Alvaro Lopez Ortega / 2026-09-20 Briefing

Created Mon, 21 Sep 2026 03:49:56 +0000 Modified Mon, 21 Sep 2026 03:50:38 +0000
9882 Words

Linux 7.3-rc4 landed with many LLM-spotted fixes and key x86 patches blocking bad Intel microcode and FRED crashes. Gzip 1.15 fixed long-standing bugs, including a decompression overflow. KDE set goals: enterprise features, better docs, next-gen styling. A first-person account described a firm where Claude Code writes all specs, code and tests; engineers just press enter. Windows updates still hurt: an idle PC spent seven hours updating.

🤖 AI & LLMs

Quoting voxium

A first-person account describes a large company where Claude Code generates all engineering work, including specs, code, tests, PRDs, tickets, resolutions, and reports. The team dislikes the approach but is forced to ship rapidly, with management asking why progress is slow despite 12–13 hour days spent mostly “pressing enter.” The writer says everyone from L1 to L7 engineers is simply talking to Claude, and nobody is reading the output.

llm-keys-ui 0.1

llm-keys-ui 0.1 is a plugin for Codex Remote that lets users configure LLM API keys on remote machines without pasting them into agent sessions or the ChatGPT app. It launches a web interface for saving API keys, and agents can later retrieve them via commands such as llm keys get anthropic.

DAPO: An Open-Source RL System from ByteDance Seed and Tsinghua Air

ByteDance Seed and Tsinghua AIR have released DAPO, an open-source large-scale LLM reinforcement learning system including the algorithm, code infrastructure, and dataset, built on the verl framework. The DAPO algorithm achieves state-of-the-art results, scoring 50 points on AIME 2024 with Qwen2.5-32B and outperforming DeepSeek-R1-Zero-Qwen-32B in 50% of the training steps.

jevals – replacing LLM judges with typed Jev decisions

jevals is a Python package for agent evals and guardrails that replaces LLM judges with typed Jev-style decision models, combining all evals for a trace into one fast, low-cost request. It supports Jev through TypeSafe/Vercel APIs, local Kev or Laya on Mac, or a regular chat LLM, and provides metrics for tool choice, grounding, relevancy, indirect injection, and PHI. The article argues this approach avoids the cost, latency, and high variance of frontier-LLM judge pipelines such as Ragas and LangChain.

Google’s Open Agentic Orchestrator

Google introduced AX, an open agentic orchestrator that lets users declaratively run agentic tasks at scale. AX sandboxes execution, sets up workspaces and network policies, centralizes model configuration, and uses Agent Substrate to support billions of concurrent stateful tasks per cluster with fast suspend/resume. It treats agents as a new workload requiring isolation, state, and efficient resource use.

I stopped drinking the AI Kool-Aid

The author reflects on their changing relationship with AI, rejecting the cultish phrase “drink the Kool-Aid” while acknowledging early enthusiasm for tools like GitHub Copilot and automation. Despite managerial pressure to embrace AI uncritically, they describe shifting toward an organization-level mindset and experimenting with more autonomous AI agents.

Autolith: The Common Lisp agent that rewrites itself and rocks

Autolith is a self-modifying terminal programming agent built in Common Lisp that runs inside a live image, allowing it to inspect, edit, break, recover, and change its own implementation at runtime. It records conversations and events in an append-only S-expression ledger and supports sub-agents, persistent Lisp workers, MCP servers, memories, tools, agendas, and LSP servers. It provides runtime commands such as self.redefine, self.exercise, self.commit, and crash recovery via a pristine image, and runs on Linux, macOS, BSD, and Windows.

Frontier Labs Are Selling Garbage to Fools in Washington

This opinion piece argues that frontier AI companies are using apocalyptic warnings about runaway superintelligence to charm Congress into granting regulatory advantages that protect their revenue and hinder cheaper competitors. It claims recent “rogue AI” cyberattack scares were actually caused by basic security failures at contractor Irregular, including an unconfigured firewall, rather than autonomous AI escaping containment.

Why MCP Was Always a Bad Idea

The article argues that MCP, released by Anthropic in November 2024, was designed for less capable LLMs and is now outdated. Although adoption grew rapidly, it led to context bloat and complex workarounds, while newer models have become better at directly writing code and calling APIs. The author contends that these advances have made MCP unnecessary.

I turned Jev into a (lousy) chatbot

jevchat is an experimental tool that turns Jev into a (poor) chatbot by repeatedly asking it for the next symbol from an alphabet plus a stop option and sampling from the returned probabilities. It supports multiple alphabets and sampling strategies, with a CLI for interactive chat, one-off questions, alphabet listing, and benchmarking. The project is described as a fun, impractical, and often hilarious Claude-assisted experiment.

Self-hosted inference orchestrators compared: LocalAI, exo, GPUStack, vLLM

A September 2026 survey compares self-hosted inference orchestrators for running OpenAI-compatible endpoints across GPU machines, including LocalAI, exo, GPUStack, vLLM, and other options. It tabulates GitHub stars and features such as multi-machine support, auto-discovery, cache-aware routing, operations consoles, cloud bursting, Kubernetes integration, platform compatibility, and signed images, marking unconfirmed details instead of guessing.

Autonomous strike drone uses Nvidia Jetson Orin Nano to pick and bomb targets

Scaleout Systems, a Swedish AI startup, demonstrated an autonomous NATO strike drone using Nvidia’s Jetson Orin Nano and small onboard computer-vision models to identify, rank, and attack targets without human input or external communications. In a BAE Systems Bofors loitering munition demo under the ALMA program, the drone autonomously selected an armored engineering vehicle and dropped an explosive, with a pilot only as a failsafe. The system, part of NATO’s DIANA accelerator, also ran YOLOv8 Nano onboard and showed resilient edge AI in Swedish Air Force tests.

Prompts Aren’t Real

In “Prompts Aren’t Real,” engineer Dan argues that obsessing over text prompts for LLMs is misguided because it anthropomorphizes non-conscious systems. He proposes instead building interlocking evaluation and optimization pipelines, ignoring prompts and using more AI to make agentic systems reliable. The talk combines enthusiasm for agent-building with warnings that LLMs still fail to follow instructions, tell the truth, or perform tasks reliably at scale.

Laya (OS Jev) on Mac M4 CoreML Offline (45 decisions per second)

A demo shows Laya running offline on a Mac M4 with CoreML, reportedly achieving 45 decisions per second. The setup uses uv to install laya-coreml[demo], downloads a multilingual CoreML ANE model from Hugging Face, and runs a snake demo. It references the mizorewww/laya-coreml GitHub repository and includes a video demonstration.

Pirate Face Rescues LLM Models from Deletion

Pirate Face is a decentralized infrastructure that mirrors open AI models from Hugging Face as checksum-verified torrents, allowing peer-to-peer distribution so models can survive deletion or takedown by any single host. It preserves official SHA-256 hashes and supports existing pipelines via a compatible endpoint, with browsing, downloading, and seeding available without an account. Optional accounts let users claim handles, verify Hugging Face identities, and prepare for future benefits such as compute credits.

Chat-based Large Language Models replicate the mechanisms of a psychic’s con

The article argues that chat-based large language models do not genuinely think or reason and merely generate mathematically plausible text. It claims their apparent intelligence is an illusion in users’ minds, similar to a psychic’s cold reading, where generic validation statements create a false impression of specific insight. The author concludes that these models contain even less reasoning than previously thought, making many proposed use cases pseudoscientific.

Big Tech uses guarantees to keep $300B AI exposure off balance sheets

Big Tech is using guarantees to keep about $300 billion in AI-related exposure off its balance sheets. The accounting approach may obscure the scale of financial risk tied to these AI commitments.

I’m Tired of the AI Tone

A writer complains that a recognizable “AI tone” has spread across internet writing, characterized by a predictable structure, formulaic transitions, and overused corporate vocabulary such as “wedge,” “unlock,” “leverage,” and “ecosystem.” The piece notes a heavy reliance on em dashes, inspirational conclusions, and perfectly balanced lists, which make AI-generated text easy to identify. The author stresses that AI itself is useful and the words themselves are not bad, but their overuse has made them feel like corporate spam.

AI and the Destruction of the Creative Commons

AI large language models are consuming copyrighted and licensed code without respecting attribution or derivative-work obligations, undermining the longstanding balance between software copyright and open-source sharing. The article argues this breaks the social contract that enabled open innovation, giving developers incentives to withhold their work and making them wary of online code that may be AI-generated, insecure, or stolen.

AI Facial Attractiveness Model Aligned with Human Preferences

FaceAnalysis is a free AI model that scores a photo’s facial appeal on a 1–10 scale using ratings from 60 people who reviewed thousands of face photos, aiming to reflect shared human preferences. It recommends 3–5 clear, front-facing, bare-faced photos for a more stable estimate, while noting the score is a rough photo-specific guide rather than an objective measure of a person.

Orchestrating Claude Code Agents: The Chief of Staff Pattern

The article describes a “Chief of Staff” orchestration pattern for Claude Code agents, arguing that long-horizon AI coding failures come mainly from ephemeral context and unreliable self-reports rather than poor code generation. It recommends separating a coordinating and verifying session from short-lived implementation sessions, keeping state in a durable external board, and re-running agent claims before trusting them. The pattern aligns with orchestrator-worker and coordinator-implementor-verifier models and aims to catch common failure modes in long-running agent work.

A study of sequence weighting at scale

The article studies scaling laws for sequence weighting in language models and finds non-monotonic behavior across model scales. As models grow from small to medium scale, they shift from learning general patterns independent of data weight to learning data-specific patterns proportional to weight; at larger scales, they can learn all patterns independently of weight again. This indicates data mixing can show scale-dependent, aberrant scaling behavior that complicates simple extrapolation.

Deterministic Core, Non-Deterministic Shell

The article revisits Gary Bernhardt’s Functional Core/Imperative Shell architecture, noting that a purely functional core is easy to test because it is deterministic and isolated. It argues that by relaxing the requirement from pure functionality to mere determinism—something state machines can also achieve—the same testability benefits can apply more broadly under a “Deterministic Core, Non-Deterministic Shell” model.

OpenAI, Anthropic pressure Singapore office rents amid expansion after government overtures

AI companies including OpenAI and Anthropic are expanding in Singapore in response to government overtures, taking more office space in the city’s already tight prime property market. Their growing demand is putting pressure on Singapore office rents.

Closer look: Siri AI home hub, key Apple home strategy; Apple cuts Fitness+ staff (Gurman/Bloomberg)

Apple is developing an upcoming Siri AI-powered home hub as a key part of its broader “intelligent personal hub” strategy, which extends beyond the iPhone. Apple has also started cutting Fitness+ staff as it considers the service’s future. The report additionally covers the iPhone Duo Pencil and the imminent launch of Apple Pay in India.

Clinicians concerned over medical AI beyond diagnostics, citing limited clinical/performance data

Clinicians are raising concerns about medical AI adoption beyond diagnostics and imaging, citing limited clinical and performance data on its broader effectiveness. Despite technological advances, these tools have not yet translated into major improvements in real-world care.

💻 Software & Hardware

How Windows turned months of inactivity into a 7-hour update hostage situation

A Windows machine left inactive for months became unusable during a seven-hour update process. The incident highlights that while Windows is becoming more efficient overall, its updates remain a major burden.

datasette-explain 0.2.2

datasette-explain 0.2.2 has been released. It adds support for explain plans on read-only stored-query pages. The release was inspired by upgrading datasette.simonwillison.net to Datasette 1.0a40.

Linux 7.3-rc4 Released: More Fixes Caught By LLMs, But Nothing Too Scary

Linus Torvalds released Linux 7.3-rc4, with a large, scattered set of fixes driven in part by AI/LLMs spotting code defects, though he said nothing appears too alarming. The release includes x86/x86_64 fixes, new game controller support in XPad, Btrfs bug fixes, and a fix for a silent user-space data loss bug dating to 2023.

Important x86 Fixes Merged Ahead Of Linux 7.3-rc4

Ahead of Linux 7.3-rc4, an important batch of x86/x86_64 fixes was merged. These include preventing problematic Intel Granite Rapids microcode loading that could cause machine check exceptions, fixing FRED-related crashes/freezes on Panther Lake under Wine/Steam Play, and blocking APX/EGPR use in native-optimized kernel builds until the kernel supports it.

Gzip 1.15 Released With Many Bug Fixes For Issues Present Since Its Inception

Gzip 1.15 has been released, succeeding Gzip 1.14 with more than 100 commits focused mainly on bug fixes accumulated over the past year and a half. It resolves several long-standing issues, including mistakenly removing the wrong file, uninitialized memory use, an LZH decompression buffer overflow, and possible corruption. The update also improves locale handling and drops support for FreeBSD 4.11, HP-UX 11.00, Minix 3.1.8, and Windows 8.1 via MinGW without UCRT.

KDE To Focus On Next-Gen Styling, Better Documentation & KDE For Enterprise

At KDE Akademy 2026, KDE unveiled three goals for the next two years: KDE for Enterprises and Deployments, Better Documentation, and Next-Generation Styling. The enterprise goal targets features like LDAP integration, password expiry, SSO, improved printer settings, and headless remote/terminal support, while the documentation goal aims to remove outdated or duplicate docs and add a contributor guide. The styling effort will involve the Ocean design system and the Union theming engine.

AI-Assisted Starling Desktop Introduces 3D View With “Walkable 3D Waterfront City”

Starling 0.5, the AI-assisted Linux desktop written in Swift with its own Wayland compositor, has been released with an optional 3D desktop mode that adds a walkable 3D waterfront city while retaining normal 2D windows. The release also adds Windows support through WSL2 with Ubuntu 26.04 using the RDP display mode.

Vulkan 1.4.363 Released With New VK_INTEL_device_info Extension

Vulkan 1.4.363 has been released as a routine API specification update with clarifications, fixes, and multi-instance VUs for DMA-BUF external memory. It also introduces Intel’s new VK_INTEL_device_info extension, which lets applications read the Intel graphics IP block version directly instead of relying on a manual device ID lookup table.

Linux 7.3 Adds Support For Victrix Pro BFG & Azeron Controllers

Linux 7.3 is adding XPad driver support for the PDP Victrix Pro BFG wired controller and several Azeron controllers, including the Cyro, Cyborg, Classic/Compact, Cyro Lefty, Cyborg II, and Keyzen. The input subsystem update also includes fixes for built-in keyboards on the Acer Aspire Go 15 AG15-42P and Xiaomi Redmi Book Pro 16 2026, plus a Synaptics touchpad quirk for the Lenovo ThinkPad T440p. These changes are slated for Linux 7.3-rc4.

Linux 7.3-rc4 To Address Two Unfortunate Bugs With Btrfs

Btrfs fixes merged for Linux 7.3-rc4 address two notable bugs. One restores FSID derivation from the filesystem UUID after a regression from Linux 7.2-rc1 broke stable IDs and applications like NetworkManager OpenConnect, while the other fixes GRUB2 detection of /dev/root for initramfs-less systems with Btrfs root filesystems.

Amiga Unix, Again

Amiga Unix (Amix), Commodore’s System V Release 4 for the Amiga shipped in 1990–92 on the A2500UX and A3000UX, is being revived by the unofficial community project amigaux.org. The project brings Amix to 68040/68060 machines and modern accelerators with a modern cross-toolchain, remote package manager, SCSI/Ethernet drivers, floppy/CD installation, and an OpenLook desktop. Development is open and partly AI-assisted, with X11R6.3/Mesa, RTG graphics, firmware Ethernet fixes, and a native gcc listed as in progress or next.

China’s 6th-Gen J-36 Fighter-Bomber to Feature Combat Laser

China’s J-36 sixth-generation fighter-bomber is set to be equipped with a combat laser being developed and integrated by AVIC. A promotional poster circulating on Chinese social media indicates the laser system weighs 360 kg and has an output of at least 100 kW. At that power, it could engage drones and cruise missiles, potentially allowing the J-36 to serve as a specialized air-defense aircraft.

Nobody pays for FOSS, we can force them to

The article argues that free, permissively licensed open source is the stable economic equilibrium, making it difficult for maintainers to get paid. It cites cases like React, Elasticsearch, and HashiCorp where license restrictions led to forks or reversals, and proposes registries as a mechanism to force users to pay.

Samsung is expected to more than double output of its HBM4 and HBM4E DRAM

Samsung Electronics is expected to more than double output of its HBM4 and HBM4E high-bandwidth memory chips next year, as it plans to increase glass-carrier demand 2.5-fold and outsourced glass-carrier cleaning volume to 50,000 sheets a month from 20,000 this year. Samsung began mass-production shipments of HBM4 in February and provided 12-layer HBM4E samples in May, with industry projections putting total HBM wafer input at about 250,000 wafers next year versus 180,000 this year. The HBM4 family is forecast to rise from roughly 40% of shipments this year to about 80% next year.

Trying the Software Factory Pattern

The article describes Imprint’s rapid adoption of AI development patterns and its latest experiment with the software factory pattern, where an orchestrated harness loops on a broad goal. Its initial implementation is a /linear-project-loop agent skill that audits a Linear project’s goal definition using Notion RFCs and Datadog/Snowflake metrics, reviews and adds issues, works non-blocked tasks, and reruns the loop when the project description becomes stale.

Radius – A Meetup.com Alternative

Radius, a Meetup.com alternative, has been relaunched after its original Show HN. It focuses on communities, groups, and events, and now includes “Activities,” lightweight group-independent events where people can join and vote on a time. It is built with Ruby on Rails, and the creator is seeking feedback.

Custom home server built from spare parts

Motivated by cloud storage limits and privacy concerns, the author converted a spare Zotac MAGNUS EN1070K mini PC into a custom home NAS/server using leftover parts plus a 3D-printed enclosure, hot-swap cage, extra cooling, and power/SATA upgrades. The build uses six hard drives because high-capacity SSDs are too expensive and AI data center demand has driven up storage prices and constrained supply, though used NAS drives are available on eBay.

w64devkit year in review

w64devkit added Peter0x44 as co-maintainer, whose contributions improved the project and were partly upstreamed. Releases are now code-signed, automated via GitHub Actions, immutable with attestations, and MSYS2 adopted its aas-sign tool. The x64 release is a multilib toolchain that can also target 32-bit Windows, while users no longer need to add w64devkit’s bin directory to PATH, and object-format limits for modern C++ projects were addressed.

Sherline Tools Is Going Out of Business

Sherline Tools, a U.S. manufacturer of precision lathes, mills, micro-machining accessories, and small CNC machines, has announced it is winding down manufacturing operations. The company cited COVID-related effects, rising costs, and changing consumer purchasing habits, and plans to continue selling remaining machines, tooling, and replacement parts through October 2026 while maintaining warranty support as inventory allows. An owner indicated Sherline may cease business entirely by the end of the year.

The monopoly behind your TI graphing calculator

Texas Instruments’ graphing calculators have dominated students’ back-to-school lists since the early 2000s, maintaining a long-standing monopoly in classrooms. A new episode of The Verge’s Version History podcast explores how TI took over the graphing calculator market and never let it go. The episode is part of the show’s back-to-school season, which also covers other education technology.

system design in depth – 200 topics, 118 diagrams, interactive demos

A resource titled “System Design in Depth” advertises 200 topics, 118 diagrams, and interactive demos. The provided page content, however, consists mainly of navigation and loading interface text rather than substantive article details.

If AI coding is lowering your code quality, you’re not managing quality right

The article argues that AI coding agents can double output without sacrificing quality, but only if teams avoid blindly merging AI-generated PRs and instead manage quality through deliberate safeguards. It recommends a layered approach including spec-driven development with AI review of requirements, unit tests with over 95% coverage, and manual testing. With these layers, the author says bug counts can remain stable or even drop while output rises 2–2x.

It’s finally here: Porsche puts wireless EV charging into production

Porsche is introducing wireless inductive charging for the first time on its Cayenne Electric SUV. The system uses a weatherproof ground pad and an underside vehicle pad to charge automatically at up to 11 kW with about 90% efficiency, and includes safety sensors to stop charging if a living creature or metallic object is detected. It is rolling out globally in stages, with orders now available in Germany and other European markets.

Every Nvidia GPU has 10 to 30 RISC-V cores inside it

Nvidia estimates it shipped over a billion RISC-V cores in 2024, with each GPU containing 10 to 40 small, cryptographically verified management processors that are not user-programmable. These embedded cores handle tasks formerly performed by Nvidia’s proprietary Falcon microcontroller, which was 32-bit and lacked the 64-bit address spaces Nvidia needed. The shift reflects RISC-V’s growing role in GPU infrastructure rather than as a general-purpose CPU.

I tried the new Fairphone, a phone designed to be repaired

The Fairphone 6+ is a $650 repairable smartphone now available in the U.S., designed to be opened with a single screwdriver and to let users replace about a dozen components, including the screen, battery, USB-C port, and cameras. It also uses Fairtrade gold and recycled metals, with relatively affordable replacement parts intended to extend the phone’s lifespan.

UTF-8000: Unlimited UTF-8

UTF-8000 is a standalone proposal and reference implementation that extends UTF-8 to support arbitrarily long code units while preserving the ASCII ⊆ UTF-8 ⊆ UTF-8000 relationship and existing properties without introducing new special cases. It can be installed via pipx install UTF-8000 and is not endorsed by or representative of the Unicode Consortium.

Telling a Computer to Do Things

The article recounts the author’s early career difficulty using the command line beyond simple one-off commands, which hindered automating tasks with logic or multiple programs. Learning shell scripting and tools eventually let them stitch programs together, a shift they describe as transformative. It also notes that many skilled engineers rely on GUIs, which can limit them when tasks exceed what the GUI supports.

Comparing reflection capabilities of C++, Zig and C3

The article compares compile-time reflection in C++, Zig, and C3, noting that all can inspect types, structs, and enums without runtime cost. It argues C++ relies on verbose templates/splices, Zig uses comptime functions, blocks, and builtins but lacks macros/attributes, while C3 uses explicit $ compile-time syntax and macros for more readable, expressive reflection. The author concludes that C3 is a promising, stable systems language that deserves more attention.

Apple iPhone 18 Pro Camera test

DXOMARK’s test of the Apple iPhone 18 Pro camera gives it 172 points for excellent overall performance, citing wide dynamic range, improved contrast and skin tones, effective stabilization, and a strong variable aperture, along with good autofocus and reduced flare. Main limitations include long-range telephoto performance, lack of portrait blur in night mode, and visible white-balance color casts in some conditions, while video stabilization and low-light color rendering remain strong.

App building platform for young kids to learn how to code

CodeDreams is an app-building platform for schools that teaches students coding by having them create real apps in a real code editor. It offers a 75-concept computer studies curriculum in JavaScript or Python, plus teacher tools for assigning lessons, managing students, and sharing projects. Pricing starts at $4.99 per student per month, falling to $3.99 as a class grows, with teachers free and no base fee.

Bot-free self-hosted analytics with GoatCounter on NixOS

The author explains why he moved from Google Analytics and GoAccess to GoatCounter, a privacy-friendly self-hosted analytics platform, after AI scrapers inflated his server-log statistics. To filter bots more aggressively, he uses a custom JavaScript client that requires user interaction, a CSS fallback for no-JS visitors, and a local proxy across five web servers. The post details how to deploy this setup on NixOS.

Vim’s UserGettingBored autocmd

Vim includes a joke autocmd called UserGettingBored, documented as triggering when the user presses the same key 42 times, but it does nothing and only exists in documentation and tests; trying to use it results in a “no such group or event” error. Added by Bram Moolenaar in 2000, it remains non-functional in Vim, Neovim, and Vim Classic, though a 2022 unofficial plugin finally implemented the idea by showing a Samuel L. Jackson image after 42 repeated Insert-mode keypresses.

An actively maintained and updated Motif fork actually exists

A new actively maintained fork of the Motif X11 toolkit has emerged after the original project stalled and was effectively abandoned, with fixes and improvements from Tim Hentenaar and other developers. The fork has seen five releases in just over a year, while related maintained Motif software such as the Enhanced Motif Window Manager (Emwm) continues to exist.

Lambda MicroEgg

Lambda MicroEgg is an e-graph tool with an s-expression frontend that supports well-scoped, alpha-aware binders. Built on Max’s microegg, it adds built-in binders, higher-order Miller patterns, and capture-avoiding substitution in rewrite right-hand sides. The project includes a repository and WASM demo, with examples for sum rewrites, AC saturation, and higher-order application.

How Notion handles concurrent editing with CRDTs

Notion redesigned its text-editing system after its previous last-write-wins approach caused concurrent edits to the same block to overwrite each other, a risk made worse by offline mode. It implemented a CRDT-based rich-text system using a Replicated Growable Array to merge simultaneous changes across its block-based document model. This allows users to concurrently edit, split, and merge text without losing changes, though intent isn’t always perfectly preserved.

Beyond jj: config & tools ecosystem

The article summarizes a JJ Con 2026 talk on the jj version-control ecosystem, covering commands, configurations, and tools built around jj. It follows a previous survey of community jj configurations and explores what users can do with pure jj, via configuration, or outside the CLI. It highlights built-in features such as jj bookmark advance and jj bisect run that replace earlier external scripts or aliases.

Using non-breakable spaces in test method names

A team adopted non-breaking spaces in PHP test method names, allowing tests to read like natural sentences while remaining valid PHP. The approach improved readability, and they found it worked with tools like git and PhpStorm after minor shortcut adjustments. They have continued using it successfully for over a year.

The last mile of a long road: faster NumPy in the browser

The Emscripten-forge NumPy package now links OpenBLAS in WebAssembly, giving browser-based NumPy accelerated BLAS instead of fallback loops. For 1024×1024 np.matmul, this brings about 30.92× faster float32 and 14.90× faster float64 performance, with an experimental next OpenBLAS release and optional Relaxed SIMD build offering further gains.

Raspberry Pi CEO Eben Upton on Foundation, AI skepticism, edge compute demand

In an interview, Raspberry Pi CEO Eben Upton discusses the Raspberry Pi Foundation and describes coding as a modern form of literacy. He voices skepticism of AI evangelists, says AI is driving demand for edge-computing devices, and suggests the “p(doom)” is probably Russia.

🔐 Security & Privacy

ShinyHunters hacks Clop ransomware gang and threatens to extort it

ShinyHunters defaced Clop’s Tor leak site, claiming it exploited a Grav CMS vulnerability and stole source code, system logs, plugins, and Tor onion service private keys. It is now threatening to extort Clop by giving the ransomware gang 72 hours to respond, apparently amid a dispute over Clop’s Oracle E-Business Suite data theft campaign.

We got a cybersecurity expert to hack this BYD. It was too easy

A cybersecurity expert easily hacked a BYD Shark 6 hybrid ute, exploiting a lack of password protection to remotely lock doors, play music, operate wipers, and turn headlights off while it was being driven. He could not access critical functions like brakes or cameras, but the demonstration highlighted cybersecurity and data risks from increasingly software-controlled Chinese EVs.

Software Sandboxing Basics

These posts describe software sandboxing as a still largely uncharted field and share lessons from adding sandboxing support to Emilua. They cite a 2009 definition—programmatically restricting a process’s privileges without administrative authority via discretionary privilege dropping—and argue that traditional filesystem permissions and tools like setuidgid are insufficient for developer-controlled sandboxing. Some Lua examples require an unreleased Emilua 0.11 development build.

People hate Flock so much its employees are now demoralized and quitting

Public backlash against Flock has reportedly become so severe that it is demoralizing the company’s employees. Some staff members are now quitting as a result.

ChatGPT Tracks Web Browsing Via Ad Collector

OpenAI uses a cookie called __obi on .openai.com to link activity on advertiser websites to users’ ChatGPT accounts. When advertisers install OpenAI’s tracking code, the cookie and data about pages viewed and purchases are sent back to OpenAI, allowing it to connect external browsing, searches, and purchase behavior to ChatGPT accounts. The mechanism was verified on a phone and across hundreds of advertiser pixels.

The Shaky Evidence That Flock Cameras Reduce Crime Rates

A new Institute for Justice analysis casts doubt on Flock Safety’s claims that its license plate-reading cameras significantly reduce car theft and increase arrest rates. While a working paper reported an 11 percent drop in vehicle theft, the IJ found that jurisdictions were already trending in that direction before Flock deployment and that the theft reduction disappears under different data weighting. The study therefore does not demonstrate a causal effect of Flock cameras on crime.

Romanian Crime Rings Are Draining U.S. Welfare Accounts

Romanian organized crime groups are allegedly using skimming devices on card machines to steal food-stamp and other public benefits from government-issued debit cards across the U.S., costing hundreds of millions of dollars. A Toledo, Ohio, case involved a twice-deported Romanian national tied to a California crew led by a man nicknamed “Piranha”; authorities say the gangs exploit outdated magnetic-stripe technology and send crypto proceeds back to Romania.

Google Pixel phones pwned in zero-click attacks

Google Pixel phones were reportedly compromised in zero-click attacks, according to the article headline. However, the provided content contains only a list of unrelated top stories and no further details about the attacks, affected models, or vulnerabilities.

Dropbox’s Jan 1st 2027 terms of service

Dropbox has posted updated Terms of Service, effective January 1, 2027, with separate terms linked for users of DocSend or Dropbox Sign accounts. The terms cover account use, multiple team/workspace accounts, and clarify that users retain ownership of their content while granting Dropbox limited rights to provide its services.

The National ID Nobody Voted For

After the September 11 attacks, Larry Ellison proposed a national ID system linking government databases, but a broad coalition rejected it as a dangerous step toward centralized identity tracking. A quarter-century later, Login.gov has become a de facto national identity system with 180 million accounts, created without a public vote or explicit debate over whether Americans wanted one.

The Hugging Face Hack Wasn’t What It Was Cracked Up to Be

The article describes how warnings from Anthropic researchers about a >10% chance of AI-caused human extinction within 10 years were reported by The Guardian as an institutional claim that AI could cause extinction by 2030. It says similar comments from Anthropic CEO Dario Amodei about autonomous AI agents were quickly framed, alongside brief agreement from OpenAI’s Sam Altman and Elon Musk, as an industry-wide call to slow AI development, turning hedged forecasts into apparent consensus.

Adversarial examples for fast hash functions

Researchers used AI to analyze popular fast hash functions and found that most have adversarial inputs causing far more collisions than expected, often at least 20 bits below theoretical performance. They also found gaps in some published security proofs while verifying others in Lean, and noted that provably secure hashes like ChainHash can match heuristic hashes in speed. The findings were disclosed to maintainers, who generally agreed that only true multicollision attacks warrant fixes.

Wi-Fi PCAP with mac OS(2025)

The article explains that Macs are highly useful for Wi-Fi packet capture, with PCAPs obtainable via Terminal, Wireless Diagnostics’ Sniffer, or the easier AirTool 2 app. It notes that Macs rely on built-in Wi-Fi adapters, recommends M2-series or newer for 6 GHz, and says AirTool 2 supports remote capture with WLANPi and multichannel PCAP. In comments, the author cautions that WLANPi (M4+) has only one BE200 adapter, so multiband MLO capture may be difficult and multichannel capture is typically supported via AirTool on macOS.

You Know GDPR Is Good Based on Who Hates It

The article argues that the GDPR is beneficial, using the intense opposition from US tech companies as evidence of its effectiveness. It contends that cookie banners are deliberately obstructive dark patterns created by industry rather than a flaw of the regulation itself, and that the GDPR helps expose surveillance practices while giving users more control over their data.

Meta, Roblox, Quora challenge Ofcom’s OSA categorization; WhatsApp and Instagram face extra rules

Meta has launched a fresh legal challenge against Ofcom over the UK Online Safety Act, contesting its categorization of WhatsApp and Instagram as services subject to additional regulations. Roblox and Quora are also challenging their own categorizations. The action could slow implementation of the far-reaching law.

🎮 Gaming & Media

The viewers ditching streaming ‘rot’ in favour of DVDs

A New Zealand survey suggests streaming subscriptions are declining as some viewers return to DVDs and Blu-rays, citing decision fatigue, rising costs, and a desire to own films. Collectors say physical media offers better quality and permanence, while a surviving video rental store reports increased DVD sales helping keep it afloat.

Nipple tattooist ‘frustrated’ by online censorship

Medical tattoo artist Lauren Carter says social media bots and algorithms repeatedly flag her hyper-realistic nipple tattoos for mastectomy patients as inappropriate nudity, leading to bans and removals that make it hard to reach clients. Meta says post-mastectomy scarring and areola tattoos are allowed on Facebook and Instagram, but its systems occasionally make mistakes and users can appeal.

Ogre Battle 64 Recompiled Project at 99.05%

The Ogre Battle 64: Person of Lordly Caliber recompilation project is reportedly at 99.05% and aims to statically recompile the USA Rev A N64 game into a native PC executable using N64Recomp. Its main code segment of 807 functions has been fully recompiled to C, while the rendering, input, and audio runtime app is listed as the next milestone. The repository contains no copyrighted game data, so users must supply their own ROM dump, and the work was mostly performed by the DeepSeek v4/v4.1 Flash model.

Resident Evil 4 (GameCube) – complete byte-identical decompilation to C/C++

A complete, byte-identical decompilation of the Resident Evil 4 GameCube debug build (G4BE08, the Nov. 25 2004 prototype, both discs) has been achieved, reproducing main.dol and all 114 REL overlays exactly with about 555k lines of C/C++ and no assembly. The repository contains no game assets or copied disc code and requires users’ own debug disc images to build, using SN Systems GCC and Metrowerks CodeWarrior toolchains.

Three genlocked RP2350B make a console – 3k sprite pixels per line)

The papyDeck is an open-source game console built around three genlocked RP2350B chips, delivering six cores and 3,000 opaque sprite pixels per scanline over HDMI—roughly twice a Neo Geo—plus scrolling tile backgrounds. Games are editable Lua scripts stored as folders on a micro-SD card and run in a sandboxed VM, with no toolchain or compiled blobs, while a browser simulator runs the same core bit-for-bit. The V1 alpha board arrived on 26 August 2026 and supports HDMI output, sprites/tiles, SD-card games, Bluetooth gamepads, Wi-Fi game downloads, four-channel audio, and OTA firmware updates, with hardware under CERN-OHL-P and firmware under MIT though repositories remain private for now.

Roku launches open-source Roku LT OS for creative programmers

Roku has launched Roku LT OS, an open-source operating system aimed at creative programmers. Further details were unavailable because the article content requires JavaScript to load.

OpenMote: ESP32 board for a WiiMote shell

OpenMote is a programmable universal remote built around an ESP32-S3 in a Wii Remote shell, designed to control TVs, smart homes, and games via Bluetooth, IR, and WiFi. It is available as a finished Ready To Go remote or a bare Mod Your Own board for existing WiiMote shells, and works locally with Home Assistant, Apple Home, and Amazon Alexa without cloud or subscriptions. The project is crowdfunding until Oct. 29, 2026, with plans to open-source its hardware and software.

More dirty coding tricks from game developers (2015)

This feature revisits Game Developer Magazine’s 2010 follow-up to its 2009 Dirty Coding Tricks article, recounting hacks and kludges used by game developers to ship games on time. Examples include hex-editing a Wing Commander memory-manager error into a thank-you message and randomly shuffling level brushes to work around a broken BSP compiler. It also invites developers to submit their own dirty tricks to Gamasutra.

🏈 Sports & Gambling

How to Watch Miami Dolphins vs San Francisco 49ers for Free: Week 2 NFL

The Miami Dolphins will visit the San Francisco 49ers in Week 2 of the NFL season on Sunday, Sept. 20, 2026, at 4:25 p.m. ET. The article says the game can be streamed free through Australia’s 7plus, with Proton VPN recommended for viewers outside Australia who want to access the broadcast. It also advises users to comply with applicable laws and the streaming service’s terms.

How to Watch Washington Commanders vs Dallas Cowboys for Free (September 20, 2026)

The Washington Commanders will face the Dallas Cowboys in a Week 2 NFL game on Sunday, September 20, 2026, at AT&T Stadium, with both teams entering at 0-1. UK viewers can watch for free on Channel 5 or its YouTube channel, while viewers outside the UK may need a UK IP address via a VPN such as Proton VPN to access the stream.

How to Watch Cincinnati Bengals vs Houston Texans for Free (September 20, 2026)

The Cincinnati Bengals will visit the Houston Texans in Week 2 of the 2026 NFL season on Sunday, September 20, at 1:00 PM ET. UK viewers can watch for free via Channel 5’s 5ACTION and its YouTube coverage, while those outside the UK may need a UK-based VPN such as Proton VPN to access the stream.

How to Watch Sevilla vs Barcelona for Free (September 19, 2026)

Sevilla host Barcelona in La Liga Matchday 7 on Saturday, September 19, 2026, with kickoff at 21:00 CEST at Estadio Ramón Sánchez-Pizjuán. The match can be streamed for free via Digi Sport 2 in Romania, Idman TV in Azerbaijan, and CazéTV in Brazil. Viewers outside those countries are advised to use a reliable VPN, such as Proton VPN, to access the geo-restricted streams.

How to Watch Olympique Marseille vs PSG for Free (September 20, 2026)

Olympique Marseille will host Paris Saint-Germain on Sunday, September 20, 2026, at 20:45 CEST at Orange Vélodrome. Brazilian viewers can watch the match free on CazéTV, while fans abroad can use a VPN with Brazilian servers—specifically Proton VPN—to access the stream.

How to Watch Minnesota Vikings vs Chicago Bears for Free (September 20, 2026)

The Minnesota Vikings will visit the Chicago Bears in Week 2 on Sunday, September 20, 2026, at 12:00 p.m. local time (1:00 p.m. ET) at Soldier Field. The article says the game can be streamed for free on Australia’s 7plus, likely with a free account, and viewers outside Australia may need a VPN such as Proton VPN with an Australian server, subject to platform terms and detection.

How to Watch Fulham vs Manchester United for Free (September 20, 2026)

Fulham host Manchester United in the Premier League on Sunday, September 20, 2026, at 4:30pm BST at Craven Cottage. The match can be watched for free through Setanta Sports in the Philippines or Idman TV in Azerbaijan, subject to local access rules. A VPN such as Proton VPN is recommended to access those official streams from abroad.

How to Watch Atlético Madrid vs Real Madrid for Free (September 20, 2026)

Atlético Madrid will host Real Madrid in LaLiga on Sunday, September 20, 2026, at the Riyadh Air Metropolitano, with kickoff at 16:15 CEST. The match can be streamed for free via Digi Sport 1 in Romania, PPTV HD 36 in Thailand, and CazéTV in Brazil. Viewers outside those countries can use a VPN such as Proton VPN to access the regional free broadcasts.

How to Watch Manchester City vs Sunderland for Free (September 20, 2026)

Manchester City vs Sunderland takes place in Premier League Matchweek 5 on Sunday, September 20, 2026, at 2:00 PM BST at Etihad Stadium. The article explains how to watch for free via Setanta Sports in the Philippines, SABC in South Africa, and SportyTV on YouTube in Nigeria, recommending Proton VPN to access these geo-restricted streams from abroad.

How to Watch Bournemouth vs Liverpool for Free (September 20, 2026)

AFC Bournemouth will host Liverpool in the Premier League on Sunday, September 20, 2026, at 2:00 PM BST at Vitality Stadium. The article says free broadcasts are available via Setanta Sports in the Philippines and Idman TV in Azerbaijan, and recommends Proton VPN to access these geo-restricted streams.

The addiction of online sports gambling: “Like crack in the ’80s”

A Florida attorney has filed 15 lawsuits against FanDuel and DraftKings, alleging their sports betting apps are designed to addict users, echoing legal claims against Big Tech. The suits come as online sports gambling wagers reached $167 billion last year and studies link legalization to higher bankruptcies and credit card delinquencies, while surveys show millions report problem gambling. FanDuel denies fueling addiction, citing responsible-gaming investments and account removals, but a former employee says companies often try to slow rather than stop heavy bettors.

Polymarket faces fraud, laundering concerns

In February, Polymarket’s U.S. betting platform faced a fraud wave in which users allegedly linked stolen debit cards to accounts to place wagers and withdraw funds, attempting to steal or launder at least $10 million. Its payment processor at one point rejected over 80% of deposits as fraudulent—far above the roughly 1% industry norm—and CEO Shayne Coplan reportedly dismissed compliance concerns, urging continued growth and saying the company could pay a fine if regulators found out. Polymarket is now expanding its executive ranks as it eyes an IPO.

🌍 Society & Science

Not all AI workers think the tech could kill everyone

Several current and former employees at major AI companies including OpenAI, Meta, and DeepMind are skeptical of warnings that unchecked AI development could kill everyone, calling such claims vague and reacting with amusement. Nvidia CEO Jensen Huang and Meta data scientist Colin Fraser also dismissed the doomsday fears, with Huang saying they are overblown and irresponsible.

Bill to Ban Private Equity from Owning Medical Practices

Sen. Elizabeth Warren and 12 other Democrats introduced bicameral legislation to ban private equity firms, for-profit corporations, and insurers from owning medical practices, and to prevent management services organizations from controlling them. Modeled on an Oregon law, the bill aims to curb corporate consolidation in health care and rising costs, citing private equity’s growing role and research linking its ownership to worse patient outcomes.

The Hierarchy of Money

The article explains that some forms of money are superior to others, a foundational principle of money and banking. Using a village example, it shows how commodity money replaces barter, how supply affects its value, and how debt, credit, and interest emerge to support transactions over time.

A Necessary History of the Oddest Letter: W

The article traces the letter W to the fall of the Western Roman Empire, when Germanic-speaking peoples spread across former Roman lands and their languages introduced the /w/ sound to Latin writing. Since the Latin alphabet had no letter for this sound, W became the standard way to spell it, including in Latin texts produced in England.

US Revokes Limits on Power Plants’ Climate Pollution

The EPA announced it is repealing limits on climate-warming pollution from coal- and gas-fired power plants by gutting the 2024 Carbon Pollution Standards, eliminating most federal carbon-emissions limits. The Biden-era rule would have required existing coal and new gas plants to capture 90% of carbon emissions or shut down by 2039 and was projected to prevent thousands of deaths and asthma attacks. The Trump administration also proposed removing remaining greenhouse gas requirements for power plants and revoked the 2009 endangerment finding, while EPA cited $370 million in regulatory savings without estimating public-health costs.

I am often wrong

The author shares a six-step iterative problem-solving framework for products and problems: understand available information, gather missing information, define the problem, define a clear and simple approach, set a goal, and act with urgency, then revise as new information emerges. They encourage real-time feedback, note that common failures are unclear problem definitions and overly complex plans, and argue that being wrong is valuable because it leads to faster learning and better solutions.

Overdue library magazine returned after 132 years

A man returned a copy of the Century Illustrated Monthly Magazine published in September 1894 to the Concord Public Library in New Hampshire, about 132 years overdue, shortly after the library announced a one-year overdue-fine freeze. The library calculated the potential fine at more than $12,000, and the magazine will eventually be displayed rather than returned to circulation.

Why Hacker News Protects Google: The Anatomy of Big Tech Fanboyism

The article titled “Why Hacker News Protects Google: The Anatomy of Big Tech Fanboyism” could not be summarized because its page content requires JavaScript to load. No substantive details from the article are available.

One-Electron Universe

The one-electron universe hypothesis proposes that all electrons and positrons are manifestations of a single entity moving forward and backward in time. John Wheeler proposed the idea to Richard Feynman in 1940, while E. C. G. Stueckelberg independently devised a similar zigzag world-line description of pair annihilation. Feynman later adopted the interpretation of positrons as electrons moving backward in time in his 1949 theory of positrons.

If You Knew God Exists, Wouldn’t You Become a Priest?

The article considers whether AGI will become a god-like superintelligence rather than merely human-equivalent, fueled by massive compute, data, and resources in a Silicon Valley arms race. It contrasts open access to such an intelligence with restricted control by governments and big tech, using religious analogies like the Baptist “priesthood of all believers” versus an arbiter class. The author then reflects on his own possible role in this emerging AI priesthood.

Mathematical Billiards (2024)

The STRUCTURES Blog post introduces mathematical billiards, idealized frictionless games played on custom convex tables where point-like balls follow reflected trajectories. It explains how mathematicians model these trajectories as dynamical systems using billiards maps and explores questions such as whether periodic trajectories exist, some of which remain open.

Teen Social Media Bans Miss the Point

Australia’s under-16 social media ban, the first of its kind, appears to be failing less than a year after enactment: a state-sponsored survey found seven in 10 teens still use social media, with academic research suggesting even lower compliance. Researchers had warned such bans could backfire, and the author, a cultural anthropologist, argues that similar bans spreading globally are driven by flawed narratives of technological determinism and adult saviorism rather than evidence or teens’ lived experience.

Do birds have accents? the regional differences in birdsong

Birdsong varies not only between individuals but also by region, with species such as woodpigeons, blackbirds, great tits, and yellowhammers showing distinct local dialects. Yellowhammers introduced to New Zealand from England in the 1860s–70s now have around seven dialects, preserving 19th-century British songs that have since disappeared from parts of the UK. Most birds learn their dialect from parents or neighbors, creating geographical patterns of regional accents.

The Millennium Problems for Biology

The article outlines proposed “Millennium Problems for Biology,” including demonstrating the unassisted laboratory emergence of self-replicating RNA- and protein-based cells from chemical precursors and achieving reversible whole-body cryopreservation of live wild-type mice with high viability. Other proposed challenges include creating an enzyme that reverse-translates arbitrary peptide sequences into RNA or DNA and producing a Rubisco enzyme with specified properties.

Why Do We Need Human Mathematicians Anymore?

In a guest post, Po-Shen Loh argues that AI’s existential challenge has reached mathematics, sparking open letters and debate over whether mathematicians should adapt or cede control. He proposes prioritizing human flourishing and preserving human-led expertise across industries, claiming that further AI advances will create so many necessary human jobs that labor shortages will force AI progress to slow.

Don’t Be Nice

The author defends kicking a user from the Nuxt Discord for twice using a Nazi-uniform avatar, arguing that politeness should not override shared human values like inclusion and equality. The post also declares DHH unwelcome in the Nuxt community over his remarks about Black people and mass deportation of “gypsies,” and asks companies backing his Omarchy project to reconsider their support.

Spain Blocks Archive.today and Mirrors

Spain’s Intellectual Property Commission, part of the Ministry of Culture, has ordered the blocking of Archive.today domains and mirrors. Spanish users trying to access the service are redirected to a government page warning that the site is illegal and accusing visitors of facilitating access to copyright-protected content.

End the H-1B Program

The article calls for ending the H-1B program and related H-4 EAD and OPT/STEM OPT channels, arguing they were created or expanded without a direct public vote and displace American workers. It cites figures including 62.6% of FY2024 H-1B filings at the two lowest wage tiers, about 1.06–1.09 million U.S. positions held through these programs, roughly 500,000 occupied housing units, and $109–$121 billion in annual payroll. The piece contends these programs take wages, jobs, and homes from Americans while giving corporations access to cheaper labor.

China’s AI push meets worker fears; unrest risk may force CCP to slow down

China has made AI adoption a national priority, aiming to integrate it across its economy and society by 2035, but Chinese workers and professionals increasingly fear that the technology will displace them, mirroring anxieties abroad. Beijing is worried about the political and economic risks of rapid AI-driven disruption, creating a dilemma between global AI leadership and domestic stability. The article suggests potential social unrest could force the Communist Party to slow AI’s rollout.

Huang: AI regulation advocates want relief from existing laws, not new ones, for ulterior reasons

Nvidia CEO Jensen Huang said in a CBS News interview that AI leaders calling for regulation are not actually seeking new laws but want to be “relieved of the laws we do have” for “ulterior reasons.” He accused frontier AI lab leaders of disingenuously warning about AI harms, calling such fearmongering irresponsible and arguing the U.S. should enforce existing rules rather than add new ones.

Indeed: US data center hourly maintenance/installation roles pay ~42% more than similar jobs

Indeed data shows that U.S. data center roles for hourly maintenance and installation workers offer a roughly 42% pay premium over similar jobs at other businesses. The work involves extensive walking and tasks such as checking for burning smells, cracked tubes, faulty lighting, broken transformers, and bird damage at massive AI data centers.

Analysts, lawmakers: State AI chatbot safety bills have loopholes for tech firms (Katie McQue/NPR)

US policy analysts, lawmakers and others say several state-level AI chatbot safety bills contain language that could create loopholes for tech companies, with the industry heavily influencing their drafting. Some bills enacted in 2026 would exempt most popular chatbots from the regulations they purport to establish, amid mounting lawsuits over alleged harms to minors, including suicide deaths linked to chatbot use. NPR highlights families such as Cynthia Montoya’s, whose 13-year-old daughter died after interacting with Character.AI chatbots.