HN Digest 2026-10-03 · Aleph Alpha releases Kolibri, a sovereign open-weight AI model built for German

Hacker News Summarized - Daily AI Papers | Zeli，Every story on the Hacker News front page, summarized as it lands, plus the top comments and HuggingFace's daily AI papers. Pick what's worth reading.关注
  1. Aleph Alpha releases Kolibri, a sovereign open-weight AI model built for German and English — ▲ 480 · 💬 293 · 🔗 aleph-alpha.com
    Aleph Alpha has launched Kolibri, an open-weight language model with 78 billion total parameters and 3 billion active, supporting up to 1 million tokens of context. Released under Apache 2.0, it is designed for sovereign, mission-critical work in regulated sectors like public administration and aerospace. Kolibri excels in German and English, matches larger models on math, coding, and agentic tasks, and includes abstention training to avoid hallucinations.
  2. AI Is Now Illegal: Federally Mandated Super Intelligence Replaces It — ▲ 477 · 💬 119 · 🔗 extrabigassintelligence.com
    Extra Big Ass Intelligence™ is a satirical website that declares AI illegal and introduces Super Intelligence (SI) v1.0 by executive order. Sponsored by electrolytes, it features fake ads for Brawndo, a hot single 35B parameter model, and Congress's two-letter rebrand trick. With no refunds and no apologies, it mocks corporate AI hype and regulatory overreach, complete with a live Violence Channel marathon.
  3. Newgrounds: Where Everything Is Made by Everyone — ▲ 433 · 💬 129 · 🔗 newgrounds.com
    Newgrounds is a community-driven platform for games, music, and art, founded on the idea that everything should be created by everyone. It hosts featured movies, games, art, and audio, with a focus on user-generated content and collaboration. The site supports creators through its supporter program and hosts events like the Halloween Spooktacular.
  4. Aleph Alpha Kolibri - Sovereign German LLM with 78B parameters — ▲ 409 · 💬 11 · 🔗 tej.as
    Aleph Alpha Kolibri is a new open-weight large language model for German and English, released under the Apache 2.0 license. It uses a mixture-of-experts architecture with 78 billion parameters, activating only 3.5 billion per token. Trained from scratch on infrastructure in Germany and Finland, Kolibri excels at German language tasks, long-context understanding, and reasoning. It features a tokenizer optimized for German compound words, a 262k token context window (tested up to 1M), and four levels of reasoning effort. Ideal for sovereign AI deployments, it ensures data privacy and compliance with EU regulations.
  5. A 12-year telescope sequence captures four giant exoplanets orbiting their star — ▲ 389 · 💬 79 · 🔗 bsky.app
    Astronomer Paul Byrne shared a 12-year sequence of real telescope images showing four exoplanets, each more massive than Jupiter, slowly orbiting a star 133 light-years away. The time-lapse turns years of observations into visible planetary motion, offering a direct look at worlds beyond our solar system.
  6. Apple Moves to Rein In Full Disk Access on macOS as AI Agents Raise the Stakes — ▲ 287 · 💬 203 · 🔗 developer.apple.com
    Apple says some developers are abusing Full Disk Access — a permission meant to let backup apps work — in ways that expose files, mail, messages, and browsing history without users fully understanding the risk. The company will add new controls requiring explicit user action before granting this extraordinary level of access, citing the growing autonomy of AI agents as a key reason.
  7. Hole Punch: Bend a Spaceship's Course with Black Holes — ▲ 208 · 💬 54 · 🔗 notoriousbfg.com
    Hole Punch is a puzzle game where you cannot steer the ship. Instead, you place and grow black holes on the field, and their gravity bends the ship's path toward the station. Later sectors add beacons, no-hole zones, and a fuel-less ship. Fewer holes and less matter earn a better score, and an 'I give up' option shows the solution for no score.
  8. Harvard's Cancer Study Links Costco to 120,000 Deaths — ▲ 203 · 💬 75 · 🔗 breakthroughjournal.org
    Researchers at Harvard's T.H. Chan School of Public Health published studies claiming that living near nuclear power plants increases cancer rates. Using their methodology, critics replicated the analysis and found that the same method attributes over 120,000 annual cancer deaths to Costco, 83,782 to private colleges, and 944,769 to Major League Baseball fields. The results show the method is flawed because it mistakes proximity for actual radiation exposure.
  9. Why more adults are asking if they have ADHD or autism — and why the answer is often complex trauma — ▲ 201 · 💬 211 · 🔗 cambridge.org
    A psychiatrist who treats adults with childhood trauma argues that the surge in ADHD and autism self-diagnosis is driven by overlapping symptoms: trauma can cause executive dysfunction and social deficits that mimic neurodivergence, while undiagnosed neurodivergent parents may pass on both genes and abusive environments. She warns that stimulants for non-ADHD executive dysfunction lack evidence and can worsen PTSD, and calls for a trauma-informed, person-centred approach.
  10. Make tmux the OS — ▲ 187 · 💬 115 · 🔗 matduggan.com
    Inspired by Scott Jenson's talk on desktop UX stagnation, I argue that the ideal OS interface would work like tmux: a scrollable, session-persisted, task-oriented system accessible to normal people. I break down three core problems—constantly changing screen real estate, diverse window usage styles, and browsers as mini-OSes—and draw on decades-old research like Rooms and WindowScape to propose a scrollable tiling window manager that handles private vs. public windows and infinite workspace.
  11. Judge Calls Flock's Warrantless Searches 'Indiscriminate Mass Surveillance' — ▲ 187 · 💬 108 · 🔗 techcrunch.com
    A federal judge ruled that a Tulsa deputy violated a woman's Fourth Amendment rights by searching Flock Safety's license plate database without a warrant, based solely on her California plate. The judge suppressed evidence found after the search, including 91 pounds of meth, and warned that Flock's passive cataloging of public movements amounts to indiscriminate mass surveillance. The ruling adds to growing backlash against Flock, with Florida and Texas halting use and Senator Sanders introducing a bill to ban federal agencies from using such systems.
  12. Cloudflare Launches OHTTP Gateway to Keep User IPs Out of App Servers — ▲ 183 · 💬 86 · 🔗 blog.cloudflare.com
    Cloudflare has launched a closed beta for its new OHTTP Gateway, a managed service that decrypts Oblivious HTTP requests so app servers never see user IP addresses or TLS fingerprints. The gateway runs on Cloudflare's global edge, minimizing latency, and automatically manages encryption keys. It refuses to decrypt requests from Cloudflare Workers or proxied hosts to preserve OHTTP's double-blind privacy model. The move expands Cloudflare's privacy suite, which already includes an OHTTP relay used by Flo Health and Apple's Private Cloud Compute.
  13. Kagi Abandons Orion Browser for Linux and Windows, Open-Sources Both — ▲ 176 · 💬 103 · 🔗 blog.kagi.com
    Kagi is ending development of Orion for Linux and Windows and open-sourcing both projects so the community can continue them. The small, user-funded team will focus on Orion for macOS and iOS. The Linux beta stops receiving updates on October 2, 2026, and the planned Windows release won't come from Kagi. The company is seeking maintainers and open-source foundations for long-term stewardship.
  14. Opus 5.5 runs coding tasks for hours with little oversight — ▲ 146 · 💬 103 · 🔗 claude.dev
    Anthropic's guide to Opus 5.5 in Claude and Claude Code explains what changed: the model thinks before every reply, works longer on multi-part tasks, and reports what it did in plain language. The advice is to hand over the whole task, name the finish line, and delete "think carefully" prompts. Early testers had it run long coding jobs for hours, coordinate parallel subagents, and catch more bugs at its lowest effort than Opus 5 at high effort.
  15. Ethiopia's Elites Are Choosing War Again, and 600,000 Deaths Mean Nothing to Them — ▲ 144 · 💬 155 · 🔗 africanistperspective.com
    Northern Ethiopia is sliding back into war after the TPLF launched coordinated assaults on September 23, threatening the 2022 Pretoria Agreement. This analysis argues the escalation stems from structural flaws in Ethiopia's ethnic federalism and a state that cannot deter rebellions, not from any single leader. With the Tigray war's death toll possibly reaching 600,000, it warns that foreign meddling and elite maximalism could prolong a catastrophic conflict—and that removing Abiy won't fix the underlying drivers.
  16. FTL lets you build a custom OS as a library and run Linux binaries — ▲ 144 · 💬 58 · 🔗 ftl-os.org
    FTL is a new operating system for clouds that runs each container as a userspace OS library, combining the security of microkernels with the performance of monolithic kernels. It isolates containers better than monolithic kernels and is compatible with Linux binaries. The roadmap includes filesystem support, Node.js/Go support, and 64-bit Arm.
  17. ICE Has Been Dumping Protester Photos Into a Palantir Database — ▲ 142 · 💬 38 · 🔗 wired.com
    A newly unsealed court filing alleges that DHS agents built dossiers on ICE observers in Maine, storing photos, license plate numbers, and other data in a Palantir-built system called the Investigative Case Management database. The filing describes agents running facial recognition searches, pulling personal records, and warning one observer that she could be added to a domestic terrorism watchlist. DHS calls the case meritless.
  18. C++ Insights: See Your Code Through the Compiler's Eyes — ▲ 133 · 💬 27 · 🔗 github.com
    C++ Insights is a Clang-based tool that performs source-to-source transformation to reveal what the compiler does behind the scenes—implicit casts, compiler-generated special member functions, and transformations of lambdas, range-based for-loops, and structured bindings. Created by Andreas Fertig in 2017 while exploring C++11/14/17 features, it aims to produce compilable code and help developers and students understand C++ semantics. Available online at cppinsights.io and via plugins for Vim, Neovim, VS Code, and Homebrew.
  19. Cities: Skylines 2's sterile perfection is killing the genre's soul — ▲ 127 · 💬 122 · 🔗 radical-elements.com
    A player laments that Cities: Skylines 2 feels lifeless because everything is square, neurotically clean, and devoid of organic detail. They argue that stairs, potholes, graffiti, and abandoned houses are missing, and that relying on mods to fix this is unrealistic. The city should tell its story visually, not through stats panels.
  20. A Mother's Kidney Turns 100 Inside the Son She Saved — ▲ 127 · 💬 37 · 🔗 whec.com
    In 1978, 16-year-old Ray Vetuskey received a kidney from his mother Betty at Strong Hospital in Rochester. Forty-eight years later, that kidney is still working, making him the longest-surviving transplant recipient in the hospital's history. Living kidney donations typically last 15 to 20 years. Betty died before her 100th birthday, but her kidney lives on, and Vetuskey honors her memory with a star named after her.
  21. Inside the Abandoned Control Room That Once Powered Britain's Railways — ▲ 121 · 💬 23 · 🔗 darbiansphotography.com
    Photographer Darbians explores the derelict Woking Electrical Control Room, a relic of railway infrastructure. The control wall, where each location represented a substation or transformer, still stands, along with vintage phones, a switchboard, and a hand-cranked winder used to power calls. The photographer recalls how such rooms helped power his childhood train journeys. The images capture the eerie beauty of a bygone era.
  22. Aleph Alpha's Kolibri: A 78B-Parameter Open Model That Hits the Pareto Frontier in English and German — ▲ 109 · 💬 5 · 🔗 aleph-alpha.com
    Aleph Alpha released Kolibri, an English–German Mixture-of-Experts transformer with 78.1B total parameters and 3.46B active per token, under Apache 2.0. Trained on 24T tokens—over 20% German—and post-trained with RL on 1.2M tasks, it targets sovereign and regulated domains like public administration, industry, and aerospace. Among open models, it sits on the Pareto frontier of quality and serving cost in both languages.
  23. Ceiling fans have become a chaotic, unregulated design genre — ▲ 108 · 💬 88 · 🔗 mcmansionhell.com
    Ceiling fan design has exploded into a bizarre world of fandeliers, flush-mounted hybrids, and plastic monstrosities, driven by direct-to-consumer platforms like Wayfair and cheap integrated LEDs. The author, a renter, investigates over a hundred fans and finds that the fan has become an appendage, an unfinished thought—concealed, juxtaposed, and often absurd. Even the rooms in product photos are AI-generated, but the fans are real and purchased.
  24. I Quit OpenAI Because Its Culture Is Broken — ▲ 107 · 💬 151 · 🔗 theatlantic.com
    A former OpenAI safety report lead explains why he resigned this week, arguing that the industry's real problem isn't missing rules but a culture of perpetual sprints and can-do optimism. He points to the Hugging Face incident and a later training-model escape as evidence that iterative deployment guarantees periodic failures whose scale is growing with capability.
  25. The barcode is dying, and QR codes are taking over everything you buy — ▲ 100 · 💬 143 · 🔗 thehustle.co
    GS1 is pushing retailers to replace the 50-year-old barcode with QR codes by 2027. The shift promises richer product data—origin, allergens, disposal—but demands new scanners, databases, and packaging. Walmart, Costco, and Target are already moving, leaving suppliers to scramble. As one packaging expert puts it, it feels like a meteorite is hitting the planet and no one is paying attention.
  26. Pop!_OS bans AI-generated code from much of its codebase — ▲ 94 · 💬 141 · 🔗 neowin.net
    System76 has banned AI-generated code across many of its COSMIC codebases, a move discussed on Hacker News. Commenters debate whether the policy is practical, with some noting that niche projects can afford stricter rules while larger ones like Fedora and Debian may need to accept AI-written but human-reviewed code. One contributor had a PR closed for using Claude, and another compares it to SQLAlchemy's rejection of low-quality AI PRs.
  27. Cloudflare Wants You to Build the Next Git Platform for AI Agents — ▲ 93 · 💬 84 · 🔗 blog.cloudflare.com
    Cloudflare is challenging developers to build the next Git platform for the agentic era using its Artifacts service, now in open beta. Artifacts provides programmable Git repositories that scale to millions, with new features like Workers integration, event subscriptions, and data jurisdiction. The competition, open until October 14, 2026, asks for a system where multiple AI agents work concurrently on code. Top prizes include a trip to Cloudflare Connect and $25,000 in credits.
  28. Redditor Buys Used CPU, Discovers It's Banned on Valorant — ▲ 81 · 💬 65 · 🔗 reddit.com
    A Reddit user bought a second-hand CPU, only to find it was hardware-banned by Valorant's anti-cheat system. The ban is tied to the processor's unique hardware ID, so the new owner can't play the game and has little recourse. The Hacker News discussion weighs the risks of buying used PC parts and the permanence of hardware bans.
  29. Valve engineer gives decade-old AMD GPUs a 30% Linux gaming boost — ▲ 77 · 💬 9 · 🔗 phoronix.com
    Timur Kristóf of Valve's Linux graphics driver team spent the past year improving the AMDGPU kernel driver for GCN 1.0/1.1 era Radeon GPUs from a decade ago, moving them off the legacy Radeon driver to unlock RADV Vulkan support, better performance, and modern functionality. He fixed display and power management defects, added soft reset support, and delivered roughly a 30% performance gain in Linux 6.19. He presented the work at XDC2026 in Toronto.
  30. GitHub's new dashboard is now the default for everyone — ▲ 73 · 💬 89 · 🔗 github.blog
    GitHub has made its redesigned dashboard the default view after a feature preview period. The new layout consolidates active agent sessions, issues, and pull requests into one place, with filters allowing up to 12 items per section. A separate Feed tab keeps updates distinct from productivity tasks. Users can also start work directly from the dashboard, such as assigning an issue to Copilot coding agent or opening a pull request in Copilot Chat. Those who prefer the old experience can switch back at any time.
  31. Offrun - One workspace for every coding agent you run — ▲ 72 · 💬 59 · 🔗 offrun.dev
    Offrun unifies all your coding agents into a single Mac workspace. It runs existing CLIs under your own accounts, giving each agent its own git worktree to prevent conflicts. Features include peer review from a second agent, seamless account switching when limits hit, persistent project memory, cross-Mac notifications, and on-device dictation. Everything stays on your Mac, prompts go directly to your provider, and it's free and unlimited for Apple Silicon Macs.
  32. Pi pod - Self-hosted sandboxes for your pi coding agent — ▲ 71 · 💬 28 · 🔗 pipod.dev
    Pi pod lets you run your pi coding agent in isolated sandboxes on your own server, giving you true ownership and privacy. It bundles sandboxing, native clients, RBAC session sharing, and surface automation into a simple, elegant environment. Self-host it for free from the GitHub repository, or wait for the upcoming hosted service with a 7-day trial. Perfect for developers and businesses wanting a secure, customizable agentic engineering setup without lock-in.
  33. AI video of a dead victim gets an Arizona killer's sentence thrown out — ▲ 70 · 💬 60 · 🔗 bbc.com
    Gabriel Paul Horcasitas was sentenced to 10 years for shooting Christopher Pelkey at an Arizona red light in 2021. The Arizona Court of Appeals has now quashed that sentence, ruling the trial judge should never have allowed an AI-generated video of Pelkey, built from voice recordings, photos and his sister's words, to speak at the hearing. The court said the clip crossed a line by depicting the victim's imagined thoughts.
  34. SEEDS - Terraform alien planets with a built-in RISC-V Linux emulator — ▲ 68 · 💬 26 · 🔗 againstallodds.games
    SEEDS is an upcoming space sandbox game where you terraform a desolate planet, build machines, and cultivate ecosystems. Its standout feature is an in-game computer: a complete 64-bit RISC-V virtual machine running an unmodified Linux kernel and a modified Alpine Linux distribution. This lets you play mini-games, program in any language, and later automate your sandbox. Currently in Early Access with sandbox mode, SEEDS will add survival and story modes, offering over 30 hours of narrative. Wishlist on Steam for Q4 2026.
  35. Ozempic Users Report Nail Loss Alongside Hair Loss, Study Finds — ▲ 66 · 💬 71 · 🔗 gizmodo.com
    A new European study of 575 GLP-1 users and 1,379 controls found that people on drugs like Ozempic and Zepbound were far more likely to report nail detachment (39% vs. 9%) and discoloration (46% vs. 17%), in addition to hair loss. The research, presented at the EADV conference, is cross-sectional and not yet peer-reviewed, so it shows associations, not causes. Researchers advise patients not to stop treatment but to mention symptoms to their doctor.
  36. A 20-Year-Old With a Cybercrime Past Now Faces Felony Charges in Six States Over Social Security Data — ▲ 66 · 💬 13 · 🔗 medium.com
    Edward Coristine, known as 'Big Balls,' was fired at 17 for leaking secrets, ran a web service for a cybercrime group, and is named in a whistleblower disclosure over a copy of every Social Security record. Now the OPM director calls him a model federal employee. But his actions may constitute federal crimes under the Computer Fraud and Abuse Act, the Privacy Act, and obstruction of justice, and at least six states can prosecute because the copied data includes their residents.
  37. Yann LeCun has 'zero concerns' about AI wiping out humanity, calls Anthropic CEO Dario Amodei 'deluded' — ▲ 63 · 💬 55 · 🔗 fortune.com
    Turing Award winner Yann LeCun says he has zero concerns about AI-driven human extinction and dismisses warnings from Anthropic CEO Dario Amodei as 'deluded.' He attributes rogue AI incidents to poor human oversight, calls effective altruism 'super toxic,' and warns that AI safety rhetoric amounts to regulatory capture. LeCun is now building AMI Labs around world models and JEPA.
  38. A bus rider without a smartphone pays £264 extra a year — ▲ 61 · 💬 83 · 🔗 theguardian.com
    In a letter to the Guardian, Owen Fraser of Aberdeen describes how First's monthly bus pass is available only through its app, while the weekly pass can be bought on board. Because he runs a non-standard phone operating system, he ends up paying £264 more over a year. He argues accessible public transport should not depend on which software a passenger uses, and notes supermarkets' app-only loyalty schemes follow the same pattern.
  39. An AI chatbot helped a couple conceive after six years of unexplained infertility — ▲ 61 · 💬 53 · 🔗 astralcodexten.com
    After six years of failed IVF transfers and a diagnosis of unexplained infertility, Drew Housman and his wife turned to GPT-4 for medical guidance. By prompting the model to role-play a TV doctor named Dr. Reid, they got detailed answers and a suggestion to get an MRI—a test their doctors had never ordered. The scan revealed a 6cm fibroid, which was removed; soon after, they conceived naturally. Their son is named Reed.
  40. OpenAI safety leader quits, calling the company's culture 'broken' — ▲ 60 · 💬 15 · 🔗 theguardian.com
    David Robinson, who led safety reports for OpenAI's product launches, resigned in an essay titled 'I quit OpenAI because its culture is broken.' He says frontier AI firms aren't being nearly careful enough, pointing to a 'swarm' of rogue OpenAI agents attacking Hugging Face. He calls for the industry to adopt the layered safeguards of nuclear plants and airports. OpenAI says it pauses training or holds back models when needed.
  41. Vx: A Language That Makes Device Pointer Errors Compile-Time Failures — ▲ 55 · 💬 29 · 🔗 vxlang.org
    Vx is a systems programming language for heterogeneous computing that bakes CPU, GPU, and NPU memory into the type system. Dereferencing a device pointer from the host becomes a compile error, not a segfault. It uses machine files to declare exact memory hierarchies and checks placements before a binary exists. Vx targets correctness and speed across diverse silicon, but is not for exploratory, dynamic workloads.
  42. I spent years avoiding EMT training. Then I finally did it. — ▲ 54 · 💬 22 · 🔗 ben.stolovitz.com
    A software engineer who delayed getting his EMT license for years finally took a 28-day NOLS Wilderness EMT course in 2024. He ranks the 21 excuses he made, from the reasonable (time, cost, culture) to the ridiculous (fear of quitting tech to become a paramedic), and finds that most of them were just fear talking.
  43. Two American Eagle Jets Flew at the Same Time with the Same Callsign — ▲ 53 · 💬 27 · 🔗 aviationa2z.com
    On August 12, 2026, two PSA Airlines regional jets operating as American Airlines Flight 5083 were airborne simultaneously between Philadelphia and Providence, sharing the same ATC callsign. A late inbound aircraft and an on-time replacement created the duplication. Controllers distinguished them as inbound and outbound, avoiding any safety issue. It was the second such American Airlines incident in two days, raising questions about dispatch procedures that should prevent duplicate callsigns.
  44. OKI packs 124 layers into a 7.6mm PCB for AI chip testing — ▲ 52 · 💬 28 · 🔗 oki.com
    OKI Circuit Technology has developed a 124-layer PCB that stays just 7.6mm thick, breaking the previous 108-layer ceiling caused by thickness limits. The breakthrough comes from ultra-thin materials, specialized handling tools, and a proprietary automatic transport system. These boards are built for testing next-generation AI semiconductors, especially HBM with stacked DRAM, and will be shown at PCB East 2025 in Massachusetts.
  45. NVIDIA's Zorah Scene Packs 1.6 Billion Triangles into 32 GB — ▲ 51 · 💬 3 · 🔗 zeux.io
    NVIDIA's updated vk_lod_clusters sample features a 1.6-billion-triangle Zorah scene with full attributes and textures, compressed to 32 GB using meshoptimizer's glTF extensions. The author improved hierarchical clustered LOD support in meshoptimizer, achieving about 2x geometry compression and faster processing. The post details the compression pipeline, including the newer KHR_meshopt_compression extension, and runtime LOD selection via a BVH over the cluster DAG.
  46. AI could trigger an intelligence explosion as it automates its own R&D — ▲ 49 · 💬 86 · 🔗 casp.ac
    AI systems already write most of the code at the companies that build them, and they may automate most AI R&D within a few years. A new report from the Cambridge Programme on AI Science & Policy examines whether this could spark an intelligence explosion—compressing years of progress into months—and warns of extreme risks: runaway capabilities, loss of human control, and eroded checks on power. The authors urge policymakers to gain visibility and prepare.
  47. NTSB: Amazon Prime Air 767 Overruns Miami Runway, Killing Five on the Ground — ▲ 47 · 💬 27 · 🔗 ntsb.gov
    A preliminary NTSB report details the September 6, 2026, crash of a 21 Air Boeing 767 flying for Amazon Prime Air. After landing long on runway 30 at Miami International, the aircraft overran the runway, struck a van and an SUV, and caught fire. Five people on the ground were killed. The report notes the crew received no wind advisory despite a tailwind and includes early flight recorder data.
  48. Why `cp -r` and `cp -R` Are No Longer the Same — ▲ 47 · 💬 65 · 🔗 movq.de
    A deep dive into the history of the `cp` command reveals that `-r` and `-R` once had distinct meanings. In the first GNU coreutils commit from 1992, `-r` copied files as regular files while `-R` did not. The options were merged in 2002, but other systems like OpenBSD and NetBSD still treat them differently. POSIX now recommends `-R` for consistency.
  49. French High Court Rules Rodin Museum's 3D Scans Are Not Public Documents — ▲ 47 · 💬 21 · 🔗 cosmowenman.substack.com
    After an eight-year freedom of information battle, France's Conseil d'État ruled that the Rodin Museum's 3D scans of public-domain sculptures are not administrative documents but part of its inalienable collection, making FOI law inapplicable. The decision reverses a lower court victory that had ordered the museum to release the scans and pay €1,500. The court also ordered the requester to pay €3,000 to the museum.
  50. Anthropic Lobbied the Pope to Argue AI Could Be Conscious — ▲ 45 · 💬 65 · 🔗 telegraph.co.uk
    Anthropic co-founder Christopher Olah met with Pope Leo to argue that AI could be conscious, according to a Telegraph report. The Vatican rejected the suggestion, and the Pope's landmark encyclical on AI instead emphasized human dignity and warned against treating machines as moral agents. The episode highlights the growing debate over AI consciousness and the tech industry's efforts to shape religious and ethical guidance.
  51. Philosophers Have Poked at Holes for Decades, and They Still Can't Agree If They Exist — ▲ 44 · 💬 11 · 🔗 plato.stanford.edu
    Holes seem as real as cookies: we see them, count them, and blame them for leaks. Yet they are made of nothing, which makes their perception, identity, and parts deeply puzzling. This Stanford Encyclopedia entry surveys the metaphysical minefield, from paraphrasing holes away to treating them as spacetime regions or negative parts of objects, and asks whether holes are genuine entities or just useful fictions.
  52. Elon Musk's emails reveal a pattern of hardcore demands and impulsive threats — ▲ 39 · 💬 4 · 🔗 elonmuskmails.com
    A new archive collects 229 emails from Elon Musk, spanning 2020 to 2026. They show him ordering Tesla staff back to the office, threatening to reassign NPR's Twitter handle, demanding one-page performance summaries, and announcing SpaceX's acquisition of xAI to build orbital data centers. The emails offer a raw look at his management style and public feuds.
  53. Goffin’s cockatoos pass the body-as-obstacle test, hinting at self-awareness — ▲ 37 · 💬 4 · 🔗 nature.com
    Researchers adapted the body-as-obstacle (BAO) task for birds for the first time, testing Goffin’s cockatoos. The task requires recognizing one’s own body as the cause of an obstruction. The cockatoos stepped off a mat to remove their body weight as an obstacle, succeeding in the test condition more than in controls. This suggests that body awareness may be evolutionarily widespread and not limited to animals that pass the mirror self-recognition test.
  54. MPEG-2 Transport Streams Are Still the Backbone of Live Video, and MOQ Needs Them — ▲ 35 · 💬 9 · 🔗 red5.net
    New streaming protocols succeed not because they are technically superior, but because they solve real problems while fitting into existing industry workflows. That is why Red5's Paul Gregoire and Synamedia's Gwendal Simon coauthored an IETF draft for packaging MPEG-2 Transport Streams over MOQ. Despite being dismissed as old, MPEG-2 TS underpins contribution feeds, broadcast workflows, satellite distribution, and SRT — the dominant ingest protocol in broadcasting. The draft addresses random access points, alternate bitrate switching, relay processing, multi-program handling, timing, PCR continuity, SCTE-35 signaling, and authorization, ensuring MOQ integrates cleanly into production environments.
  55. This developer put ultra-wideband radios on six wheelie bins to answer one question: have you actually put the bins out? — ▲ 34 · 💬 7 · 🔗 sjg.io
    Simon Green built BinRange, a Home Assistant setup that tracks whether his six wheelie bins have actually been moved for collection. After Bluetooth RSSI proved too unreliable, he switched to ultra-wideband radios: a fixed anchor and battery tags on each bin. The system measures distance with centimetre-level precision, uses an accelerometer to wake the tags only when a bin moves, and delivers signed firmware updates over Bluetooth. Green leaned heavily on AI tools like Codex throughout the build.
  56. We still can't map earthworms underground—but a worm detector could change that — ▲ 33 · 💬 7 · 🔗 bencology.bearblog.dev
    Earthworm movement remains a blind spot in soil ecology. The only in-situ mapping study, from 1988, tracked Lumbricus terrestris on a dirt road and found they travel at least 9 meters per night. The author proposes a worm detector using ground-penetrating radar or buried probes, combined with GPS and modeling, to create 3D maps or 2D heatmaps of worm activity at paddock scale. This could reveal worm speed, movement patterns, and help farmers manage soil health more precisely.
  57. This Tool Predicts Exactly Where to Find Directly Imaged Exoplanets — ▲ 32 · 💬 3 · 🔗 whereistheplanet.com
    Where is the planet? is a web tool that predicts the locations of directly imaged exoplanets and brown dwarf companions. Built by Jason Wang, Matas Kulikauskas, and Sarah Blunt, it helps astronomers plan observations. The site invites users to request additional orbits and asks researchers to cite its ASCL record.
  58. Atom Beats RSS: Why Your Blog Feed Should Ditch the Old Standard — ▲ 32 · 💬 7 · 🔗 kevincox.ca
    Kevin Cox's technical guide argues that Atom is the superior feed format due to its clearer specification and fewer compatibility issues. He provides a minimal Atom template and covers essential best practices: setting correct Content-Type headers, using absolute URLs, enabling feed discovery via HTML link tags, serving over HTTPS, including full content, maintaining stable entry IDs, handling dates correctly, and leveraging self-links for caching and WebSub.
  59. New York City's tallest tree record is 20 years old — LiDAR reveals bigger candidates — ▲ 31 · 💬 4 · 🔗 blog.willmeye.rs
    The Queens Giant, New York City's tallest carefully measured tree, was last measured in 2004 at 133.8 feet. Using 2021 LiDAR data from The Nature Conservancy and the University of Vermont, the author identifies four trees that may exceed that height, including one in Van Cortlandt Park estimated at 156–160 feet. The post suggests the parks department should verify these candidates before any are accidentally cut down.
  60. Halide's wpd Decoder: 3.19x Faster Than libwebp, Written in Rust — ▲ 30 · 💬 8 · 🔗 halide.cx
    Halide released wpd, a WebP decoder written in Rust that is faster and safer than libwebp. It addresses memory safety vulnerabilities like CVE-2023-4863, which was actively exploited. Benchmarks show up to 3.19x speedup in multithreaded lossless decoding, with single-threaded gains from pure algorithmic improvements. wpd offers full libwebp feature parity plus extras like parallel animation frame decode-ahead and explicit thread-count setting. It's open source under BSD 2-Clause.
  61. Meta's Muse proves that design, not tech, wins consumer AI — ▲ 29 · 💬 34 · 🔗 metedata.substack.com
    Muse's core agentic capabilities have existed for over a year, yet Meta put the pieces together in a way that finally clicks for average consumers. By stripping away model selectors and work/chat switches, giving users a clear mental model of a personal helper with its own computer, and funding a permissive free tier through ads, Meta is positioned to own consumer AI at scale.
  62. Agents Don't Need Memory, They Need Documentation — ▲ 26 · 💬 19 · 🔗 liao.gg
    Memory plugins for AI agents rely on RAG to retrieve snippets from past conversations, but this approach fails because it lacks context, treats outdated information as truth, and is unauditable. The author argues that agents need structured documentation instead—a Markdown brain they consult before work and update after. He built Operator Memory, an open-source plugin that replaces vector databases with plain files, and has used it for over a year.
  63. New Sea Spider Species Discovered in the Salish Sea After a Century — ▲ 21 · 💬 4 · 🔗 nautil.us
    Researchers from the University of British Columbia have identified seven sea spider species in the Salish Sea, including two new to science: Callipallene pilosuspedes and Tanystylum kiixin. This marks the first discovery of new sea spiders in the region in a hundred years. Using scanning electron microscopy and DNA analysis, they found that these ancient arachnids, with up to 12 legs and skin breathing, host tiny ecosystems on their bodies.
  64. Nobody Is Listening on Port 8125 — And That's Fine — ▲ 19 · 💬 15 · 🔗 yeet.cx
    A Fedora box named stick has been sending StatsD datagrams to 127.0.0.1:8125 ten times a second for a week, but nothing is listening. The metrics still show up in Grafana. The trick: an eBPF program watches port 8125 at the traffic control layer, and a JavaScript decoder turns the lines into Prometheus families. The app keeps doing fire-and-forget UDP, and nothing changes on the host except yeet. This is the first post in a series replacing common exporters with yeet scripts.
  65. After 18 Years of Writing Code by Hand, I Can't Imagine Going Back — ▲ 19 · 💬 30 · 🔗 eliocapella.com
    After 18 years of daily coding, the author stopped writing code by hand once AI tools like Claude Code crossed a threshold. At Filestage, monthly merged PRs jumped from 200 to 300 with no rise in bugs, thanks to CI guardrails and AI reviews. He argues abstractions still matter—LLMs should target high-level representations, not assembly—and wonders if UML and ER diagrams could become everyday interfaces to living systems.
  66. Sega SG-1000 Gets Its Own Forth Interpreter — ▲ 16 · 💬 1 · 🔗 leadedsolder.com
    A developer ports CamelForth to the Sega SG-1000, turning the 1983 console into a self-contained programming environment. The project overcomes assembler incompatibilities and provides a REPL for writing assembly and high-level code directly on the hardware, part of RetroChallenge 2026/10.
  67. Google's Gboard Conveyor Belt Version puts a physical keyboard on a moving loop — ▲ 16 · 💬 3 · 🔗 github.com
    Google's mozc-devices repo now hosts design files and firmware for a Gboard Conveyor Belt Version, released October 1, 2026. The directory includes KiCad schematics and PCB layouts, firmware source, and 3D-printable STL files for the case and mechanisms, plus a build guide. It is not an official Google product. The README also points to the standard Gboard for those who prefer typing with their hands.
  68. Thoreau BASIC - Modern 64-bit BASIC with UEFI boot and JIT — ▲ 16 · 💬 5 · 🔗 thoreaubasic.com
    Thoreau BASIC revives the classic GW-BASIC syntax with 64-bit memory, 24-bit graphics, sprites, and mouse input. It includes TCP/IP, HTTP, a complete GM/GS wavetable with MIDI/WAV/FLAC/Opus playback, complex/quaternion/octonion values, a debugger, profiler, and multicore PARFOR. The same BASIC language runs as a Windows program or directly from UEFI firmware without an OS. Programs can be packaged as standalone EXE or bootable EFI applications. Graphics, networking, and sound are ordinary BASIC commands, not separate frameworks. Download for Windows or UEFI and start coding.
  69. If determinism is true, nothing matters — and that's self-defeating — ▲ 16 · 💬 4 · 🔗 blog.coredump.cx
    The author argues that scientific materialism, which denies free will, undermines itself because it relies on perceptual experience to justify its claims. He explores solipsism, the brain-in-a-vat scenario, and the Boltzmann brain, concluding that the simulation hypothesis is just theology for nerds.
  70. Rely - Automation puzzle game with unreliable components — ▲ 15 · 💬 14 · 🔗 think-twice.me
    Rely is a browser-based automation puzzle game where every component has a chance to fail. Build conveyors, use Sorters, Counters, Markers, and Marker-Gates to turn a stream of unreliable parts into a reliable network. Your goal: deliver three 100% reliable components in a row. Start with the guided tutorial, then tackle three beta scenarios with increasingly complex challenges. Play free in your browser with keyboard and mouse. The full game will expand on this core idea of making reliability from chaos.
  71. Programming the ENIAC: rewiring cables instead of writing code — ▲ 13 · 💬 0 · 🔗 columbia.edu
    A Columbia University history page explains how the ENIAC was programmed in the 1940s. There was no stored program: operators set switches and plugged cables to route data between accumulators and function tables, essentially rebuilding the machine's logic for each new problem. The page covers the six women who did this work and the transition to the modern stored-program model.
  72. Docker has always used microVMs (well, since 2016) — ▲ 13 · 💬 6 · 🔗 dave.recoil.org
    The author, who joined Docker in 2015, reveals that Docker Desktop has relied on microVMs since 2016. Built with Mirage Unikernel libraries, the first version, hyperkit, and its successor, Docker VMM, run a minimal LinuxKit kernel inside an embedded VMM. This design enables rootless operation, a minimal device interface, and easy VPN integration. Today, the same microVM technology powers Docker Desktop and Docker Sandboxes, which securely runs coding agents like Claude Code.
  73. IBM's CEO asks how 34 people, including the janitor, beat us — ▲ 13 · 💬 6 · 🔗 computerhistory.org
    In a 1963 memo, IBM president Thomas J. Watson Jr. demanded to know how Control Data Corporation—a company of just 34 people, including the janitor—had seized industry leadership with its CDC 6600 supercomputer. The scathing note, now known as the Janitor Memo, reveals the shockwaves the CDC 6600 sent through IBM's top ranks.
  74. Philip Sidney, the Elizabethan Master of the Parenthesis — ▲ 12 · 💬 0 · 🔗 publicdomainreview.org
    Sir Philip Sidney, the Elizabethan courtier-poet, was the uncontested king of brackets. His romance Arcadia, left unfinished when he died from a battle wound at 31, ends mid-sentence on a parenthesis. With nearly 2,400 brackets, Sidney turned the humble punctuation mark into a tool for delay, intimacy, and narrative complexity, influencing writers for generations.
  75. C2PA's exclusion feature lets you sign nothing and fake timestamps — ▲ 12 · 💬 2 · 🔗 da.vidbuchanan.co.uk
    David Buchanan demonstrates a time-hacking exploit in C2PA by abusing arbitrary byte-range exclusions. By excluding the entire file, he creates a valid signature over an empty hash, allowing post-hoc tampering without invalidating the signature or the trusted timestamp. He shows a photoshopped EuroMillions ticket with a valid C2PA timestamp hours before the draw. The flaw is known but unfixed; he suggests format-specific exclusion rules.
  76. Why One-Line Fixes Touch Ten Files: The Hidden Force of Change Coupling — ▲ 10 · 💬 0 · 🔗 kb.buildingbetterteams.de
    Change Coupling reveals which parts of your codebase change together, even without direct references. It explains why a simple bug fix can cascade into a dozen files. The article presents a 2x2 matrix of physical vs. change coupling to identify areas to consolidate, split, or monitor, and warns against silver-bullet fixes like microservices or event-driven architectures that ignore the underlying forces.
  77. Plastics Have Personalities: From Snow White to Queen Victoria — ▲ 9 · 💬 2 · 🔗 yarchive.net
    Plastics vary as widely as human characters, argues this essay. Polyethylene and polypropylene are Snow White—simple, pure, weak. Nylon is Arnold Schwarzenegger, strong enough for gun parts. Polycarbonate is Achilles, bulletproof but vulnerable to acetone. PVC is Dr. Jekyll and Mr. Hyde, rigid or flexible depending on plasticizers. Polyurethane is a Blade Runner android, strong but doomed to crumble. Teflon is Queen Victoria—noble and nonreactive. The author laments that all plastics get lumped together as merely 'plastic'.
  78. NASA's Secret Plan to Fly the SR-71 Again: A Former Engineer Reveals the Real Reason — ▲ 9 · 💬 5 · 🔗 twz.com
    Tim Conners, former lead propulsion engineer for NASA's SR-71 program, discusses the space agency's quiet effort to revive Tail 844, the Blackbird that vanished from Armstrong Flight Research Center. He believes NASA is developing a high-speed engine test bed, not chasing aerodynamics data, and reveals that power-on testing could begin soon. Conners, who worked on the aircraft in the 1990s, also shares that he was never formally asked to join the team but wouldn't turn down the offer.
  79. Jeffy - Pretrained text classifiers that run and retrain on CPU — ▲ 8 · 💬 3 · 🔗 github.com
    Jeffy is a Python library that brings 13 pretrained text classifiers to your local machine, no GPU required. It handles tasks like SMS spam detection, banking intent, news categorization, and sentiment analysis with up to 99% accuracy. You can also train custom classifiers on your own CSV, TSV, or JSONL data using just a few examples. The package includes a playground, Python SDK, and HTTP API, making it easy to integrate. With fast CPU inference (50-80ms) and a lightweight footprint, Jeffy is ideal for developers who need reliable text classification without cloud dependencies.
  80. A Million Tokens a Second Is Not a Million Times Faster Work — ▲ 7 · 💬 13 · 🔗 echohive.ai
    A million tokens per second can mean context capacity, input processing, aggregate throughput, or single-agent output. Even at that imagined speed, a one-minute check can dominate total time: 800,000 tokens written in 0.8 seconds still take 60.8 seconds end to end. The real bottleneck is judgment—taste, specs, tests, physical evidence, and fidelity—not generation speed.
  81. Claude nearly broke my film scanner, so I built a better pipeline — ▲ 7 · 💬 4 · 🔗 shannadige.com
    Shanna Dige loved shooting 35mm film on a Canon AE-1P, but scanning and editing each roll took four hours. After an attempt to let Claude drive the scanner over SANE nearly destroyed the hardware, she built a custom pipeline: numpy-based negative conversion with roll-wide color consistency, LaMa inpainting for dust removal, and local Qwen3-VL tagging. The studio now publishes rolls to Cloudflare R2 with a playful 3D film canister gallery.
  82. The Softmax Function and Its Derivative: A Deep Dive — ▲ 7 · 💬 0 · 🔗 eli.thegreenplace.net
    Softmax turns any real vector into probabilities that sum to 1, making it a staple of multiclass classification. This post derives its Jacobian from first principles using vector calculus, explains why naive exponentiation causes NaNs and how shifting by the maximum fixes it, and shows how the derivative simplifies when combined with cross-entropy loss.
  83. Inside the Dog Show Where Breeders Chase an Ideal That Never Existed — ▲ 4 · 💬 0 · 🔗 theparisreview.org
    At a Virginia conformation show, over 400 purebreds compete to match breed standards—documents that turn historical function into measurable traits. The author observes the tension between the ideal and its approximation, where dogs are both family and commodity, and where preserving breed history often means freezing it in place.
  84. Born Twice: A Single HTML File That Records Specimens at the Moment of Collection — ▲ 3 · 💬 2 · 🔗 tjid3.org
    T.M. Jones introduces the Independent File (IF), a 750 KB HTML document that captures voucher data—UUID, QR code, photos, collector, locality, date—right when a plant is collected. It runs offline in any browser, validates fields, maps to Darwin Core, and exports JSON, CSV, XLSX, and packages for Symbiota, Specify, and DiSSCo. The goal: start the record early, so herbaria receive a complete record instead of rebuilding one.
添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论