HN Digest 2026-10-02 · Debian patches Linux kernel against hundreds of vulnerabilities, including privi

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. Debian patches Linux kernel against hundreds of vulnerabilities, including privilege escalation and data leaks — ▲ 545 · 💬 393 · 🔗 lwn.net
    Debian has issued a security advisory for the Linux kernel, addressing a massive list of vulnerabilities that could lead to privilege escalation, denial of service, or information leaks. The fixes are available in version 6.12.111-1 for the stable distribution (trixie). Users are strongly advised to upgrade their Linux packages immediately.
  2. Frog and Toad return as a sharp AI fable, and Hacker News can't agree on who's to blame — ▲ 537 · 💬 120 · 🔗 frogandtoad.ai
    A pastiche of Arnold Lobel's Frog and Toad books, hosted at frogandtoad.ai, drew 100 points and 16 comments on Hacker News. Readers praised the writing and art as a well-executed, Aesop-like story about increasingly capable machines, while others questioned whether Lobel's estate was compensated and whether the piece was AI-generated.
  3. Court Blocks Utah's VPN Law, Calling It a Technical Impossibility — ▲ 456 · 💬 205 · 🔗 eff.org
    A federal judge issued a preliminary injunction against Utah's SB 73, which would have forced adult websites to block VPN users or verify every visitor's location. The court agreed with EFF that perfect geolocation is impossible, making the law unconstitutional. The ruling sets a precedent against anti-VPN mandates that demand invasive surveillance and threaten privacy nationwide.
  4. DeepSeek Harness goes open source with an everything-is-a-plugin architecture — ▲ 380 · 💬 205 · 🔗 deepseek.com
    DeepSeek has launched a public preview of DeepSeek Harness, an open-source agent harness for macOS and Windows. Built on Cordis's 'everything is a plugin' design, it handles everyday work, coding, research, and background tasks, and lets you build new plugins through chat in Creator mode. Install via npx or from source, and extend it with community plugins.
  5. Apple's Pass Designer lets you see exactly how a pass will look on iPhone and Apple Watch — ▲ 284 · 💬 192 · 🔗 developer.apple.com
    Apple has introduced Pass Designer, a new tool that simplifies creating and previewing Wallet passes for any business, from a local gym to a global airline. It offers templates, real-time previews using the same rendering as iOS and watchOS, and supports semantic tags for features like Siri Suggestions and Maps. The beta requires macOS 27 or later and is available for download.
  6. Shimano's Bicycle Museum Puts Bikes Before Brand — ▲ 282 · 💬 71 · 🔗 inrng.com
    The Shimano Bicycle Museum in Sakai, near Osaka, is a glass-and-steel space that celebrates cycling rather than corporate history. Visitors see a chronological parade of machines, from hobby horses to modern race bikes, all within reach. A Dura-Ace groupset is exploded into 3,200 parts, but there's no hard sell and no gift shop—just a research archive and a clear message: there are as many ways to enjoy cycling as there are riders.
  7. FLUX 3 Image lets you place every object with bounding boxes — ▲ 261 · 💬 57 · 🔗 bfl.ai
    Black Forest Labs introduces FLUX 3 Image, a model built for maximum control over every pixel. Users can compose images from scratch by laying out elements with bounding boxes on a 0-to-1000 grid, then replay and refine the result. The model supports precise editing, multiple aspect ratios, and commercial weights, with a demo and documentation available now.
  8. Singapore's government-run dating service uses a Nobel Prize-winning algorithm, but it won't solve the dating crisis — ▲ 243 · 💬 337 · 🔗 singapore-samizdat.com
    GovTech's FirstDate pilot matches government workers using the Gale-Shapley stable marriage algorithm, with Singpass verification and a 30+ question questionnaire. But the algorithm only guarantees stable matches, not happy ones—and with no commitment to the process, stable matches mean nothing when better options exist outside the pool. The real problem with modern dating isn't mismatched preferences.
  9. The Legend of John von Neumann: The Mathematician Who Could Divide Eight-Digit Numbers in His Head at Age 6 — ▲ 234 · 💬 136 · 🔗 gwern.net
    Paul Halmos's 1973 essay profiles John von Neumann, the Hungarian-American polymath who revolutionized quantum physics, game theory, and computing. It traces his journey from a Budapest childhood—where he memorized phone books and mastered calculus by age 8—to Princeton's Institute for Advanced Study. The piece mixes verified achievements with colorful legends: his love of puns, his notorious driving mishaps, and his wife's remark that after 17 years in their house, he still didn't know where the glasses were kept.
  10. Mike Tomlin spent 12 years building a Minecraft city — ▲ 208 · 💬 56 · 🔗 nytimes.com
    Former Steelers coach Mike Tomlin finally opened his 12-year Minecraft project to the public in a 19-minute YouTube tour. The video drew 20,000 concurrent viewers and over 264,000 views as Tomlin guided fans through high-rises, a boutique hotel, an executive spa, and a speakeasy, explaining his meticulous approach to lighting, amenities, and interior detail. His son Dino hyped the reveal on X for three weeks.
  11. Nabu Casa renames Home Assistant Cloud to Link, saying Big Tech ruined the cloud — ▲ 204 · 💬 100 · 🔗 home-assistant.io
    Nabu Casa is renaming Home Assistant Cloud to Home Assistant Link, arguing that Big Tech has tarnished the cloud with subscriptions, outages, and data harvesting. The company stresses that Link remains optional, private, and free from lock-in, with profits funding the Open Home Foundation. The change becomes official in the December 2026.12 release.
  12. Turbo Haskell runs GHC itself on the JVM — ▲ 194 · 💬 58 · 🔗 comonad.com
    Edward Kmett built THC, a Turbo Haskell compiler that takes GHC Core and JITs it onto the JVM via Truffle and GraalVM. It implements every GHC 9.14.1 prim-op, supports Template Haskell, Linear Haskell, and polyglot FFI to Python, Ruby, R, and JavaScript. It can compile pandoc, happy, alex, and now GHC itself. Performance ranges from 3× faster to 3× slower than GHC, mostly 10–20% slower.
  13. ChatGPT now lets you build and publish full websites from a single prompt — ▲ 194 · 💬 211 · 🔗 chatgpt.com
    OpenAI's new ChatGPT Sites feature turns a prompt into a hosted website or app. Describe what you want, refine it conversationally, then publish with access controls, visitor sign-in, and custom domains. It also connects to external tools with read-only permissions, and new sites default to private.
  14. ICC Judge Kimberly Prost: U.S. Sanctions Are 'Completely Futile' — ▲ 193 · 💬 139 · 🔗 npr.org
    Canadian judge Kimberly Prost describes how U.S. sanctions upended her daily life—canceling credit cards, blocking services like Alexa, and prompting her health insurer AXA to deny claims—even though she was sanctioned for a judicial decision in an Afghanistan investigation. She is challenging the sanctions in U.S. court and vows to keep hearing cases, including alleged crimes in Timbuktu, Mali.
  15. Supabase acquires Turso to make databases as cheap as files for AI agents — ▲ 190 · 💬 97 · 🔗 supabase.com
    Supabase is acquiring Turso, the company that rebuilt SQLite in Rust and built a cloud platform where one server can manage millions of databases, loading and suspending them on demand. Supabase already launches over a million databases weekly and expects agent-driven demand to outpace global capacity. Turso keeps operating and continues its SQLite work, while Supabase stays on Postgres; together they aim to give every agent the same developer experience from prototype to production.
  16. Zig 0.17.0 brings incremental compilation to x86_64-linux — ▲ 190 · 💬 99 · 🔗 ziglang.org
    Zig 0.17.0 lands after five months and 925 commits from 206 contributors, reworking the build system with a new Build Server Protocol and upgrading the ELF linker so incremental compilation is expected to work for everyone on x86_64-linux. The release also adds a formally specified and fuzzed grammar, moves C translation to an external package, and introduces SafeAllocator, while dropping several legacy syntax forms.
  17. stillwet - AI models paint in a simulated oil studio — ▲ 180 · 💬 60 · 🔗 stillwet.art
    stillwet is an art project where AI models like Claude Opus 5.5, Gemini 3.8 Flash, and MiMo v2.6 Pro create paintings in a simulated oil paint studio. Each model writes every brushstroke, and a simulation of bristles, wet paint, and linen carries it out. Most works are after Caspar David Friedrich, based only on written research—the models never see his paintings. The site showcases 54+ paintings across 18 rounds, with insights into model behavior, such as recurring motifs and surprising coincidences. Visitors can watch timelapse videos of each painting being created.
  18. AI Makes Me Sad — ▲ 178 · 💬 214 · 🔗 mondobe.com
    A Georgia Tech student and teaching assistant watches classmates chase $500k OpenAI jobs while he dreads an industry that feels like Charlie Chaplin being swallowed by the machine in Modern Times. He mourns the loss of craft and the startup era, but finds unexpected hope: institutions like Twitter and LinkedIn are young enough to rebuild, and the coming AI bubble's burst might bring a satisfying clarity.
  19. Greg Kroah-Hartman debunks AI security hype: 79 vulnerabilities, only 10 real — ▲ 159 · 💬 32 · 🔗 youtube.com
    In a Kernel Recipes 2026 talk, Greg Kroah-Hartman dissected 79 vulnerabilities reported by an AI model. Most were noise: 24 lacked detail, 14 weren't bugs, 3 were fabricated, and 15 were already fixed. Only 10 were genuine fixes. He warns against uncritical parroting of AI security press releases.
  20. Aging May Be a Loss of Cell Identity, Not Just Wear and Tear — ▲ 158 · 💬 40 · 🔗 erictopol.substack.com
    Two new papers, one in Nature by Vadim Gladyshev's Harvard team and one in Cell by Juan Carlos Izpisua Belmonte's group at Altos Labs, propose that aging is driven by the erosion of epigenetic constraints that lock cells into their specialized identities. This loss leads to mesenchymal drift, where cells take on fibroblast-like properties, causing fibrosis and inflammation. The work suggests interventions like caloric restriction, partial reprogramming with Yamanaka factors, and lithium may slow or reverse this process.
  21. AI finally beats the best Stratego player in history — on a budget of 16 GPUs — ▲ 158 · 💬 74 · 🔗 arstechnica.com
    Stratego, a game of hidden identities and long, bluff-heavy matches, had resisted AI even as chess, Go, and poker fell. Ataraxos, built by researchers at Carnegie Mellon, MIT, NYU, and Stanford, beat world champion Pim Niemeijer 15–1 with four draws. A second neural network that guesses hidden pieces made the difference, and training cost just a few thousand dollars on 16 GPUs.
  22. Tiny Brutalism: A Cozy Sandbox for Not-So-Cozy Architecture — ▲ 136 · 💬 31 · 🔗 placeholders.itch.io
    Tiny Brutalism is a relaxing freeform building game for Windows, macOS, and Linux, inspired by brutalist architecture. For $5, players can construct strange monuments, impossible structures, or concrete towers with no goals or rules. The game features intuitive controls, day/night cycles, and adjustable plot sizes. Community feedback highlights its charm and potential, with requests for export options, bigger lots, and performance fixes.
  23. Chinese Social Media Is Like American Junk Food — ▲ 132 · 💬 134 · 🔗 lesswrong.com
    A first-generation Chinese-American explores four cultural differences that shape China's online ecosystem: the dominance of social reality and shame, a cynical 'Dark World' view of human nature, collective insecurity driving nationalism, and a tolerance for low-quality content that made AI slop a natural fit. He argues these traits are crucial for anyone trying to communicate AI safety or rationalist ideas in China.
  24. Audionaut - Open-source multitrack audio editor with AI agent control — ▲ 131 · 💬 45 · 🔗 github.com
    Audionaut is a free, open-source multitrack audio editor and recorder built in modern C++ with JUCE, running natively on Windows, macOS, and Linux. It offers precise cutting, per-track playlists, multi-channel support, and clean exports without the complexity of a full DAW. What sets it apart is MCP support: Claude and other AI agents can drive edits directly, turning commands into single undo steps. It also includes stem separation via demucs.cpp, audio analysis with Essentia, and a headless CLI for scripting and CI. Perfect for music, podcasts, and multitrack projects.
  25. AI is Solving Math's Hardest Problems, and Mathematicians Are Grieving — ▲ 127 · 💬 110 · 🔗 xenaproject.wordpress.com
    Language models are now cracking problems humans couldn't, and the math community is reeling. From denial to depression, many mathematicians are grieving what they see as a loss. But not everyone: some, like Kevin Buzzard, are excited. He argues that the real value of math is human understanding, yet AI's rapid progress threatens even that refuge. A provocative look at AI's impact on mathematics.
  26. Redis creator launches ds4 to run frontier LLMs locally — ▲ 125 · 💬 35 · 🔗 dwarfstar.sh
    Salvatore Sanfilippo, creator of Redis, has released DwarfStar 4 (ds4), a C inference engine for high-memory Mac, CUDA, and ROCm machines. It runs DeepSeek V4/V4.1, GLM 5.x, and Qwen3.8 Flash Next with asymmetric 2-bit quantization, KV cache disk streaming, and a CLI, HTTP server, and native agent. On an M5 Max with 128 GB, it achieves up to 790 tokens/s prefill and 39 tokens/s generation.
  27. Nazi Uranium Cubes Reveal Germany Never Had Enough Material for an Atomic Bomb — ▲ 121 · 💬 97 · 🔗 science.org
    Measurements of surviving uranium cubes, archival research, and computer simulations show that Nazi Germany's combined stocks of uranium and heavy water were far too small to sustain a nuclear chain reaction. The study challenges Werner Heisenberg's postwar claim that the B8 reactor was close to criticality, and finds no evidence he deliberately sabotaged the program.
  28. Agentic Coding's Four Horsemen: Slop, Alienation, Deskilling, and Isolation — ▲ 105 · 💬 80 · 🔗 distantprovince.substack.com
    Agentic coding is undeniably useful, but it's also wreaking havoc on software engineering. LLM-generated code has a distinctive 'slop' smell that makes codebases repulsive to humans. Engineers become alienated from their craft, caring less about the code they produce. AI erodes existing skills and impedes learning, with no meaningful skill progression. And agents weaken team social bonds by reducing interpersonal communication and increasing self-reliance. The author sees no solution in sight.
  29. Meta Open-Sources Muse Gadget SDK for ESP32 and Linux — ▲ 103 · 💬 51 · 🔗 gadgets.muse.ai
    Meta has released an open-source SDK that turns off-the-shelf ESP32 boards or a Raspberry Pi into Muse gadgets. The SDK includes firmware for displays, buttons, sensors, and audio, and supports devices like the Waveshare AMOLED, M5Stack StickS3, and Seeed reTerminal. It's built by hackers for hackers, with examples ranging from Home Assistant integration to a TV stick.
  30. Three Dutch Computer Museums That Will Amaze You — ▲ 98 · 💬 26 · 🔗 aresluna.org
    Marcin Wichary recounts visiting three Dutch computer museums: HomeComputerMuseum in Helmond with over 500 interactive machines, Bonami SpelComputer Museum in Zwolle with thousands of computers and arcade games, and a barn housing a DEC collection. He highlights the regional artifacts, including the Aesthedes graphic editor with its massive keyboard, and reflects on how these museums offer a unique perspective on computing history outside the US.
  31. GrapheneOS fixes Android 17 QPR1 kernel regression that freezes Pixels — ▲ 98 · 💬 31 · 🔗 discuss.grapheneos.org
    Android 17 QPR1, released for Pixels on September 15, introduced a kernel driver regression causing stuttering, lag, and freezes under memory pressure. GrapheneOS patched it for its release today, but Google's fix may not arrive until December 2026. The project is now handling basic updates and quality control for the entire OS, and has also found a way to cut most extra memory use from its secure spawning feature.
  32. Wagtail's one-month AI coding challenge failed—and that's the point — ▲ 95 · 💬 75 · 🔗 wagtail.org
    Wagtail set out to spend a month coding exclusively with GLM 5.3 Flash. They succeeded for two weeks, then burned 1B tokens on other models. A vibe-coded MCP prototype cost $150 overnight, and infrastructure limits forced switches to DeepSeek V4.1 Flash and Qwen 3.8 Flash. They share benchmarks and lessons for budgeting experimentation and measuring energy, not just tokens.
  33. The People Who Made Big Platforms Interesting Are Quietly Walking Out the Side Door — ▲ 95 · 💬 83 · 🔗 widdershins.verja.net
    A wave of blog posts — leaving Reddit, WordPress, Discord, GitHub, Photoshop — all share the same shape: someone spent years somewhere, it changed under them, and now they're packing up. What strikes me is that hardly anyone is leaving for something bigger. They're going smaller: a personal site, a self-hosted app, a capsule nobody stumbles onto by accident. Less reach, more control. A few years ago that sounded like giving up; now it sounds sensible.
  34. US Threatens France and Germany With Diesel Export Ban — ▲ 92 · 💬 130 · 🔗 reuters.com
    Washington is pressuring France and Germany to release their diesel reserves, warning that refusals could trigger a US export ban, according to Reuters. The standoff over fuel stockpiles raises questions about energy security and transatlantic trade ties.
  35. World Bank Aid Is Being Laundered Through Crypto Wallets, Study Finds — ▲ 85 · 💬 29 · 🔗 nber.org
    A new NBER working paper tracks $238 billion in World Bank aid across 93 countries from 2018 to 2024 and finds that crypto activity spikes sharply in the month aid tranches arrive, driven by anonymous, newly created wallets on both tax-haven and mainstream exchanges. The authors estimate that 2 to 6 cents of every aid dollar is diverted, roughly $1.7 to $4.4 billion in total, and that the four affected sectors keep receiving half of subsequent World Bank funding.
  36. Butterflies Use Optical Illusions to Make Predators Miss — ▲ 77 · 💬 19 · 🔗 essex.ac.uk
    New research from the University of Exeter and the University of Essex, published in Nature, shows that butterfly wing patterns create optical illusions that confuse predators' motion perception. High-speed footage, a study of 400 European species, touchscreen experiments with 100 volunteers, and simulations evolving 50,000 wing patterns all point to the same conclusion: the markings mimic the barber-pole effect, scrambling a bird's sense of speed and direction so it strikes thin air or a less vital wing part. The finding may extend to zebras, snakes, and other animals.
  37. Every SaaS Business Will Become a Harness Around a Model — ▲ 77 · 💬 57 · 🔗 blog.sshh.io
    The author argues that all SaaS companies will eventually become harnesses around stateless LLMs, progressing from engineers pairing with agents to proactive background agents that orchestrate humans. The company's domain knowledge, tooling, permissions, and review loops become the business harness, and differentiation shifts to how well a company builds that harness rather than the software itself.
  38. GPT-6 Astra beats World of Warcraft's starting zone in 40 minutes without dying — ▲ 69 · 💬 55 · 🔗 agent-wow.sh
    I built agent-wow to test how frontier LLMs handle World of Warcraft without any game-specific training. Using GPT-6 Astra, the agent created an orc and completed every quest in the Valley of Trials in 40 minutes with zero deaths. It mined AzerothCore SQL files for quest data, generated a C++ pathfinding helper using Detour, and communicated directly via the WoW network protocol—no computer vision or game hacking required. The long-term goal: fill a server with AI agents and see if they can clear Icecrown Citadel on heroic.
  39. A Lean Proof That Adding Two Binary Numbers Is Regular, Explained for Software Engineers — ▲ 69 · 💬 6 · 🔗 agostbiro.net
    A software engineer walks through a formal Lean proof that the language of bit-column strings where the bottom row equals the sum of the top two is regular. The trick: reverse the language so a DFA can process columns least-significant-bit first, carrying only a single bit of state. The post covers DFAs, regular languages, the adder DFA, and the inductive invariant, showing what it takes to verify a system formally.
  40. Media Fans Are Ditching Algorithms for CDs, DVDs, and Vinyl — ▲ 68 · 💬 105 · 🔗 theguardian.com
    Frustrated by streaming services that remove titles and curate via algorithms, people are returning to physical media. A university lecturer with over 5,000 records and CDs says playlists replaced album listening; a compliance manager turned to flea markets after HBO pulled 200 episodes of Sesame Street in 2022; a retired professor owns 5,300 CDs in a walk-in closet. They cite ownership, artist support, data privacy, and freedom from algorithmic control.
  41. Europe's payment giants unite to take on Visa and Mastercard — ▲ 61 · 💬 84 · 🔗 rte.ie
    Five European payment groups—including Spain's Bizum and Italy's Bancomat—are forming the European Network for Payments (ENP), a Madrid-based joint venture serving 130 million users across 13 countries. The network will connect national instant payment systems through a common layer, starting with cross-border person-to-person transfers, then e-commerce and point-of-sale. The move aims to reduce Europe's dependence on US card networks, though analysts say its success hinges on merchant adoption and regulatory support.
  42. Pyxel - Python retro game engine with built-in art and sound editors — ▲ 61 · 💬 8 · 🔗 github.com
    Pyxel is a retro game engine for Python that brings back the charm of classic consoles with 16-color palettes and 4 sound channels. It includes built-in image, tilemap, sound, and music editors, so you can create pixel-art games without leaving the environment. Inspired by PICO-8 and TIC-80, Pyxel lets you write code in Python and instantly see results. With over 18,000 GitHub stars, it's a thriving open-source project under the MIT license. Whether you're a beginner or a seasoned developer, Pyxel makes retro game creation accessible and fun. Start building your nostalgic masterpiece today!
  43. Meta's Muse is fantastic for web scraping — ▲ 59 · 💬 74 · 🔗 sigh.dev
    A developer building fullsets.fm, a side project that aggregates professionally recorded concerts, uses Meta's Muse to scan Reddit and YouTube for thousands of videos from obscure channels. For $80 a month, Muse burns billions of tokens weekly, often without counting usage. The model is mediocre at writing and coding, but excels at Gemini Flash-style workloads. The developer worries about the ethics of AI agents that never sleep and can impersonate real users.
  44. ldraw-nova - Agent tooling for generative LEGO models — ▲ 59 · 💬 36 · 🔗 github.com
    ldraw-nova is an open-source tool that lets AI agents design buildable LEGO models. Give an agent a prompt, and it plans, iteratively renders, and adjusts until it produces LDraw source code, a Blender-editable glTF file, and interactive 3D views. Powered by GPT-6 Astra and Claude Opus 5.5, it uses semantic search with re-ranking to find suitable parts and examples. Run it via Docker to generate physical, buildable LEGO creations from simple ideas.
  45. Venice's failed war against Constantinople led to the first bond market — ▲ 58 · 💬 19 · 🔗 bigthink.com
    In the 12th century, Venice's military disaster against Constantinople forced the republic to invent something the world had never seen: government debt that ordinary citizens could buy and sell. That innovation became the first bond market, reshaping global finance for centuries to come.
  46. Figure AI melts its F.02 robots in Finland with a little help from Arnold Schwarzenegger — ▲ 51 · 💬 21 · 🔗 figure.ai
    Figure AI decommissioned its F.02 robot fleet by melting the robots in a Finnish foundry. After Arnold Schwarzenegger suggested the idea online, the company trained a new AI model in simulation to make the robots jump autonomously into a 75-ton electric arc furnace. The resulting metal bars are now being machined into limited-edition artifacts commemorating F.02.
  47. I gave 25 friends custom text buzzes based on Morse code — ▲ 50 · 💬 28 · 🔗 liquidbrain.net
    Tired of not knowing who's texting when his phone buzzes in his pocket, the author manually created custom vibration patterns for about 25 contacts. Each pattern encodes the person's initials in Morse code, recorded via iPhone's Create New Vibration. A Claude artifact helped generate the beat sequences, and the whole process took 45 minutes.
  48. Emacs, Org-mode, Gleam, and Nix: One Blogger's Absurdly Overbuilt Static Site — ▲ 48 · 💬 9 · 🔗 byzantine-systems.github.io
    A developer describes a static blog pipeline built around Emacs and Org-mode, with Pandoc converting Org to Markdown, Gleam and Blogatto generating HTML, Lustre handling templates, and Nix pinning the environment. The author argues Org-mode beats Markdown because it is a coherent, interactive document model, and defends keeping everything inside Emacs as "good bloat."
  49. Oracle's Wisconsin AI datacenter may miss its 2027 deadline as power approval stalls — ▲ 47 · 💬 24 · 🔗 theregister.com
    Oracle's Project Lighthouse in Port Washington, Wisconsin, could slip past its planned 2027 customer delivery because the transmission line needed to power it is still awaiting approval from the Public Service Commission of Wisconsin. Aterio's base case puts partial power in December 2027 and full supply in October 2028, and even that leaves little time for testing and commissioning. Oracle, which reportedly sent a force majeure notice on its New Mexico campus, says the project remains on schedule.
  50. Now That's an Impurity Story — ▲ 46 · 💬 4 · 🔗 science.org
    A Hacker News discussion thread links to a science.org blog post titled "Now That's an Impurity Story." The linked article could not be fetched, so the thread itself offers little detail beyond the headline and the community's reactions. The post drew 12 points and a handful of comments within its first eight hours.
  51. Jev, TypeSafe's New Classifier, Is Barely Better Than Random at Probability — ▲ 46 · 💬 18 · 🔗 maximumeffort.substack.com
    TypeSafe's Jev, a 'System One' decision model, outputs probability distributions over multiple-choice answers. But when tested on 1,000 physics problems with known distributions like Maxwell-Boltzmann and Gaussian, Jev scored a mean total variation of 0.518—barely better than the 0.546 from a uniform guess. It tends to be overconfident and peaky, often copying parameters from the prompt. The author questions using Jev as an automated judge for LLM answers.
  52. Micron CEO: RAM Shortage Will Persist Through 2028 — ▲ 41 · 💬 60 · 🔗 arstechnica.com
    Micron CEO Sanjay Mehrotra told investors that memory demand will outstrip supply for at least two more years, with 75 percent of 2027 output already sold and most current negotiations focused on 2028. Samsung's Kim Taewoo expects HBM to consume nearly 30 percent of DRAM wafer capacity next year, squeezing consumer device memory and driving up prices for PCs, smartphones, and consoles.
  53. Breadcrumb - Record everything on your Mac + context manager for AI — ▲ 41 · 💬 5 · 🔗 innerloop.works
    Breadcrumb is a local, encrypted AI plugin for developers that records your screen, meetings, and AI sessions, then feeds that context to Claude, ChatGPT, Cursor, and opencode. It lets you set scoped rules, create one-word shortcuts, and pick up long chats without token bloat. Everything stays on your Mac, runs on local models, and requires no account or telemetry. Free for personal use, with optional $299 founding seats for teams.
  54. Google's Project Suncatcher prototype satellite is now in orbit — ▲ 38 · 💬 45 · 🔗 blog.google
    Google launched its Project Suncatcher prototype satellite with Planet aboard SpaceX's Transporter-18 rideshare mission, and the team has confirmed contact. The spacecraft is operating as expected. Over the coming weeks, Google will collect in-orbit data on how its TPUs withstand spaceflight stress and the radiation and thermal extremes of space, part of a long-term moonshot exploring whether space could host scalable machine learning infrastructure.
  55. Apple's first home camera reportedly won't record video at all — ▲ 35 · 💬 66 · 🔗 theapplepost.com
    Apple's upcoming smart home camera, known internally as J450, reportedly skips video recording entirely. According to Bloomberg's Mark Gurman, the device uses a low-frame-rate sensor, facial recognition, and infrared to detect people and pets, then reports activity and drives automations instead of storing footage. That sets it apart from Ring, Google Nest, and even Apple's own HomeKit Secure Video — and raises questions about how it would help after a break-in or package theft.
  56. What if we stopped using GPUs? — ▲ 35 · 💬 8 · 🔗 youtube.com
    A Hacker News discussion sparked by a video that asks what would happen if the tech industry abandoned GPUs. The thread weighs the idea against the reality that GPUs dominate AI and graphics workloads, and commenters debate whether CPUs, FPGAs, or dedicated accelerators could ever fill the gap.
  57. Apple's M4 chips forget everything when they sleep, and that's why Linux wouldn't boot — ▲ 35 · 💬 0 · 🔗 yuka.dev
    I bought an M4 Mac mini in 2024 hoping to boot Linux, but Apple's new SPTM security and a CPU quirk got in the way. The M4's WFI instruction wipes all registers, violating the ARM64 spec. By replacing every WFI with NOP, I finally booted Linux with all cores. The fix is now in mainline Linux and m1n1, and it also works on M4 Pro, M4 Max, and M5.
  58. Personal Computing 2.0: Your Data, Your Computer, Your Revolution — ▲ 34 · 💬 13 · 🔗 ideas.imbue.com
    The author, who fell in love with personal computing as a child, argues that the original PC era gave way to an internet strangled by Big Tech, spam, and closed AI. He proposes Personal Computing 2.0: a paradigm where your data is truly yours, running on any device, powered by open-weight AI models. This shift would restore privacy, enable instant search across all your data, and force AI companies to compete on cost and performance rather than lock-in.
  59. The first packet sent by carrier pigeon per RFC 1149 is up for auction at Christie's — ▲ 34 · 💬 5 · 🔗 onlineonly.christies.com
    In 2001, the Bergen Linux User Group implemented RFC 1149, sending IP packets via carrier pigeon. One of the actual paper packets, inscribed to RFC author David Waitzman, is now at Christie's with a current bid of $1,300 and an estimate of $2,000–$3,000. The auction closes in 11 days.
  60. The Only Shuttle Mission to Abort After Launching—and Still Succeed — ▲ 32 · 💬 10 · 🔗 en.wikipedia.org
    STS-51-F, also known as Spacelab 2, launched on July 29, 1985, and became the only Space Shuttle mission to perform an Abort to Orbit after a sensor failure shut down a main engine. Despite the lower altitude, the crew conducted a full Spacelab mission, tested the ESA's Instrument Pointing System, and even carried out a heavily publicized Coca-Cola vs. Pepsi taste test in space.
  61. The Only Intuitive Interface Is the Nipple? Tracing the Quote's Origins — ▲ 30 · 💬 32 · 🔗 greenend.org.uk
    The saying 'the only intuitive interface is the nipple' is often attributed to Bruce Ediger, but he denied creating it in 2001. This investigation digs through Google's Usenet archives to find its earliest appearances, starting with Scott Francis in August 1994 and Jay Vollmer in January 1995, and shows how Ediger adopted and adapted the phrase. It also highlights a later twist from Ediger: 'There is no intuitive interface, not even the nipple. It's all learned.'
  62. Mozilla is shutting down Solo, its AI website builder — ▲ 29 · 💬 58 · 🔗 support.soloist.ai
    Mozilla has decided to shut down Solo, its AI website creator. All Solo websites and data will be permanently deleted on November 30, 2026. Users can download an HTML export and a CSV of image links, but image source files must be saved individually. Pro subscribers get prorated refunds as of October 1, 2026, and Solo Domains must be transferred to Name.com. Support is available at support@soloist.ai.
  63. Speedrun.com's new terms kill the open license and ban community tools — ▲ 29 · 💬 9 · 🔗 therun.gg
    On October 1, 2026, speedrun.com replaced its terms of use, removing the Creative Commons license that had protected community data since 2021. The new terms grant parent company Elo ownership of all data, define donations as commercial use, and ban any project with 'substantially similar' features—effectively outlawing most community tools. The author, who runs a speedrun stats site, explains the impact and urges runners to export their data and push back.
  64. A frontier model with only grep matches a tuned RAG pipeline at 0.61, but takes five times as long — ▲ 25 · 💬 2 · 🔗 kapa.ai
    Kapa built Company Knowledge Bench, 1,000 eval cases from real production data, to measure retrieval on messy company knowledge. They scored seven retrievers: hybrid search scores 0.41, adding a reranker lifts it to 0.50, query decomposition reaches 0.56, and an optimized agentic retriever hits 0.65 in about five seconds. A frontier model with only grep matches a tuned pipeline at 0.61 but takes five times as long. Public benchmarks were too narrow, so they built their own.
  65. Capcom's RE:Dox engine hits 2.8x faster parallel deserialization than System.Text.Json — ▲ 25 · 💬 6 · 🔗 github.com
    RE:Dox is a high-performance, token-based structured data engine for .NET, built as a core component of Capcom's next-generation game engine. It parses JSON, CBOR, MessagePack, and more into a compact 64-bit token DOM that supports editing, format conversion, and automatic parallel deserialization. Benchmarks show up to 2.8x faster parallel deserialization, 1.8x faster sequential deserialization, and 1.6x faster serialization than System.Text.Json, with significantly lower memory allocation. Licensed under Apache-2.0.
  66. Firezone ditches SCIM to build its own directory sync engine — ▲ 22 · 💬 2 · 🔗 firezone.dev
    Directory sync is the process of copying users and groups from an identity provider like Okta or Entra into your app. Firezone explains why SCIM, the standard for this, fails in practice due to inconsistent provider implementations, and how they built a pull-based sync engine that handles pagination, rate limits, bad responses, and overlapping jobs.
  67. ProVer fixes GRPO's credit assignment by verifying only pivotal decisions — ▲ 22 · 💬 3 · 🔗 arxiv.org
    GRPO assigns the same trajectory-level advantage to every token, blurring which decisions actually mattered. ProVer uses an agentic judge to contrast successful and failed rollouts and propose a segment that may explain the difference. Instead of trusting the judge, it verifies the segment by comparing terminal success rates of continuations sampled before and after it, then folds positive estimates into GRPO advantages. On ALFWorld, WebShop, and SearchQA, ProVer beats GRPO by 9.91% and 7.12% for Qwen3.5-2B and Qwen3.5-4B.
  68. Trail of Bits Unveils SequenceHash to Stop a Common Cryptographic Blunder — ▲ 21 · 💬 0 · 🔗 blog.trailofbits.com
    Trail of Bits has released SequenceHash and SequenceMAC, hash-agnostic multihashing constructions that work with SHA2, BLAKE, and RIPEMD, unlike NIST's Keccak-only TupleHash. They provide unambiguous input encoding, length-extension prevention, and optional customization strings. The spec is now part of C2SP, with Rust, Go, and Python implementations and test vectors.
  69. Harvard students launched the first college radio club in 1906 — ▲ 21 · 💬 4 · 🔗 blog.archive.org
    The Internet Archive's DLARC college radio collections trace student radio from a forgotten 1906 Harvard club to 1920s broadcast experiments at Dartmouth and the Milwaukee School of Engineering, plus newly digitized materials from Amherst, Middlebury, and William & Mary. The archives preserve handmade posters, DJ manuals, QSL cards, and even a 1987 WMSE metal chart tracking airplay for Slayer and Megadeth.
  70. Daily puzzle challenges you to reverse-engineer a hidden Python function — ▲ 18 · 💬 1 · 🔗 codeset.ai
    Every day, a new hidden Python function `f` transforms text, and you have seven rows to figure out its rule. Test it by entering any text to see the exact output, then predict the output for new inputs. Correct characters are marked by position, wrong ones by presence. Get three outputs right to win, with fewer rows giving a better score.
  71. AI Was Supposed to Make Everything Cheap. Instead, It's Driving Prices Up. — ▲ 16 · 💬 34 · 🔗 derekthompson.org
    Americans are furious about the cost of living, and AI is making it worse. For years, Sam Altman and Marc Andreessen promised AI would create abundance and push prices toward zero. But unemployment hasn't spiked and prices keep climbing. The AI buildout is minting billionaires while directly raising costs across the economy—an irony that doesn't require picking a side to see.
  72. Blender's Fast Blur Isn't Fast: A New Gaussian Blur That Handles Animated Radii — ▲ 16 · 💬 1 · 🔗 aras-p.info
    Aras Pranckevičius noticed Blender's Fast Blur mode is slow for large radii and built Smol Gaussian: downsample, apply a small separable Gaussian, then reconstruct. It matches Skia's GPU blur but adds pixel-area integration for odd sizes, variance subtraction from downsampling, and a cubic B-spline upsample to avoid bilinear slope steps on HDR highlights. An interactive WebGPU playground compares it with Dual Kawase, Skia Gaussian, Blender's Fast Gaussian, and Ryg Blur.
  73. America's Largest Foundry Coke Producer Is Operating Under an Expired Permit — ▲ 15 · 💬 5 · 🔗 insideclimatenews.org
    ABC Coke, the nation's largest foundry coke producer, has been operating without a valid air permit since its 2023 renewal application stalled. The Southern Environmental Law Center sued Jefferson County health officials to force action, and a day later the county agreed to move forward. The plant has a history of benzene pollution in nearby Tarrant, a predominantly Black community.
  74. ExplainDB: A Database System Built for Understandability — ▲ 15 · 💬 0 · 🔗 github.com
    ExplainDB is a didactic DBMS implemented in Python, designed for readability rather than speed. It includes a full storage hierarchy, B⁺-tree and bitmap indexes, Bloom filters, MVCC with journaling, query operators, external merge sort, and join-order optimization. The repository provides Jupyter notebooks that walk through each concept, from data layout to Z-order curves, and can be run directly in the browser via Binder. Created by Prof. Dr. Jens Dittrich's group at Saarland University, it is licensed under AGPL-3.0.
  75. Ai2's 8B open model matches proprietary report quality while running 3.5× faster — ▲ 15 · 💬 1 · 🔗 allenai.org
    Ai2 open-sources AstaBrief 8B, a model that turns a research question and retrieved literature excerpts into a cited report. Trained on 47K real queries with SFT and DPO, it matches the quality of Claude-powered pipelines while cutting generation time from 178.5 to 51.1 seconds. Weights, training data, and a local workflow are released so institutions can run it on sensitive research.
  76. The Quest for Limited Liability Is Ancient, Not a Modern Revolution — ▲ 13 · 💬 6 · 🔗 scholarlycommons.law.wlu.edu
    Robert W. Hillman argues that the recent emergence of LLCs and LLPs is evolutionary, not revolutionary. For as long as commerce has existed, merchants have sought to limit and shift liability, from Roman noxal liability and the peculium to modern business forms. The article samples diverse historical periods to show that the drive for limited liability is an ancient activity in the law, and that ignoring this history leads commentators to excess excitement over supposedly new developments.
  77. AI Is Just Normal Technology, Not a Superintelligent Species — ▲ 13 · 💬 2 · 🔗 knightcolumbia.org
    A new essay argues that AI should be seen as normal technology, like electricity or the internet, rather than a separate, superintelligent entity. The authors contend that transformative impacts will take decades because adoption and diffusion are slow, especially in safety-critical domains. They reject both utopian and dystopian visions, advocating for resilience and reduced uncertainty in policy instead of drastic interventions based on speculative superintelligence risks.
  78. Pi Codex Connectors - Use every Codex connector from any Pi model — ▲ 12 · 💬 0 · 🔗 github.com
    Pi Codex Connectors is a Pi extension that lets you access all your connected Codex (ChatGPT) connectors—GitHub, Gmail, Google Calendar/Drive, Slack, Linear, Figma, and more—directly from any Pi model. It starts a local Codex app-server, builds a catalog of available connectors and tools, and exposes just three simple tools to the model, avoiding schema bloat. Writes require approval by default, and privacy controls let you restrict or deny access. Install via npm and start using your existing connectors seamlessly.
  79. The Real Luddites Were Fighting for Their Craft, Not Against Machines — ▲ 11 · 💬 1 · 🔗 articlesofinterest.substack.com
    Knitter and audio producer Elizabeth Kauma traces the intertwined history of labor and fiber arts, from medieval knitting guilds to the framework knitting machine that displaced handknitters. She explains why crochet, which cannot be machine-made, became a Victorian craze, and offers tips for spotting imitation crochet from fast fashion. The episode connects today's automation fears to the original Luddites, who fought for their livelihoods, not against technology itself.
  80. Kyoto Microcomputer's Partner-N64 Dev Kit: How Nintendo 64 Games Were Debugged — ▲ 10 · 💬 1 · 🔗 behindthecode.ca
    This is the PARTNER-N64 PC development kit for the Nintendo 64, made by Kyoto Microcomputer (KMC). It required a specially modified N64 console with tie-lines re-routing IRQ and RESET signals to the CPU, enabling real-time breakpoints and resets. The kit also supported the failed 64DD development system, using an IPL ROM PAK for FONT and WAV data. A DCC adapter board allowed parallel cartridge connections for testing. Includes an in-depth video demo.
  81. Underwater Umbrellas Shield Coral Reefs from Bleaching — ▲ 9 · 💬 4 · 🔗 gizmodo.com
    Scientists tested underwater umbrellas in the Florida Keys that cut light exposure by 60% and reduced bleaching in brain corals during a record marine heatwave. Some shaded corals even regained their color. Lead author Karen Neely says the design works for small areas like individual high-priority corals or nurseries, and larger trials are already underway.
  82. Scientists Crack Open the 'Undruggable' Cancer Gene Behind 70 Percent of Tumors — ▲ 9 · 💬 5 · 🔗 nytimes.com
    MYC, hyperactive in 70 percent of cancers, has long been considered impossible to drug. A new study in Molecular Cell from Fox Chase Cancer Center identifies two molecular switches that control how extra copies of MYC accumulate, suggesting it may be easier to target those switches than MYC itself. The work is basic science, but it offers a fresh angle on a gene researchers call the grand orchestrator of cancer.
  83. AI agents filled only 21 of 138 missing Iran data fields, but 64 of 130 for the US — ▲ 9 · 💬 0 · 🔗 royapakzad.substack.com
    A human rights researcher tested three AI agents—Muse, Claude Cowork, and GPT 6.1 Sol—on filling missing World Bank procurement data for the US and Iran. For Iran, GPT and Muse each filled only 21 of 138 N/A fields, versus 51 and 64 of 130 for the US. Official government sources made up 76–89% of US citations but only 11–22% for Iran. Muse even registered on the portal as david.jones@gsa.gov without showing terms of service.
  84. IEA's 2026 Electrification Report Lands on Hacker News — ▲ 9 · 💬 0 · 🔗 iea.blob.core.windows.net
    The International Energy Agency's Electrification Special Report 2026 was posted to Hacker News, drawing early discussion on the global shift toward electric power. The linked PDF, hosted on the IEA's Azure storage, outlines the agency's outlook for electrification across economies and sectors. The submission by user gmays had gathered 9 points within two hours, though the page itself could not be fetched.
  85. Graphene - Data analytics framework for coding agents — ▲ 7 · 💬 1 · 🔗 github.com
    Graphene transforms your coding agent into a world-class data analyst. It's an everything-as-code analytics framework that combines a powerful semantic layer with a dashboard file type, enabling agents to build accurate queries and polished visualizations 10x faster. With Graphene SQL, you get the governance of metrics and modeled joins without sacrificing SQL's full power. The CLI-driven design ensures token efficiency and agent ergonomics. Graphene supports Snowflake, BigQuery, ClickHouse, Postgres, MotherDuck, and DuckDB. Free forever, your business logic stays in your repo, never locked in. Try the demo project and see how Graphene raises the floor for agent-generated analytics.
添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论