std::hive is a new C++26 container that provides stable pointers and iterators during insertions and erasures, O(1) amortized insert/erase operations, and good cache locality through multiple memory blocks and a skipfield. It reuses erased slots, supports bidirectional iteration, and is derived from plf::colony.
- C++26: std::hive — sandordargo.com
Anatomy of a Test
The article argues that testing advice often relies on vague concepts or toy examples, so it walks through a real Gleam test from a bookmark app. The test verifies that start_job marks a job as running and removes it from pending jobs, while the author explains Gleam syntax like use, pipes, and let assert, along with file structure and naming conventions. The piece advocates for realistic test examples and clearer test descriptions.
- Anatomy of a Test — jonathan-frere.com
zopt: low-ceremony command line parsing for Zig
zopt is a low-ceremony, lazy command line argument parser for Zig that scans argv on query rather than parsing it upfront. It acts as a thin wrapper around std.process.Args, deferring help text, error handling, and default values to user code for simpler tools. It supports flags, options, repeated values, typed parsing, and best-effort positional arguments.
- zopt: low-ceremony command line parsing for Zig — codeberg.org
Pointing at the error: compiler-style diagnostics in uutils coreutils
Starting with version 0.11.0, uutils Coreutils uses the ariadne library to report parse errors in a compiler-style format when stderr is a terminal. The new diagnostics display the offending argument with a caret marking the exact character or range and include contextual help. This improves error reporting for commands with small-language arguments such as tr, cut, chmod, sort, env, and test.
Rust SIMD on the GPU
Rust’s portable SIMD (core::simd) can now target GPUs by mapping SIMD vectors directly to GPU warp lanes, treating a warp as a wide vector unit. This enables the same SIMD code to compile for both CPU architectures and GPU kernels without source changes, completing a parallelism hierarchy where GPU warps act like CPU threads with SIMD lanes.
- Rust SIMD on the GPU — vectorware.com
Gleam and BEAM- Looking beyond the JVM
The provided content is not an article but a security verification page from Vimeo, requiring human confirmation to access. No news or substantive information is available to summarize.
- Gleam and BEAM- Looking beyond the JVM — vimeo.com
Bespoke: A Programming Language for People Who Say Please
The article humorously proposes “Bespoke,” a statically typed programming language that enforces British etiquette and verbose courteous syntax, replacing typical imperative commands with formal addresses to the compiler and polite requests. It outlines features such as apology-requiring mutations, diplomatic control flow, and British orthography, while noting the implementation is less advanced than its etiquette manual and unrelated to a prior esolang of the same name.
- Bespoke: A Programming Language for People Who Say Please — blog.hofstede.it
🔧 Hardware & Infrastructure
Rebuilding a 1995 GPS Time Server so I don’t get Telstra’d
The author refurbished a 1995 TrueTime XL-AK GPS time server using a Raspberry Pi 5 and a GNSS HAT, creating a stratum 1 NTP time server with Chrony, PPS, and the original LCD/LED display. The project was partly inspired by a similar GPS server causing Australia’s Telstra outage, and the author plans to deploy it at VCF Midwest before potentially adding it to the NTP Pool.
- Rebuilding a 1995 GPS Time Server so I don’t get Telstra’d — jeffgeerling.com
A bizarre Commodore 64 peripheral, a mime, and some pretty bad ads
Mimic Systems promoted its Spartan, an Apple II+ emulator for the Commodore 64, with bizarre mime-themed ads for nearly two years before finally shipping it in 1986, by which time poor reviews cited long delays and bugs. Despite the product’s commercial failure, a fully equipped Spartan still sold for more than $5,500 at an auction in October 2024.
- A bizarre Commodore 64 peripheral, a mime, and some pretty bad ads — buttondown.com
Terpstra Keyboard
The Terpstra Keyboard is a musical instrument featuring 280 color-changing continuous controllers, as highlighted on its promotional site. The page provides options to learn about the product, purchase it, try it out, and engage with its community of backers and forum users.
- Terpstra Keyboard — terpstrakeyboard.com
BepiColombo sheds its ride and starts the final glide to Mercury
After eight years in space, the BepiColombo Mercury probe has separated from its Mercury Transfer Module as it approaches the planet. The uncoupling marks the start of the final phase of its journey to enter orbit around Mercury.
- BepiColombo sheds its ride and starts the final glide to Mercury — theregister.com
What filesystem are you running on your NAS/Backups, and why?
The article asks readers to share which filesystem they use for their NAS and backup setups, including the reasons behind their choice. It also inquires about how they manage multiple disks, whether through filesystem-native RAID, mdadm, other software solutions, or hardware RAID.
Nvidia plans DLSS 5 for RTX 40-series GPUs after RTX 50 tuning.
Nvidia will officially bring DLSS 5 to older RTX 40-series GPUs after first tuning RTX 50 performance, but has not confirmed support for RTX 30 or 20-series. Gamers will only get a basic on/off toggle, not the granular controls available to modders via leaked versions. Nvidia also declined to reconfirm whether previously announced DLSS 5 games will still ship with the technology.
📈 Tech Business & Startups
Waymo racked up 8,300 SF parking tickets and nearly $1M in fines
Waymo’s robotaxis have received 8,309 parking tickets in San Francisco since the start of 2025, racking up nearly $1 million in fines—a sharp increase from just 589 tickets in 2024 as the fleet rapidly expanded. The citations, concentrated at busy pickup spots like Fisherman’s Wharf, are viewed as a cost of scale rather than a safety concern, though critics note the fines have done little to alter the company’s behavior.
Seeking Designer for a Small Company
An aspiring startup founder is seeking a freelance human graphic designer to create logo/branding and a launch website, explicitly requiring no generative AI use and a portfolio. Interest in spaceflight, retrofuturism, and strong font opinions are a plus. The post asks for recommendations or self-submissions with contact details.
Betaworks’ 5.5% Hugging Face stake worth $650M; A.Capital could gain $1.5B
Betaworks wrote Hugging Face’s first $150,000 check after selecting the startup for its 2016 accelerator. That stake has grown to 5.5%, now worth about $650 million at Hugging Face’s $11.9 billion acquisition price.
Scan.com raises $220M Series C ($90M equity) for AI imaging referral matching
Scan.com, an AI-powered medical imaging network, raised $220 million in combined equity and debt financing, including a $90 million Series C equity round led by Noteus Partners and $130 million in debt facilities. The company will use the funding to expand its U.S. imaging provider network and invest in its API and agentic AI infrastructure for scheduling and diagnostic routing.
- Scan.com raises $220M Series C ($90M equity) for AI imaging referral matching — mobihealthnews.com
India’s Slice, now a small finance bank, raises $100M at ~$450M valuation, down from $1.4B
Slice raised $100 million at a valuation of roughly $450 million, a 68% drop from its earlier $1.4 billion fintech valuation, in a round led by Neo Wealth. The downround reflects its transition into a licensed small finance bank, now assessed on banking metrics rather than fintech growth multiples, with part of the proceeds going to secondary share sales.
- India’s Slice, now a small finance bank, raises $100M at ~$450M valuation, down from $1.4B — moneycontrol.com
🌍 Society & Environment
The Luxuries in Life
The author reflects on Labor Day as the symbolic transition from summer to fall, noting that at age 60 they now focus more on what truly matters. They share Amy’s list of “real luxuries,” including time, health, a quiet mind, meaningful relationships, and rest without guilt.
- The Luxuries in Life — feld.com
Bob and Van
The author fondly recalls former neighbors Bob and Van, who were kind and helpful but whose political views he never knew. He argues that people were better off before everyone’s political opinions became public, and opposes communities that demand condemnation of others based on politics in everyday, non-political settings.
- Bob and Van — marco.org
How the Tobacco Industry Drove the Rise of Ultra-Processed Foods (2025)
In 1963, R.J. Reynolds’ purchase of Hawaiian Punch marked the tobacco industry’s entry into food, leading to the aggressive acquisition of major brands and production of ultra-processed foods marketed to children. These foods now account for roughly 60% of calories in Americans’ diets and are linked to rising obesity and chronic disease, according to UCSF professor Laura Schmidt.
- How the Tobacco Industry Drove the Rise of Ultra-Processed Foods (2025) — vcresearch.berkeley.edu
How the Disaster of “Forever Chemicals” Was Kept Secret
PFAS “forever chemicals” are now in nearly everyone, but 3M knew decades ago that they had entered the public’s blood and concealed it. After years of investigation, reporter Sharon Lerner finally got a former 3M scientist to explain how the company kept the contamination secret for so long.
- How the Disaster of “Forever Chemicals” Was Kept Secret — propublica.org
DHH Nazism Funded by 1Password VP Who Wrote “Honest Security”
Jason Meller, who authored Kolide/1Password’s “honest.security” ethics principles, is accused of violating them by dismissing criticism of David Heinemeier Hansson, a Rails Foundation figure with documented Nazi and white-supremacist rhetoric. The article details DHH’s offensive posts—praising a white-nationalist march, endorsing “remigration,” and attacking Romani people—while 1Password funds his foundation and Meller serves as a director. This exposes hypocrisy between 1Password’s stated values of transparency and trust and its financial and leadership ties to DHH.
- DHH Nazism Funded by 1Password VP Who Wrote “Honest Security” — flyingpenguin.com
Pentagon rescinds new testosterone screening policy without explanation
The Pentagon temporarily rescinded its new clinical guidance on mandatory testosterone screening just one day after its release, citing needed updates. The withdrawn guidelines included screening questionnaires for men aged 30 and older and women, potentially leading to blood tests and hormone therapy. An interim guidance remains in effect, with final guidance promised shortly.
- Pentagon rescinds new testosterone screening policy without explanation — arstechnica.com
Global warming will exceed 1.5-degree limit, UN says
The UN has acknowledged that global warming will surpass the 1.5-degree Celsius Paris accord threshold in the next few years. Instead of abandoning the goal, its new report recommends an “overshoot” strategy: peak temperatures, then bring them back below the limit by ending fossil fuel use and removing carbon dioxide from the air. This path would still bring decades of worsening weather disasters, rising seas, and other climate impacts.
‘Guilt by association’: why 7 int’l aid workers were killed in Israeli airstrike
A Guardian investigation found that seven World Central Kitchen aid workers killed in an Israeli airstrike in April 2024 were targeted under an IDF practice of guilt by association, not simply a mistake. A military prosecutor declined criminal proceedings, despite commanders justifying the strike by claiming the workers had “mixed together” with armed guards, which they equated with Hamas ties. The investigation highlights how this policy endangers civilians, amid over 70,000 Palestinian deaths in the Gaza war.
I want to see the same moon as you
The author recalls how he and his wife, separated by 2,000 miles, shared the same moon, likening it to the early internet’s shared experiences like a common Instagram front page. He laments that personalized algorithms have isolated users into separate feeds, removing that sense of collective online experience.
NYC PIT Crew
New York City’s Public Interest Technology (PIT) Crew is an in-house team of engineers, designers, and researchers creating digital public goods to improve government interactions for New Yorkers. Their goal is to make these experiences dignified and delightful, and they encourage others to join or subscribe for updates.
- NYC PIT Crew — nyc.gov
Kai-Fu Lee Q&A: China open-model edge, AI jobs, US chip export curbs, Chinese AI in developing world
Chinese AI companies are narrowing the gap with US rivals despite longstanding restrictions on advanced chips, according to AI pioneer and investor Kai-Fu Lee. He argues that AI’s rapid improvement and falling costs will drive unprecedented adoption, forcing companies to rethink hierarchies as AI workers take on larger tasks. Lee also warns that many CEOs treat AI as “theater” and have not grasped how fundamentally jobs and organizations will change.
- Kai-Fu Lee Q&A: China open-model edge, AI jobs, US chip export curbs, Chinese AI in developing world — bloomberg.com
US promised Nvidia chips for Armenian data center to broker preliminary peace deal with Azerbaijan
U.S. negotiators offered Armenia expanded access to Nvidia’s AI chips for a data-center project as an incentive to help secure a preliminary peace deal with Azerbaijan last year. The previously unreported inducement highlights what officials describe as “chip diplomacy.”