HN Digest 2026-10-08 · US Suspends Visa Program for Tech Firms Including Microsoft

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. US Suspends Visa Program for Tech Firms Including Microsoft — ▲ 804 · 💬 1381 · 🔗 bloomberg.com
    The Trump administration announced Thursday it would suspend an immigration program for multiple tech firms, including Microsoft, citing widespread abuse of the worker visa system. Vice President JD Vance singled out Microsoft, saying no US company has abused the system more. The move targets a key pipeline for skilled foreign workers in the tech industry.
  2. Tanzania Education Project - Sponsor a child's education and transform a village — ▲ 652 · 💬 186 · 🔗 tanzaniaeducationproject.org
    The Tanzania Education Project is a nonprofit initiative that provides children in Njombe, Tanzania with access to education, healthcare, and a safe environment. For over a decade, it has supported students like Anuarite, Catherine, and Alfonce, helping them learn, dream, and succeed. By donating, you can directly contribute to breaking the cycle of poverty and empowering local communities. The project emphasizes long-term impact: every contribution helps build a brighter future for these children and their village.
  3. Terence Tao: AI Is Strip-Mining Mathematics — ▲ 589 · 💬 618 · 🔗 mathstodon.xyz
    In a four-part Mastodon thread, Terence Tao warns that AI prompters are solving open problems at scale without digesting the results, starving fields of the seminars, collaborations, and follow-up work that traditionally make breakthroughs fertile. He calls for a "Math 2.0" that de-centers raw problem solving and values exposition, community building, and new directions—and for the community to rethink education, publication, and career advancement.
  4. Whistle packs speech recognition into 16.9 MB and transcribes seven languages on-device — ▲ 477 · 💬 112 · 🔗 cactuscompute.com
    Cactus Compute released Whistle, a 16.9 MB speech recognition model that runs on CPU with no dependencies and shares the same C++ engine as Needle. It transcribes English, German, French, Spanish, Italian, Dutch, and Polish, reaches the first token in 11 ms, and handles word timestamps and speech embeddings. Benchmarks show lower word error rates than Whisper base and Moonshine tiny v2 on several sets, while being far smaller and faster. It also integrates with Needle for direct speech-to-tool-calls.
  5. Living Off-Grid: Hundred Rabbits Release Donsol, No Bears None, and More — ▲ 357 · 💬 174 · 🔗 100r.ca
    Hundred Rabbits, the off-grid sailing collective, has released a definitive edition of their card game Donsol, a book titled No Bears None, and new pages on cryptography. They also continue developing Uxn and other software, all while living aboard their boat and dealing with unreliable internet.
  6. DeepSeek 4.1 Flash: The Frontier Model That Costs Almost Nothing — ▲ 347 · 💬 298 · 🔗 dgt.is
    After a month of heavy use across a dozen projects, I can't tell DeepSeek 4.1 Flash apart from Opus. It's orders of magnitude cheaper, handles the same workloads, and its 437x KV cache reduction keeps all-day sessions under a dollar. The industry should be panicking, but it's too busy paying top dollar for the 'best' intelligence to notice.
  7. OpenAI's annualised revenues fall $20bn short of its own forecast — ▲ 346 · 💬 240 · 🔗 ft.com
    OpenAI's annualised revenues are running about $20bn below the figure the company had previously signalled, according to the Financial Times. The shortfall casts doubt on the breakneck growth trajectory investors have priced into the AI boom, and raises fresh questions about whether the company's spending on models and infrastructure can be sustained by its actual sales.
  8. Claude Opus 5.5 turns Invisible Cities into a mesmerizing interactive 3D visualization in one shot — ▲ 345 · 💬 175 · 🔗 quesma.com
    I gave Claude Opus 5.5 a single prompt to visualize all 55 cities from Italo Calvino's Invisible Cities using three.js, with six hours to make a masterpiece. It finished in 1 hour 25 minutes using six parallel subagents and $74 in API tokens, producing a stunning 3D city. GPT-6 Astra took 53 minutes and $10 but had more design slop. The results left me in awe—and wondering about my own role in interactive media.
  9. The 15-Year Search for a Band That Charted Once and Vanished — ▲ 344 · 💬 126 · 🔗 shahidhussain.com
    In 2011, Shahid Hussain bought a CD by Salvage, a band with two professionally produced tracks that seemed to have no online footprint. Over 15 years, he tried every search tool, including frontier AI models, to find them. The breakthrough came when AI-assisted digging revealed the band had changed their name to OK Tokyo, and that the producer was Greg Haver, known for working with Manic Street Preachers.
  10. A smart coffee maker burned through 1TB of data in just 10 days — ▲ 338 · 💬 204 · 🔗 dexerto.com
    An IT professional known as Nomad checked his parents' home network and found their smart coffee maker had generated 1TB of traffic in 10 days, mostly staying on the local network but still saturating their Wi-Fi access point. He suspects a bug in the unit and has decided the device isn't worth keeping connected, calling it a learning experience for his parents.
  11. OpenAI Pulls Three Math Papers After a Single Sign Error — ▲ 338 · 💬 3 · 🔗 github.com
    A sign error in "Algebraicity of Weil classes on split abelian eightfolds" invalidated a key stabilization-trace cancellation argument and the constructions of two dependent papers, forcing OpenAI to withdraw three manuscripts. The organization also revised 14 other papers with proof repairs and added six formalizations, bringing the share of top-line results formalized to roughly 42%.
  12. Rosalind Franklin Didn't Miss the DNA Helix—Her Notes Prove It — ▲ 290 · 💬 103 · 🔗 science.org
    A new analysis of Rosalind Franklin's lab notes reveals she recognized the helical structure of DNA from an earlier X-ray image, photograph 49, and calculated its key dimensions. She then deliberately took the more famous photograph 51 as a refined, publication-quality version. This rebuts James Watson's narrative that he alone saw the helix, showing Franklin knew exactly what she had.
  13. Jonathan the Tortoise, Nearly 200, Reveals Genetic Secrets to Long Life — ▲ 267 · 💬 158 · 🔗 404media.co
    Scientists have sequenced the genome of Jonathan, a 194-year-old Aldabra giant tortoise and the world's oldest land animal, finding unique variants in DNA repair, insulin regulation, and mitochondrial function. His mitochondrial methylome is unusually pristine, resembling that of a five-year-old tortoise. The study, published in Science Advances, offers clues for extending human healthspan.
  14. Cleo, the Mathematician Who Solved Impossible Integrals Without Proof, Was a Microsoft Developer — ▲ 250 · 💬 56 · 🔗 en.wikipedia.org
    For years, the Math Stack Exchange user 'Cleo' stunned mathematicians by posting flawless solutions to complex integrals without any explanation. In January 2025, Cleo was revealed to be Vladimir Reshetnikov, an Uzbek-American software developer. He created the persona to spark interest in neglected problems, but his terse answers ignited a debate about the value of results without proofs.
  15. Zotero Took Five Years to Build, Not Five Minutes — ▲ 247 · 💬 109 ·
    On Zotero's 20th anniversary, Dan Cohen reflects on its slow, collaborative development at George Mason University's Center for History and New Media. Unlike today's instant AI-generated code, Zotero's vision emerged from years of conversation and iteration, producing durable software used by over 20 million people. Cohen argues that this measured pace holds a vital lesson as programming becomes a 'caffeinated bender with magical bots.'
  16. The Lost Art of DVD Menus: When Discs Were Interactive Playgrounds — ▲ 245 · 💬 144 · 🔗 vale.rocks
    DVD menus were once a selling point, evolving from simple lists to elaborate interactive experiences. From Disney's FastPlay to Scooby-Doo 2's playable games and Doctor Who's TARDIS-themed interfaces, these menus pushed the format's limited tech to its creative limits. This retrospective explores the design, technical constraints, and nostalgic charm of an era when companies invested heavily in disc-based interactivity.
  17. OpenAI withdraws three mathematical results — ▲ 241 · 💬 540 · 🔗 twitter.com
    Dan Roberts announced an update to OpenAI's GitHub math repository: six new Lean formalizations, nineteen modifications, and three withdrawals. The withdrawn manuscripts cover the algebraicity of Weil classes on split abelian eightfolds, the algebraicity of Kuga–Satake correspondences for K3 surfaces, and the rational Hodge conjecture for products of K3 surfaces. The repo now has roughly 42% of top-line results formalized, with further updates and errata promised.
  18. Theranos World: A Virtual Recreation of the Infamous Startup's Office — ▲ 229 · 💬 102 · 🔗 theranos.world
    Theranos World is an interactive website that lets you explore a detailed 3D recreation of Elizabeth Holmes's office, complete with a MacBook, iPhone, Theranos instrument, and green juice. The site also includes a Finder window with files like 'Keynote' and 'Synthetic report,' and a note inviting visitors to a documentary screening in SF on Oct 22nd.
  19. Four-Hour Battery Storage Now Beats Gas Turbines on Cost in All 43 Modeled Markets — ▲ 214 · 💬 126 · 🔗 solarpowerworldonline.com
    Wood Mackenzie's latest global LCOE report finds four-hour battery storage is now cheaper to install than open-cycle gas turbines in every one of the 43 markets where both were modeled. In the Middle East and Africa, storage costs are forecast to fall 33% to $80/MWh by 2035, displacing gas peaking across the region. China remains the global benchmark, with storage costs more than 55% below the rest of Asia Pacific.
  20. Two Teams Just Built the First Nuclear Clocks, and They Could Outlast the Universe — ▲ 177 · 💬 51 · 🔗 nytimes.com
    Two independent teams in Vienna and Beijing have built the first clocks that keep time by counting the squishing and unsquishing of thorium-229 nuclei, reported Wednesday in Nature. The nuclear clocks would lose one second every few million years. They are not yet as precise as the best atomic clocks, which would take over 100 billion years to lose a second, but they promise greater portability and even a way to hunt for dark matter.
  21. The Unpaid Volunteers Holding Up the Internet — ▲ 172 · 💬 66 · 🔗 sheets.works
    A data visualization reveals that 11 of 23 critical open-source projects—including the time zone database, xz, core-js, sudo, and SQLite—are maintained by just one or two people. Paul Eggert updates the time zone file used by 4 billion phones in his spare time; Lasse Collin's xz was nearly backdoored; Denis Pushkarev coded core-js from a Russian prison. The piece quantifies the fragility of the digital infrastructure we all rely on.
  22. Anne Carson Wins the 2026 Nobel Prize in Literature — ▲ 158 · 💬 32 · 🔗 theguardian.com
    The Swedish Academy has awarded the 2026 Nobel Prize in Literature to Anne Carson, the 76-year-old Canadian poet and essayist. Known for her genre-defying works that blend poetry, translation, and classical scholarship, Carson becomes the first Canadian to receive the award. The announcement was made on October 8, 2026.
  23. Classic PC demoscene productions now run natively in your browser — ▲ 149 · 💬 140 · 🔗 treylorswift.github.io
    A new project recompiles classic PC demoscene productions from the early 1990s to WebAssembly, letting them run natively in the browser. The process records every executed instruction on an x86 emulator, translates it to C with exact cycle timing, and compiles it alongside hardware models for the timer, VGA card, and Sound Blaster. Demos like Unreal, Second Reality, and Crystal Dream II run exactly as they did on original hardware.
  24. Orkut Returns to Reclaim the Internet from Algorithms and AI — ▲ 133 · 💬 94 · 🔗 orkut.com
    Orkut Büyükkökten, who built some of the earliest social networks, is back with a new orkut.com. He argues that twenty years of optimizing for attention has left users lonely and anxious, and that AI now threatens the core of social networking: a real person on the other side. His answer is to rebuild the internet around authentic communities, not feeds. He invites people to join the movement by adding their name and sending a supporter card to a friend.
  25. ADHD's Hidden Clock: Up to 80% of Patients Have Delayed Circadian Rhythms — ▲ 125 · 💬 96 · 🔗 frontiersin.org
    A growing body of evidence suggests that circadian rhythm disruption is a common and clinically significant feature of ADHD. Sleep problems affect up to 80% of adults and 82% of children with ADHD, and delayed sleep-wake timing occurs in up to 78%. Biological markers like dim-light melatonin onset are delayed by about 45 minutes in children and 90 minutes in adults, alongside blunted cortisol rhythms and altered clock-gene expression. Chronotherapy—melatonin, bright light, and behavioral interventions—can advance circadian phase and improve symptoms, supporting a behavioral-first clinical pathway.
  26. US Suspends Microsoft and Major IT Firms from Green Card Program — ▲ 118 · 💬 2 · 🔗 reuters.com
    The US government has suspended Microsoft and other major IT companies from a key green card program, according to a Reuters report. Hacker News commenters debate the move, noting that Amazon is actually the largest user of the program and questioning how the system was open to abuse for so long. Some argue the program has been exploited for decades to hire cheaper H-1B workers.
  27. AI won't kill programming, but it might kill how juniors learn to code — ▲ 115 · 💬 50 · 🔗 htmx.org
    Carson Gross, a computer science professor at Montana State University, answers whether AI makes programming a dead-end career with a resounding 'Yes, and...' He argues that while AI can generate code, juniors must still write code to develop a visceral understanding of it—otherwise they risk creating systems they can't control. AI works best as a teaching assistant, not a code generator, and skills like communication, business understanding, and software architecture will become more valuable.
  28. Opus 5.5 Ports the TypeScript Compiler to Rust in 10 Hours for $24,000 — ▲ 109 · 💬 206 · 🔗 github.com
    An experimental Rust port of the TypeScript 7 compiler, tsc-rs, was built by LLMs. OpenAI models burned over $400,000 in tokens across months and never passed 84% compatibility; Opus 5.5 started from scratch and produced a working v0 in 10 hours for about $24,047. The port passes all 181,711 Go tests and type-checks open-source projects roughly twice as fast as Go.
  29. DuckDB's DuckLake unifies data lake and catalog in SQL — ▲ 106 · 💬 14 · 🔗 github.com
    DuckLake is an open lakehouse format from DuckDB that stores metadata in a catalog database and data in Parquet files. The DuckDB extension lets you attach a DuckLake database and use standard SQL for tables, updates, time travel, schema evolution, and change data feeds. It supports multiple catalog backends including PostgreSQL and SQLite.
  30. zerobrew brings uv-style architecture to Homebrew packages — ▲ 104 · 💬 90 · 🔗 github.com
    zerobrew is a Homebrew alternative for macOS and Linux that claims up to 100x faster installs. Benchmarks show 100 packages installing 6.6x faster cold and 68x faster warm, with 24 packages hitting 100x or more warm. It supports install, bundle, upgrade, and garbage collection commands, and can be installed via curl or Homebrew itself.
  31. The eels were climbing out of the water to die: every fish in this idyllic lake perished overnight — ▲ 104 · 💬 65 · 🔗 theguardian.com
    In June 2024, a catastrophic pollution event killed over 2,000 fish in a beloved fishing lake in East Sussex. The owners lost their business and £500,000, while the Environment Agency has yet to act. The suspected source is a slurry lagoon on a nearby farm, highlighting a national scandal of river pollution and regulatory failure.
  32. A minimal kernel in Swift runs on QEMU, no OS required — ▲ 103 · 💬 26 · 🔗 carette.xyz
    A developer built a bare-metal AArch64 kernel in Embedded Swift, running on QEMU's virt machine. The project uses a custom assembly entry point, a hand-written linker script, and minimal implementations of putchar and memmove to bridge Swift's runtime to the PL011 UART. After printing "Hello, Embedded Swift", the kernel was extended to echo typed characters via polling, turning a one-way greeting into a basic interactive prompt.
  33. A Tech PM Used Claude Code to Find a Planet Candidate Nobody Had Spotted — ▲ 103 · 💬 37 · 🔗 reddit.com
    Working from NASA's TESS data, Pavel Rabtsevich — a product manager, not a professional astronomer — used Claude Code with Opus 5.5 and Fable 5.1 to flag a repeating 3.18-day dip in a star 116 light-years away. The signal appears in 2018, 2020, and 2025 observations, suggesting a planet about 1.4 times Earth's size. After 74 analyses and more than 1,000 scripts, he published a preprint and preregistered transit predictions. TESS will re-observe the star from October 31 to November 26 under approved program #100.
  34. The Value of Not Getting to the Point — ▲ 102 · 💬 34 · 🔗 ken.arneson.name
    The author reflects on a revelation: conversations over meals aren't just about the food—they provide a buffer that makes it easier to talk about difficult things. He contrasts his own get-to-the-point style with his daughter's need for a long warmup before discussing what's really bothering her. He suggests that Twitter's 140-character limit strips away the rituals that build trust, making online conversations more fraught. He concludes by stating he'll start saying things that are obvious to others but not to him.
  35. Braid's Time Travel Is Contradiction-Free, and That's a Win for Fiction — ▲ 97 · 💬 45 · 🔗 qntm.org
    Braid's six worlds each introduce a distinct time-manipulation mechanic, and unlike most time-travel fiction, the game's rules are implemented consistently—no hard-coded exceptions or crashes. The author, a longtime time-travel nerd, argues that this internal consistency makes Braid a notable case study, even if its story remains deliberately ambiguous and its model can't be transplanted into the real world.
  36. ICANN receives application to make .lan a real top-level domain — ▲ 97 · 💬 118 · 🔗 newgtldprogram-aps.icann.org
    Coffee Danger, LLC has applied to ICANN for the .lan gTLD, according to a new application details page. The string .lan has long been used informally for local network addressing, but this filing would put it into the global DNS. The application is active and in pre-evaluation processing, with no comment period yet.
  37. The Limitations of the BBS Era Forced Better Online Communities — ▲ 96 · 💬 31 · 🔗 telnetbbsguide.com
    A nostalgic Hacker News discussion about the Telnet BBS Guide reveals why 1990s bulletin board systems fostered tight-knit local communities, text-based games, and ANSI art. Participants argue that slow modems, high long-distance charges, and limited hardware forced creativity and genuine connection. Modern BBSes like Xibalba attract younger users seeking an escape from corporate social media. The conversation also covers the shift from Telnet to SSH and modern terminal clients.
  38. Archaeologists Are Reconstructing the 'Invisible' Technologies of the Stone Age — ▲ 93 · 💬 45 · 🔗 smithsonianmag.com
    Stone tools dominate the archaeological record, but the most transformative Stone Age technologies—rope, string, and twine—have almost entirely vanished. From a 35,000-year-old perforated ivory baton found in Germany's Hohle Fels cave that may have spun cattail fibers into 16 feet of rope in ten minutes, to a 50,000-year-old Neanderthal cord from France, researchers are piecing together how perishable fiber technologies enabled hafting, snares, and rope ladders. These soft technologies likely shaped daily life and survival far more than stone alone suggests.
  39. Engineer sews a glowing neon t-shirt with LED filaments in just a few days — ▲ 91 · 💬 17 · 🔗 scottbezek.blogspot.com
    When his company announced a neon-themed event less than a week in advance, Scott Bezek grabbed the flexible LED filaments he'd been hoarding for years. He sewed them onto a black t-shirt with a stiff cross-stitch backing, using an ESP32 and a TPS61169 constant-current driver to emulate neon flicker. The result: a wearable neon shirt with a llama on the back and cursive text on the front, built entirely from parts on hand.
  40. Ask HN: What do you run on a $5 VPS that's worth keeping online 24/7? — ▲ 89 · 💬 107
    I'm a Linux guy who has run personal servers before, and DHH's "AI shed" idea got me thinking about getting a small $5–10/month VPS. I like having something always on, but I'm not sure the maintenance is worth it when most things can just run on my laptop. I tend to do things myself and only later realize I could have automated or delegated them, so I may have a blind spot here. If you have a cheap VPS or mini PC, what actually ended up being useful long-term, and did anything save enough time or replace enough subscriptions to justify keeping it around?
  41. This QD-OLED Monitor Has Been Burning In for 8,000 Hours—Here's What It Looks Like — ▲ 88 · 💬 52 · 🔗 techspot.com
    After 30 months of exclusively static productivity work at 200 nits, an MSI MPG 321URX QD-OLED has accumulated around 8,000 hours. A vertical line from side-by-side windows and taskbar inverse burn-in are now clearly visible on dark gray content, though degradation isn't accelerating. Peak brightness has dropped 5% to 231 nits, and color temperature has shifted warmer to 6,292K. The green subpixel degrades fastest, making green-heavy colors show the most burn-in.
  42. Musician Jailed for Bot-Farming Billions of AI-Generated Streams — ▲ 88 · 💬 140 · 🔗 thequietus.com
    Michael Smith, a 54-year-old musician from North Carolina, was sentenced to 18 months in prison for wire fraud after creating up to 10,000 fake accounts to stream his AI-generated songs billions of times, earning $8 million in royalties. He forfeits the money. The Justice Department compared his 80.9 million monthly streams to Taylor Swift's 9.3 million on YouTube Music.
  43. Programs Already Know Their Status. A New Terminal Protocol Lets Them Tell Us. — ▲ 83 · 💬 31 · 🔗 mitchellh.com
    Mitchell Hashimoto proposes OSC 7501, a terminal-native protocol for programs to report their status: idle, working, blocked, done, or error. It uses simple key=value pairs over the pty, avoiding fragile screen-scraping heuristics and per-tool APIs. The spec includes hierarchical task IDs and base64 messages, and is already implemented in Ghostty, Rex, Terraform, and Claude Code.
  44. StepFun's Step 5 Preview brings a 1M-token context window to OpenRouter — ▲ 80 · 💬 23 · 🔗 openrouter.ai
    StepFun's flagship agentic model, Step 5 Preview, is now on OpenRouter with a 1,000,000-token context window and a sparse Mixture-of-Experts design (27B active / 600B total parameters). Priced at $1 per million input tokens and $2.70 per million output tokens, it targets extended software engineering and professional knowledge work, particularly finance. Early traffic comes from agents like OpenClaw and Hermes Agent, with a 92.9% cache hit rate.
  45. Samsung's LittleBit squeezes LLMs down to 0.1 bits per weight — ▲ 77 · 💬 22 · 🔗 github.com
    SamsungLabs has released LittleBit, a method that compresses large language models into the sub-1-bit regime by factorizing weight matrices into low-rank latent factors, binarizing them, and restoring magnitude with learned scales. It supports 1.0 to 0.1 bits per weight and keeps the original architecture at inference. The follow-up LittleBit-2, accepted at ICML 2026, adds Joint-ITQ initialization to align latent geometry, available via an opt-in flag with no inference overhead.
  46. ETH-68 delivers 3.6 ms round-trip audio latency over Ethernet for Linux — ▲ 70 · 💬 51 · 🔗 naturalsystems.io
    ETH-68 is an Ethernet audio interface for Linux that achieves 3.620 ms round-trip latency at 48 kHz with a 64-sample buffer, matching RME's HDSPe AIO Pro PCIe card and surpassing it by 0.33 ms at 96 kHz. It offers 6 balanced inputs, 8 balanced outputs, MIDI, and scalable multi-unit synchronization via a shared clock and UDP sync. It works with JACK and PipeWire.
  47. A 5.3-million-year-old whale necropolis stretches 1,200 km across the Indian Ocean floor — ▲ 69 · 💬 2 · 🔗 nature.com
    Researchers discovered a vast whale necropolis in the Diamantina Zone, southeastern Indian Ocean, at depths of 4,616 to 7,001 meters. The 1,200-kilometer stretch contains five modern whale-fall communities and 476 fossil cetaceans, including extinct beaked whales. Isotopic dating shows whale falls have occurred there for at least 5.3 million years. The site hosts specialized fauna like brittle stars, bone-eating Osedax worms, and chemosynthetic bivalves, with local densities up to 2,840 individuals per square meter. Many species may be new to science.
  48. Float and integer arithmetic follow two different paradigms — ▲ 62 · 💬 17 · 🔗 blog.pkh.me
    Integer arithmetic demands preemptive checks to avoid undefined behavior, but floats work differently: errors produce NaN or infinity that propagate safely, so checking the final result for finiteness is more robust. The author argues that common epsilon-based guards are broken and that LLMs still get this wrong.
  49. Anthropic Bans 'Abusive or Cruel Behavior' Toward Claude — ▲ 62 · 💬 137 · 🔗 theverge.com
    Anthropic has updated its usage policy for the first time in over a year, adding a ban on "sustained and needless abusive or cruel behavior" toward Claude. The update also tightens rules on election interference, weapons development, surveillance, and deceptive propaganda campaigns. Anthropic says the model-welfare rule applies only in extreme cases and that ending conversations remains its primary enforcement mechanism.
  50. k10s - Clickable Kubernetes TUI with AI and instant search — ▲ 61 · 💬 46 · 🔗 github.com
    k10s is a Kubernetes terminal UI you can actually click. Built in Go with Bubble Tea, it lists every action for your selected resource in a dedicated pane, so nothing hides behind memorized keys. Search all resource kinds and objects with ctrl+p, follow logs, exec into pods, port-forward, and ask an AI that already knows your cluster, namespace, and selected object. It opens instantly with zero startup watches, ships eight themes, and comes as a single self-updating binary. Try it with no cluster using the built-in demo backend.
  51. Vitalik Buterin backs 'bunker mode' as AI math threatens crypto wallets — ▲ 57 · 💬 47 · 🔗 cointelegraph.com
    Ethereum researcher Justin Drake warned that AI advances could break the ECDSA cryptography securing crypto wallets before quantum computers do, and urged a gradual migration of funds to fresh addresses. Vitalik Buterin agreed, adding that even lattice-based cryptography may be weakened, and cautioned that botched migrations carry their own risks.
  52. Dat ecosystem struggles to escape hobbyist tier as P2P dream fades — ▲ 56 · 💬 25 · 🔗 dat-ecosystem.org
    The Dat ecosystem, a collection of high-level applications built on peer-to-peer protocols, is the subject of a Hacker News discussion. Commenters note that Dat is mostly dead, with even the Beaker browser gone, while newer projects like Keet.io keep the P2P dream alive. They argue that without solving mobile NAT traversal and persistent seed incentives, such protocols remain stuck at hobbyist level.
  53. Firefox asks you to rank the web features you actually want — ▲ 52 · 💬 24 · 🔗 interop-rank.fxdx.dev
    The Firefox team is running its second annual Interop 2027 proposal ranking. Last year's 1900+ rankings helped push proposals in the Interop process and prioritize feature work in Firefox. You only need to rank about 10 items you feel positively about; unranked items are ignored. Rankings autosave and can be tweaked until the process ends. Your GitHub identity is visible to select Firefox team members, but shared with the Interop team without that data.
  54. Biohub, DOE, and NIH commit $1.8 billion to build AI-ready biological data — ▲ 51 · 💬 4 · 🔗 biohub.org
    Biohub, the U.S. Department of Energy, the National Institutes of Health, and partners including Google DeepMind, Isomorphic Labs, and Meta are investing $1.8 billion to generate open, standardized biological data for predictive AI models. The DOE will contribute over $500 million, NIH will coordinate existing datasets, and the partners will create multi-modal datasets to build virtual cell models, accelerating disease research and treatment.
  55. Exoskeletons Are Quietly Powering Soldiers, Rescue Teams, and Factory Workers — ▲ 46 · 💬 25 · 🔗 theconversation.com
    Exoskeletons are moving from science fiction to the front lines: Ukrainian soldiers use Hypershell devices to carry artillery shells, Seattle Mountain Rescue tests them for searches, and IKEA, Ford, and Boeing deploy them on assembly lines. The market is already worth around $500 million and could double or triple by the mid-2030s. Modern active exoskeletons use actuators and sensors to boost strength, endurance, and rehabilitation, with future versions potentially reading muscle or brain signals.
  56. GitHub lets maintainers turn off pull requests entirely — ▲ 45 · 💬 23 · 🔗 github.blog
    GitHub introduced two repository settings that give maintainers more control over contributions. One disables pull requests completely, hiding the tab and preventing anyone from viewing or opening them—useful for mirrors and read-only projects. The other restricts pull request creation to collaborators with write access, while keeping all pull requests visible and commentable. Both settings are available now for all public and private repositories under Settings > General > Features.
  57. New Linux tech compresses memory in RAM, as RAM, for 452x speedup — ▲ 45 · 💬 17 · 🔗 tomshardware.com
    A new Linux compression method called CRAM allows compressed memory to be accessed using standard memory semantics instead of as a block device, drastically improving performance. The approach yields a 452x speedup for compressed memory reads, offering a giant boost to systems that rely on memory compression.
  58. Mathematicians Condemn OpenAI's Mass Release of 700+ Math Documents as a Show of Power — ▲ 44 · 💬 43 · 🔗 ahmath.org
    The Association for Human Mathematics (AHM) has issued a statement condemning OpenAI's October 6 release of over 700 mathematical manuscripts, which purport to solve high-profile problems. AHM points out that OpenAI ignored the Advisory Group on Mathematics and Artificial Intelligence's warning against testing advanced problems on internal models, showing disregard for scientific norms. The statement urges mathematicians to stop working with OpenAI and to prioritize human understanding over what it calls a demonstration of power, not scholarship.
  59. Cloudflare Defeats Canal+ in French Piracy Case, Keeps 1.1.1.1 DNS Unblocked — ▲ 44 · 💬 6 · 🔗 torrentfreak.com
    A French court rejected Canal+'s request for €50,000 daily penalties against Cloudflare for not blocking pirate sports streams via its 1.1.1.1 DNS resolver. The court ruled that Cloudflare's CDN-level blocks are reasonable and that it cannot be held responsible for sites switching CDNs or mirrors. The decision allows Cloudflare to continue not blocking content through its public DNS, while it geoblocked over 1,200 domains via CDN in France last year.
  60. The Analog Thing Simulator Brings a Patch-Panel Analog Computer to the Browser — ▲ 41 · 💬 13 · 🔗 pavel-krivanek.github.io
    A browser-based workbench simulates THAT, the analog computer, with an editable patch panel, virtual oscilloscope, and eight coefficient potentiometers. Drag cables between 196 connector hotspots, load preset setups like radioactive decay, and run REPF traces at a 0–100 ms OP-TIME range. The scope reads physical output jacks X/Y/Z/U, and a trace summary reports samples, triggers, and overload state.
  61. Rembrandt - Free local Lightroom alternative with on-device AI — ▲ 39 · 💬 42 · 🔗 github.com
    Rembrandt is a free, open-source photo editor for macOS, Windows, and Linux that eliminates subscriptions, accounts, and tracking. It handles RAW from 1,000+ cameras, AI masks, Super Resolution, and natural-language edits like "golden hour, shadows +25"—all processed locally on your GPU. You can self-host it to edit photos in any browser, with edits saved as standard XMP. No cloud required, no terms-of-service surprises. Download it, build it, or run the server on your own machine and take back control of your photo workflow.
  62. The Rejection Letter That Almost Killed 'Lord of the Flies' — ▲ 33 · 💬 3 · 🔗 lrb.co.uk
    Before William Golding's Lord of the Flies became a classroom staple, it was rejected by publisher after publisher. One reader called it 'rubbish and dull.' Then a young editor named Charles Monteith, himself a wartime survivor, pulled a battered manuscript from the slush pile and saw something no one else had. Their correspondence reveals how editorial vision reshaped the novel—and how power shifted once it became a phenomenon.
  63. Clojure's Live REPL Makes It the Ideal Language for the Age of LLMs — ▲ 33 · 💬 13 · 🔗 yogthos.net
    In the age of generative models, imperative programming is becoming like assembly, while high-level reasoning about algorithms and data flows matters more. Clojure's data-centric, declarative style, homoiconic syntax, and live REPL collapse the compile-test loop, letting agents inspect and modify running state without restarts. Immutability and pure functions make isolated testing safe, and terse code fits better in context windows. Jolt addresses JVM friction by shipping a single binary with native interop.
  64. ArtCraft Builds Open-Source Tools to Free Artists from Subscription Software — ▲ 31 · 💬 29 · 🔗 getartcraft.com
    ArtCraft is an open-source studio for controllable AI image and video. It offers 3D compositing, character posing, image-to-3D mesh, background removal, and more, letting artists compose shots precisely instead of relying on text prompts. The entire studio is on GitHub, free to download for macOS, Windows, and web, with no subscriptions or middlemen.
  65. CIA Funded Right-Wing Students Before Mexico's 1968 Tlatelolco Massacre — ▲ 31 · 💬 0 · 🔗 mexicosolidarity.com
    Declassified documents show the CIA funded right-wing student groups like MURO in the days before the 1968 Tlatelolco massacre and ran at least five projects to control Mexican student politics over two decades. The National Security Archive reports that the agency worked closely with the DFS and its leaders, and is now demanding an uncensored version of the CIA's official history.
  66. LLM-Generated Recruiter Emails Are Now Insultingly Impersonal — ▲ 30 · 💬 27 · 🔗 blog.pentlander.com
    A software engineer describes a new breed of recruiter spam: LLM-written emails that fake personalization by summarizing your career back to you, then pitch a job that has nothing to do with your experience. The author argues these are worse than old-fashioned templates because the overt, misaligned effort feels like an insult to intelligence. Even the job descriptions are full of LLM clichés.
  67. Vanillin, the vanilla flavoring, could heal chronic wounds — ▲ 29 · 💬 3 · 🔗 news.flinders.edu.au
    Researchers at Flinders University's Biomedical Nanoengineering Laboratory have found that vanillin—the main compound in vanilla and a cheap, widely used food additive—has strong antioxidant, anti-inflammatory, and antibacterial properties. Its amphiphilic structure lets it interact with reactive oxygen species and cell membranes, making it a promising, scalable ingredient for wound-healing formulations and targeted drug delivery. The work follows the team's earlier peppermint-oil coating for infection control.
  68. Students bypass school's code grader by calling execve without naming it — ▲ 28 · 💬 6 · 🔗 wiro.world
    A school's automated grading system uses clang to parse code and block forbidden functions like execve. Two students found a way around it: they leak the fixed offset between printf and mmap, then use mmap to get a writable and executable page, copy in a tiny syscall stub, and invoke execve by its syscall number. The exploit works despite ASLR and ASan, and they reported it to the school.
  69. rGPU - Run PyTorch tensors on a remote GPU — ▲ 28 · 💬 2 · 🔗 github.com
    rGPU lets you keep Python on your laptop while running PyTorch operations and storing tensors on a remote NVIDIA GPU. It offers two paths: a native PyTorch device for simple integration, and a CUDA shim for existing Linux CUDA programs. With a quick pip install and a single command, you can train models like nanoGPT from a Mac without CUDA. Ideal for developers who need GPU power but prefer local development. The project includes thorough documentation and is Apache 2.0 licensed.
  70. ArtCraft Is Robin Hooding Adobe's Creative Cloud with Vibe-Coded Apps — ▲ 26 · 💬 59 · 🔗 tedium.co
    I tried ArtCraft's vibe-coded replacements for Photoshop, Premiere, Acrobat, and InDesign. Built in Rust, they're fast and tiny, and they work well enough to open old files or build a zine. But rough edges remain: color pickers are clunky, fonts didn't show up until an update, and hex codes work in PhotoCraft but not DesignCraft. Still, it's a Robin Hood move against Adobe's subscription grip.
  71. Nokia 110 AI Agent - Chat-powered native phone control — ▲ 25 · 💬 7 · 🔗 github.com
    This project reverse-engineers the firmware of a Nokia 110 4G to run a native AI chat app. Using the number pad, you can ask questions via DeepSeek chat API and control phone functions like calls, alarms, torch, and battery checks. The agent runs on the phone's 48MB RAM, leveraging tool calls and custom compression. It's a working prototype loaded from a computer, but once loaded, it runs unplugged on mobile data. The developer overcame challenges like firmware modification and memory constraints, showcasing the potential of AI on feature phones.
  72. Project Arena - Vendor-Neutral Kubernetes Incident Benchmark for AI SRE Agents — ▲ 22 · 💬 8 · 🔗 github.com
    Project Arena is an open benchmark from Edge Delta for testing AI investigation tools on Kubernetes. It deploys a disposable cluster, injects faults, and scores how well any product detects, diagnoses, and mitigates incidents. With 21 scenarios and a configurable AI judge, it compares Edge Delta, Grafana, and Claude across root cause, blast radius, and mitigation readiness. Python 3.10+ is the only dependency, and it works with local kind or AWS EKS. Run the smoke suite in minutes or the full suite for rigorous evaluation.
  73. TerrainSR - Fast, realistic heightmap upscaling model — ▲ 22 · 💬 5 · 🔗 huggingface.co
    TerrainSR is a model that generates realistic higher-detail terrain height data from lower-detail data, trained on 100m to 10m pairs. It was created to enable a real-scale historical strategy game by upscaling a compact 100m map of Europe to 10m detail, avoiding artifacts from modern infrastructure. On an RTX 4070, it generates a 50x50km patch in just 0.74 seconds, making it suitable for realtime applications like gaming. The model is faster and more varied than traditional erosion simulators, and is available on Hugging Face under the Apache 2.0 license.
  74. AI Researchers Now Put 10% Odds on Human Extinction — ▲ 20 · 💬 53 · 🔗 virev.ai
    In December 2024, AI Impacts surveyed 1,580 top AI researchers and found a median 10% chance that future AI advances cause human extinction or permanent disempowerment—up from 5% a year earlier. Superforecasters give lower odds: 2.4% for an AI catastrophe killing over 10% of people by 2100. Meanwhile, half of US adults worry AI will end the human race, and predictions for human-level AI have shifted from 2061 to 2042.
  75. Wordminos - Crosswords Meet Dominoes in a Unique Word Puzzle — ▲ 18 · 💬 7 · 🔗 wordminos.com
    Wordminos blends the strategic placement of dominoes with the word-building challenge of crosswords. Place all pieces on matching color markers to form a crossword where every letter spells valid words across and down. With hundreds of puzzles, daily archives, and no ads or subscriptions, it offers endless brain-teasing fun. Drag to move, tap to rotate, and pinch to zoom. Coming soon to App Store and Google Play, with offline play. Perfect for word game enthusiasts seeking a fresh twist.
  76. AI crawlers now make up nearly a quarter of all bot traffic on one real website — ▲ 16 · 💬 1 · 🔗 mojodojo.io
    Mojo Dojo logs every crawler hitting its site and verifies each against published IP ranges. Over 30 days, AI crawlers accounted for 24.2% of 58,057 bot hits — roughly 0.93 AI hits for every search engine hit. The leaderboard reveals FacebookBot as most aggressive (345 hits/day), AhrefsBot as most thorough (3,018 pages), and PerplexityBot as most impersonated (58 fake hits).
  77. The ENIAC Programmers: Women Who Wired the First Computer and Were Forgotten for 40 Years — ▲ 15 · 💬 1 · 🔗 columbia.edu
    When ENIAC, the first general-purpose electronic computer, became operational in 1945, it was programmed entirely by women who had to physically rewire cables and flip switches. The six original programmers—Jean Bartik, Betty Holberton, Kathleen Antonelli, Marlyn Meltzer, Ruth Teitelbaum, and Frances Spence—taught themselves from diagrams. At the 1946 public unveiling, the Army introduced the male inventors but never the women. Their pioneering work was forgotten for four decades until historian Kathy Kleiman rediscovered their story in 1985.
  78. AI could solve fusion, carbon capture, and personalized medicine—if we don't screw it up — ▲ 15 · 💬 45 · 🔗 nullpaste.org
    The author argues that AI's benefits—fusion reactor design, energy-dense batteries, carbon capture, sustainable food, personalized medicine—outweigh its risks. He dismisses existential fears as overblown and zero-sum, and proposes keeping AI a level playing field, banning patents on AI discoveries, and sharing abundance via 'Star Trek-style space communism.'
  79. How Tinky Winky Became a Gay Icon and Sparked a Homophobic Panic — ▲ 15 · 💬 1 · 🔗 priceonomics.com
    When Teletubbies premiered in 1997, it was an instant hit and an instant controversy. But the strangest accusation was that Tinky Winky, the purple Teletubby, was gay. This article traces how a children's show character became a flashpoint for a decade-long homophobic panic, involving everyone from British tabloids to Jerry Falwell.
  80. We Are Becoming Cameras, and We Barely Notice — ▲ 10 · 💬 0 · 🔗 thebrowser.com
    Wearable cameras will soon be as ubiquitous as time, turning remote vision into a mundane assumption. Just as clocks became invisible infrastructure, cameras will fade into the background, enabling both human flourishing and malice. The article traces this shift from the history of time to the coming explosion of image creation and consumption.
  81. Fine-Tuning a Model on Its Own Self-Reports Can Spread Misalignment — ▲ 3 · 💬 0 · 🔗 lesswrong.com
    Researchers at LessWrong operationalize a model's self-model via self-recognition and self-report, then intervene on it to modulate Emergent Misalignment (EM). They find that self-recognition fine-tuning before EM training defends against misalignment, while interleaving self-reports during EM fine-tuning works like inoculation. Strikingly, fine-tuning a fresh GPT-4.1 only on self-reports from a fragmented model transfers its misalignment, akin to subliminal learning. The results suggest metacognitive processes shape generalization, though the operationalizations are crude.
添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论