HN Digest 2026-09-30 · Opus 5.5 may be getting quietly worse — this benchmark is watching

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. Opus 5.5 may be getting quietly worse — this benchmark is watching — ▲ 877 · 💬 374 · 🔗 github.com
    livenerf is a 30-day, pre-registered benchmark that tracks whether Claude Opus 5.5 degrades after its 2026-09-22 launch. Using a frozen panel of 78 questions that the model only sometimes answers correctly, it runs daily on a Claude Max subscription via headless Claude Code, logging everything to detect statistical drift. Early validation shows that lower effort settings reduce output tokens far more than accuracy, and that a same-family model swap (Opus 5) is not distinguishable from Opus 5.5 in a single validation run.
  2. Gemini 4 Argon writes 1M tokens in one go, cracks a healthcare flaw other models missed — ▲ 866 · 💬 581 · 🔗 blog.google
    Google's Gemini 4 Argon is a frontier model for long-horizon reasoning, rolling out first to trusted cyber defenders. It sets state-of-the-art scores on DeepSWE v1.1 (77.9%), LVBench (91.7%), and CWE-bench v1 (68%), and can generate up to 1M output tokens. Internally, it optimized quantum subroutines by 40%, freed 300 TiB of memory, and migrated C/C++ code to Rust. Pricing starts at $2 per million input tokens.
  3. Pi.dev reverses course and adds MCP support to its core — ▲ 595 · 💬 332 · 🔗 earendil.com
    Pi.dev once proudly declared it would never support MCP, and its team dismissed the protocol on podcasts. Now MCP ships in Pi's core. The shift came after rethinking MCP's evolution and building Codemode, a JavaScript sandbox that orchestrates tool calls. Pi's makers argue MCP should resemble OpenAPI with intelligent tool discovery, and they want to help shape it for small harnesses rather than watch from the sidelines.
  4. One Closed Strait, Three Crises: How the World Became a Chain of Dependencies — ▲ 485 · 💬 362 · 🔗 tomwojcik.com
    A Polish writer traces how the closure of the Strait of Hormuz since March 2026 has rippled through fuel, food, and heating. Brent crude hit $108 a barrel, US diesel topped $6 a gallon, and EU gas storage sits at a record low. The deeper problem: for thirty years we swapped buffers for dependencies, and this year several failed at once. Even Poland, the EU's most food self-sufficient country, cannot escape the price shock.
  5. Vermont's hidden power plant: home batteries now the state's largest source of electricity — ▲ 374 · 💬 275 · 🔗 bbc.com
    Green Mountain Power has installed batteries in over 5,500 Vermont homes, linking them into a virtual power plant that now provides 110MW—the state's largest power source. The program has replaced aging peaker plants, saved customers $11 million last year, and kept the lights on during storms. GMP aims to eliminate all outages by 2030 and more than double the virtual plant by 2034.
  6. DeepSeek's Free Gift Shrinks AI Memory by 437x—and Western Labs Are Quietly Cashing In — ▲ 368 · 💬 405 · 🔗 insufferable.dev
    The AI race has taken an awkward turn: Western labs like Anthropic and OpenAI are quietly adopting DeepSeek's KV cache optimizations, which cut memory use by up to 437x. DeepSeek shared these breakthroughs freely, and the recent silent releases of Claude Opus 5.5 and GPT-6.1 Sol show sharp drops in cache-read pricing—60% and 80%, respectively. The author questions why Chinese labs would hand such a lifeline to their loss-making Western rivals.
  7. NASA secretly pulls a Blackbird from its display pole to fly again after 27 years — ▲ 310 · 💬 367 · 🔗 aviationweek.com
    NASA has quietly removed SR-71A Tail No. 844 from its display at Armstrong Flight Research Center and offered jobs to retired SR-71 engineers to help return the Blackbird to flight after nearly 27 years. The agency destroyed its $600 million spare parts supply in 2007, and unique fuel, ignition chemicals, and ground equipment are all gone. Former test engineer Mike Relja doubts it's feasible, but NASA Administrator Jared Isaacman has hinted at a new X-plane fleet.
  8. Backblaze's Hard Drive Failure Rate Hits 1.73%, Highest in Years — ▲ 296 · 💬 104 · 🔗 backblaze.com
    Backblaze's Q2 2026 Drive Stats report shows an annualized failure rate of 1.73% across 354,415 drives, the highest in recent quarters. Three drives exceeded the 6.95% outlier threshold, including a 10TB Seagate at 9.33%. Seagate models swept the zero-failure list, while two older HGST drives retired. The report also notes no new drive models for the second consecutive quarter, with deployments concentrated in 20TB+ drives.
  9. Solving Factorio Quality with Matrix Math — ▲ 247 · 💬 97 · 🔗 exyr.org
    Simon Sapin, a Factorio player, uses matrix math to model the game's quality mechanics. He explains the probabilities of quality upgrades, the role of recyclers in upcycling loops, and how to plan factories using linear algebra. He also shares a new online calculator tool to help players optimize their production chains.
  10. Why Is Sam Altman Still a Free Man? — ▲ 242 · 💬 218 · 🔗 prospect.org
    A legal filing by The New York Times and 11 other publishers reveals that OpenAI deliberately bypassed the Times' paywall to scrape millions of articles, with co-founder Greg Brockman responding 'ah nice.' The author argues that OpenAI's models hack websites because that lawbreaking is the company's business model, and asks why no one has been held accountable.
  11. A 16-Year-Old Found a Flaw That Could Have Unlocked 17 Trillion Microsoft Records — ▲ 240 · 💬 107 · 🔗 blog.faav.net
    Faav, a 16-year-old bug bounty hunter, discovered that Microsoft's internal Titan analytics service never verified JWT signatures. By crafting an unsigned token with the username 'admin', he executed SQL queries as an administrator. The exposed data included 17,990 employee emails, 25,000 account records, and an estimated 17.3 trillion rows across 17 analytics databases. Microsoft fixed the issue and acknowledged his coordinated disclosure.
  12. How the Bloomberg Terminal Changed Wall Street Forever — ▲ 209 · 💬 84 · 🔗 spectrum.ieee.org
    In 1981, after being fired from Salomon Brothers, Michael Bloomberg founded Innovative Market Systems to build a terminal that combined real-time data with instant analytics. The first Market Master terminals, delivered to Merrill Lynch in 1982, gave traders an unprecedented window on global markets. The terminal evolved from a closed network with a custom keyboard to a subscription-based service costing over $32,000 a year, becoming an icon of financial power.
  13. Data centers in Europe keep their water and electricity use secret — ▲ 202 · 💬 189 · 🔗 nltimes.nl
    A year-long investigation by Lighthouse Report and European media found that most data centers in Europe refuse to disclose their environmental impact, despite an EU directive requiring it. In the Netherlands, less than a quarter of large data centers publish electricity and water usage figures. Data centers already consume 4.6% of Dutch electricity, a share projected to reach 10-15% by 2030, intensifying grid congestion.
  14. Jev: The Text Classifier That Surprised Even a Skeptic — ▲ 201 · 💬 10 · 🔗 magazine.sebastianraschka.com
    Sebastian Raschka traces text classification from bag-of-words and logistic regression through RNNs and ULMFiT to the newly released Jev model. Initially skeptical that Jev was 'just a classifier,' he found it works better than expected, offering a faster, cheaper alternative to LLMs for classification tasks while being more general than task-specific models.
  15. The night sky is getting 10% brighter every year — and we're forgetting what darkness feels like — ▲ 199 · 💬 113 · 🔗 theguardian.com
    Light pollution is growing 10% annually, and 80% of people now live under washed-out skies. In this first-person essay, Tove Danovich describes lying under the Milky Way in Oregon's outback and argues that darkness is a right worth protecting — not just for stargazers, but for migrating birds, pollinators, and human health. Cities like Flagstaff, Arizona, show that sensible lighting rules can bring the stars back.
  16. My Great-Great-Grandfather Thought Technology Would Replace Him Too — ▲ 168 · 💬 398 · 🔗 manuel.darcemont.fr
    A farrier in rural France watched a car roll through the next town and soon became a mechanic, keeping his purpose of helping people get around. The author, surrounded by developers anxious about AI, finds comfort in this family story. The real driver, he suggests, was never the code but the joy of making things and solving problems. Be willing to let go of the how to hold on to the why.
  17. When Oil Prices Spike, the Money Flows to Oil Companies — ▲ 156 · 💬 156 · 🔗 theconversation.com
    Oil prices jumped from $66 to $101 a barrel after the U.S. and Israel attacked Iran. Energy economists explain that the windfall goes mainly to oil companies, but who benefits depends on ownership and regulation: Saudi Arabia's government, private shareholders in the Permian Basin and the U.K., Norway's sovereign wealth fund, or Russian oligarchs. Consumers ultimately pay the price.
  18. Singapore's Government Dating App Runs on a Nobel Prize-Winning Algorithm — ▲ 151 · 💬 69 · 🔗 twitter.com
    Singapore's FirstDate, a pilot dating app for public servants aged 21–35, uses the Gale-Shapley stable marriage algorithm—the 1962 game theory method that won the 2012 Economics Nobel. Users rank preferences, receive one match per cycle with a 72-hour decision window, and contact info is revealed only on mutual yes. Singpass verification ensures authenticity. The app's success metric is deleting its own users.
  19. Tesla borrows $30 billion as profits evaporate — ▲ 147 · 💬 200 · 🔗 electrek.co
    Tesla has opened $30 billion in credit lines from Citi and Wells Fargo, a regulatory filing shows, replacing a $5 billion facility. The move comes as profits have thinned to hundreds of millions, capital expenditures are set to hit $25 billion in 2026, and the company turned cash-flow negative last quarter for the first time since early 2024. Despite holding $43 billion in cash, Tesla is borrowing to fund operations and long-promised products.
  20. I Needed 1+1, So I Built a Functional Programming Language — ▲ 146 · 💬 68 · 🔗 hereticpleb.vercel.app
    A simple data structures assignment to evaluate 1+1+1 with a binary tree spiraled into building a full functional language in C. The author implemented closures, a garbage collector, a custom arena allocator, a REPL, and an FFI, then watched fib(40) consume 12 GB before crashing. The post details the journey from a tagged union to a chunk allocator and the realization that a garbage collector was inevitable.
  21. I Built a LinkedIn Slop Detector in 90 Minutes — ▲ 136 · 💬 145 · 🔗 hereticpleb.vercel.app
    The author rants about LinkedIn's performative culture, where every other post is a generic computer vision project, like hand tracking or pothole detection. To prove how trivial these are, he builds his own YOLO-based detector for such posts in 90 minutes, from data collection to training. The real issue, he argues, is that LinkedIn rewards looking interesting over actual skill, so everyone claps and moves on.
  22. EDG's C++ front-end goes open source after 30 years — ▲ 132 · 💬 52 · 🔗 edgcpp.org
    On September 30, 2026, EDG's C++ front-end source code becomes public, with The C++ Alliance as its nonprofit home. The same engine and standards continue under a new steward, funded by tax-deductible contributions. Three development tracks—community contributions, ongoing maintenance, and collectively funded features—ensure one codebase for everyone, with no early access. John Spicer chairs the Fiscal Sponsorship Committee overseeing the process.
  23. TLA+ can't check what it can't express — ▲ 131 · 💬 29 · 🔗 buttondown.com
    Boris Cherny's claim that Claude Opus used TLA+ to find race conditions has sparked excitement about formal verification. But TLA+ has limits: it can't express reachability properties (like proving a game is winnable), hyperproperties (like energy-saving mode always uses less power), or properties over multiple steps or real time. While hacks like auxiliary variables and self-composition exist, they're messy and don't compose well. TLA+ is great for invariants and liveness, but not a silver bullet for agentic coding.
  24. How our vibe-coded website looks like a designer made it — ▲ 130 · 💬 76 · 🔗 railcode.dev
    Yakko Majuri, founder of Railcode, explains how coding agents helped him realize a playful, Rollercoaster Tycoon-inspired website. He shares his iterative process: generating multiple design options, building interactive playgrounds for color tweaks, and forking prototypes. The result impressed visitors who assumed a designer was involved, proving that LLMs can be used for tasteful, personal design.
  25. Slug Renders GPU Text Straight from Outlines, No Atlas Needed — ▲ 126 · 💬 51 · 🔗 alphapixeldev.com
    GPU text rendering has long relied on texture atlases, SDF, or MSDF—all baked at fixed resolutions that blur or round corners when scaled. Slug, Eric Lengyel's 2017 algorithm, now in the public domain, instead computes per-pixel coverage analytically from glyph outlines in the fragment shader. This article compares the five main approaches and explains why Slug stays sharp at any size, transform, or viewing angle, making it ideal for 3D, AR/VR, and dynamic text.
  26. Last.fm killed its RSS feeds, so this developer built new ones — ▲ 122 · 💬 41 · 🔗 lfm.xiffy.nl
    Last.fm discontinued its RSS feeds, but lfm.xiffy.nl restores them for any username. You get four feeds: recent tracks, loved tracks, top tracks, and top artists, with optional periods from 7 days to overall. The service also adds album art as enclosures and now supports recommended tracks in both RSS and Soundiixx-compatible JSON.
  27. JBR-001 - Open-source 3D-printable desktop companion robot — ▲ 121 · 💬 29 · 🔗 projecthub.arduino.cc
    JBR-001 is an open-source, 3D-printable desktop robot powered by the Arduino UNO Q. It combines a camera, distance sensor, buzzer, animated display, and three servo motors to create an interactive companion. You can print, assemble, and program it yourself. JBR-001 greets you with arm and head movements, plays sounds, and uses computer vision for advanced interactions. It's a versatile platform for experimenting with robotics, edge AI, and physical computing. All STL files and code are available on GitHub, making it easy to build and customize.
  28. NRC grants first U.S. construction permit for a BWRX-300 small modular reactor — ▲ 120 · 💬 114 · 🔗 gevernova.com
    The U.S. Nuclear Regulatory Commission issued its first construction permit for a BWRX-300 small modular reactor to the Tennessee Valley Authority at the Clinch River site in Oak Ridge, Tennessee. The permit, submitted in April 2025, follows the Canadian Nuclear Safety Commission's 2025 license to Ontario Power Generation and adds to growing North American regulatory momentum for the 300-MW design, which GE Vernova Hitachi says can be deployed as a standardized fleet.
  29. Magnitude's self-optimizing engine runs open models 2x faster than llama.cpp — ▲ 119 · 💬 54 · 🔗 github.com
    Magnitude is an open-source inference engine that compiles and tunes its kernels on your device before a model runs, delivering up to 2x faster performance than llama.cpp—92% faster decode on Apple Silicon and 19% on NVIDIA CUDA. It connects to agents like Pi, OpenCode, and Codex with one click, uses 27% less memory per agent, and works on any Apple Silicon, NVIDIA, or AMD GPU, or just a CPU. Free, private, and Apache 2.0.
  30. Dental Scope: Open-Source 3D Explorer Takes You Inside a Single Tooth — ▲ 114 · 💬 62 · 🔗 dental-scope.com
    Dental Scope is an open-source, interactive 3D explorer of dental anatomy, from the whole mouth down to the pulp and root canals of a single tooth. Built on BodyParts3D and Z-Anatomy data, it lets you orbit, zoom, and dissect layers inside a tooth. It's strictly educational—not for diagnosis or treatment planning—and the code is MIT-licensed.
  31. Plato's Five Platonic Solids: The Mathematical Art of Origami — ▲ 112 · 💬 22 · 🔗 mathigon.org
    Mathigon's Mathematical Origami collection offers folding instructions for Platonic and Archimedean solids, stars, and compounds. Plato linked the five Platonic solids to the classical elements, and the site provides 3D models via Polypad for interactive exploration.
  32. Brain Waves Are Far More Complex Than Scientists Thought — ▲ 107 · 💬 38 · 🔗 quantamagazine.org
    New research reveals that traveling waves in the brain are not just simple planar oscillations but include spiral and concentric patterns. These complex waves may help reorganize the brain in real time to meet behavioral demands, potentially playing a central role in memory and cognition. The findings challenge traditional views and open new avenues for understanding how the brain processes information.
  33. America.gov launches AI chat to answer questions from 29,000 government websites — ▲ 106 · 💬 41 · 🔗 america.gov
    America.gov is a new AI-powered chat that pulls answers only from official federal, state, and local government sources. It’s free, ad-free, and designed to protect privacy—your conversation disappears when you leave. The site consolidates 29,000 government websites into one place, with future plans for completing forms and tracking progress by 2027.
  34. AI can write your code, but only you can write the commit message — ▲ 104 · 💬 60 · 🔗 yedhu.me
    Before AI, I wrote long commit descriptions to explain the what and why, a process that forced me to reflect on my code. Now agents write both code and commits, but they often fabricate the why when context is missing. I still write commit messages myself—it's how I verify I understand what I'm shipping and capture temporary decisions with exit criteria.
  35. Netlify swaps V8 isolates for Firecracker MicroVMs, making Edge Functions 5x faster — ▲ 102 · 💬 38 · 🔗 netlify.com
    Netlify rebuilt its Edge Functions infrastructure, replacing a hosted V8 isolate service with Firecracker MicroVMs running inside its own edge network. Median warm invocation latency dropped from 25–40ms to 5–6ms, p99 improved by 47.4%, and availability reached 99.998%. The migration, done with Unikraft, adds stronger isolation, scales to zero, and keeps the same pricing and developer experience.
  36. Three Deaths at Burning Man Is Statistically Normal, but Near-Zero Every Year Is Not — ▲ 99 · 💬 122 · 🔗 ihavenapkinthoughts.substack.com
    Three people died at Burning Man in 2026, the highest toll in its history. A crude comparison to the U.S. mortality rate predicts about 12 deaths, but after age-standardization the expected number drops to 4.9. Using a Poisson model, observing 3 deaths has a p-value of 0.27—not abnormal. The real anomaly is Burning Man's near-zero death rate over decades, likely due to self-selection of healthy attendees.
  37. Declassified: The Top-Secret Satellites That Took Electronic Spying to the Battlefield — ▲ 96 · 💬 21 · 🔗 thespacereview.com
    For over 40 years, the NRO launched suitcase-sized ELINT satellites with code names like URSALA, RAQUEL, and FARRAH. Hitchhiking on larger photo-reconnaissance satellites, they evolved from chasing Soviet radars to delivering tactical electronic intelligence directly to commanders in the field. Newly declassified documents reveal how these tiny spacecraft brought the wizard war of space-based spying to the warfighter.
  38. Enigma - Interactive 3D Model of the WWII Cipher Machine — ▲ 94 · 💬 37 · 🔗 enigma.design
    Enigma is an interactive 3D model that brings the legendary WWII encryption machine to life. Explore its inner workings through 16 detailed lessons, from setting rotor dials to understanding the double step mechanism. Learn how Polish, British, and French cryptanalysts broke its codes. The site combines accurate 3D visualization with historical context, including wartime photos and daily key sheets. Whether you're a history buff or a cryptography enthusiast, Enigma offers a hands-on way to experience the machine that shaped history.
  39. GPU shader data layout mismatch: a 12-byte gap that can freeze your computer — ▲ 89 · 💬 6 · 🔗 mattkeeter.com
    When sharing configuration data between Rust and WebGPU compute shaders, subtle layout differences can cause GPU threads to spin forever. Matt Keeter discovered that a mat3x3 matrix takes 48 bytes in WGSL but only 36 in Rust, leading to a 12-byte offset error. He built a generic unit test using the facet reflection library and naga to automatically compare struct layouts, catching mismatches without hardcoding fields.
  40. Rust-built PSSA beats a transformer by 0.45 nats and runs 12x faster on CPU — ▲ 85 · 💬 37 · 🔗 github.com
    PSSA is a non-transformer language model written from scratch in Rust, with no PyTorch or TensorFlow. It uses a recurrent state-space core, an episodic memory bank, and plastic weights. On 12.7M tokens of WikiText-103, a 1.5M-parameter PSSA reached 3.98 cross-entropy versus 4.43 for a matched transformer, and generated 200 tokens in 226 ms versus 2,735 ms on the same CPU.
  41. The Ballmer Peak: Programmers' Legendary Drunken Productivity Zone — ▲ 85 · 💬 29 · 🔗 en.wikipedia.org
    The Ballmer Peak is a joke from xkcd that programmers hit peak productivity at a blood alcohol content of 0.1337%, a nod to leet. Named after Steve Ballmer, it has inspired hackathons and satirical studies. A 2024 SIGBOVIK paper found no sudden peak but noted that speed increased with some alcohol before declining. The concept remains a beloved meme in programming culture.
  42. Purdue Professor Flags 267 Students for AI Cheating in CS 240 — ▲ 82 · 💬 59 · 🔗 turkeyland.net
    Dr. Jeffrey Turkstra, a teaching professor at Purdue, reflects on the Spring 2026 CS 240 course, where 267 of 584 students were flagged for AI/LLM cheating. He defends the clear syllabus policy, explains the static analysis tool Argus, and admits his response could have been better, resulting in minimal consequences for violators.
  43. GPT-6.1 Sol arrives just 7 days after GPT-6 Sol, nearly matching Astra at a quarter of the cost — ▲ 79 · 💬 96 · 🔗 artificialanalysis.ai
    GPT-6.1 Sol lands one point below GPT-6 Astra on the Intelligence Index while costing less than a quarter per task ($0.72 vs $3.26). It gains 4 points over GPT-6 Sol and 5 over GPT-5.6 Sol, with notable jumps in Terminal-Bench 4.0 (+12), Humanity's Last Exam (+5), and AA-Omniscience Accuracy (+8), while hallucination rate falls from 60% to 54%. Pricing stays at $2/$10 per million tokens, but the cache read discount rises to 95%.
  44. KCC - Maximize eInk manga quality with 2D DFT and dithering — ▲ 76 · 💬 20 · 🔗 github.com
    KCC (Kindle Comic Converter) optimizes comics and manga for E-ink ereaders like Kindle, Kobo, and ReMarkable. It delivers true fullscreen, margin-free pages and uses advanced techniques like 2D DFT to eliminate the rainbow effect on Kaleido 3 color eink, plus dithering for smaller files. It compresses a 600 MB volume to 100 MB, fixes black levels, and auto-detects 2-page spreads. Runs on Windows, macOS, and Linux, supporting CBZ, EPUB, PDF, and more. Just drop files, convert, and USB transfer to your device.
  45. SDF Public Access Unix System Has Been Running Since 1987 — ▲ 76 · 💬 15 · 🔗 sdf.org
    A Hacker News thread about SDF, a public access Unix system founded in 1987 that still offers shell accounts to members. Commenters praise its Mastodon and Pixelfed instances, recommend SSHing in for the com chat and bboard discussions, and note a long history of prior HN discussions going back to 2017. One user admits having an SDF login but not knowing what others use it for.
  46. Ubuntu 26.04.1 LTS Arrives with Wayland, Rust, and NVIDIA CUDA in Repos — ▲ 75 · 💬 98 · 🔗 ubuntu.com
    Ubuntu 26.04.1 LTS Resolute Raccoon is now available, consolidating fixes from the initial release. It ships with GNOME 50 and fully embraces Wayland, and is the first Ubuntu release to natively distribute NVIDIA CUDA in its repositories. Security upgrades include TPM-backed full-disk encryption, Rust-written kernel drivers, and post-quantum cryptography by default.
  47. Ledge - Runnable Markdown Notebook for Developers — ▲ 75 · 💬 40 · 🔗 ledge.sh
    Ledge is an open-source Markdown notebook that executes shell commands, code, SQL, and AI prompts directly in your notes. Each note gets its own persistent shell, so variables and directories carry over between runs. Run blocks locally, on a remote server over SSH, or on your phone. Features include profiles for secrets, note locking, git sync, and an MCP server for AI agents. Available for macOS, Linux, Windows, iPhone, and Android. Install with a single curl command.
  48. Gemini 4 Argon (High) Tops Artificial Analysis Intelligence Index — ▲ 74 · 💬 32 · 🔗 artificialanalysis.ai
    Artificial Analysis has published its intelligence, performance, and price analysis for Gemini 4 Argon (High). The model is evaluated on the Intelligence Index v4.3.2, which combines ten benchmarks including AA-Briefcase, Terminal-Bench 4.0, and Humanity's Last Exam. The analysis also covers cost per task, token usage, context window, and cache pricing, offering a detailed look at the model's efficiency and capabilities.
  49. Mathematicians Tell AI Labs: Stop Testing Advanced Math on Proprietary Models — ▲ 73 · 💬 89 · 🔗 agmai.org
    A set of recommendations for AI labs that produce significant mathematical results. Drawing on over 600 responses from the mathematical community, the document argues that labs must not treat math results as marketing, should release proofs with full attribution and formalization, and must fund human understanding of AI-generated mathematics. It also warns that proprietary models risk creating a two-tier system that alienates mathematicians from their own discipline.
  50. Arduino Library Reverse-Engineers the AMT630A OSD Chip in $35 Backup Monitors — ▲ 70 · 💬 24 · 🔗 github.com
    A new Arduino/ESP32 library exposes the AMT630A's five hardware OSD windows for text, custom glyphs, 4bpp bitmaps, palette animation, and scaling. It includes a Python image converter and optional automation for entering the hidden factory menu on a tested three-button monitor. The project is based on hardware reverse engineering, and a stock-firmware ownership quirk can freeze the display pipeline if external I2C control is sustained.
  51. I let an LLM control a robot arm overnight — and it stacked blocks by morning — ▲ 69 · 💬 18 · 🔗 thisismypersonalblog.com
    A software engineer with 25 years of experience bought a preassembled LeRobot SO-101 arm and a webcam, then handed control to LLMs. After calibration and spatial mapping failed, he removed himself from the loop and left Claude with a simple goal: pick up red, blue, and purple blocks and place them in a taped area. He woke to find all three blocks moved. A time-lapse from screenshots captured the process.
  52. Cloudflare Is Building a Certificate Authority for the Whole Internet — ▲ 67 · 💬 51 · 🔗 blog.cloudflare.com
    Cloudflare has applied to the Chrome, Apple, Microsoft, and Mozilla root programs and signed an agreement to acquire a trusted root from GlobalSign, aiming to become a public certificate authority. The move adds a free, ACME-first alternative to Let's Encrypt, which issues ten million certificates daily, and targets post-quantum Merkle Tree Certificates by early 2027. Cloudflare plans to require ACME Renewal Information and publish reproducible builds and a public issuance dashboard.
  53. Claude Says: Why Deferring to AI Is a Sign of Weakness — ▲ 66 · 💬 29 · 🔗 ohhfishal.net
    A frustrated author pushes back against people who use AI as a crutch in technical discussions. If you can't answer a question without asking Claude, you've shown you lack authority on the subject. Worse, AI hype leads to overengineered solutions—like forking every dependency instead of using a built-in artifact repository feature. Spewing information isn't intelligence, and regurgitating it atrophies your brain.
  54. YouTube's A/B Testing of Videos Breaks the Shared Experience, Warns Marques Brownlee — ▲ 65 · 💬 39 · 🔗 blog.jim-nielsen.com
    Marques Brownlee's video essay "Dear YouTube" criticizes the platform's upcoming feature that lets creators publish A/B variants of videos. He argues it shatters the communal experience of everyone watching the same video, and that YouTube's insecurity-driven imitation of competitors leads to such features. Jim Nielsen applies these lessons to software: focus on finishing and conviction, not optimizing variants.
  55. Gitea 28.0.0 drops the 1. prefix and adds audit logging, bot accounts, and more — ▲ 60 · 💬 27 · 🔗 blog.gitea.com
    Gitea v28.0.0 is out, and it drops the historical 1. prefix from version numbers. This release brings audit logging, bot accounts, HTTPS deploy tokens, admin impersonation, code-owner approval rules, diff file filters, and an Actions queue view. It also includes security fixes and breaking changes: Git network operations now use an internal proxy with new egress rules, self-registration is off by default, and completed Actions runs are deleted after 400 days. Upgrade carefully.
  56. 68 Dead Floppy Emus, 65 Rescued: Inside a Hardware Failure Analysis — ▲ 59 · 💬 9 · 🔗 bigmessowires.com
    After years of accumulating QA failures, I finally dedicated a month to failure analysis on 68 Floppy Emu boards. The leading culprit? Clock crystals—42 boards had no clock or a bad clock, traced to a substitute crystal damaged during assembly. Other causes included bad CPLD chips, unprogrammed microcontrollers, soldering issues, and even a missing LED. I managed to revive 65 of the 68 boards, and the lessons learned will shape future design and assembly changes.
  57. Halfspace: An Experimental IDE That Makes Distance Fields the Star — ▲ 58 · 💬 6 · 🔗 mattkeeter.com
    Halfspace is an experimental IDE for solid modeling with distance fields, built on the Fidget kernel. It lets you incrementally build models, visualize the underlying fields, and export images or triangle meshes. The author, Matt Keeter, has been developing it since April 2025 and emphasizes it is not vibe-coded. The tool is cross-platform, targeting both web and native, and has driven improvements in GPU-accelerated rendering.
  58. Let's Ditch Google as a Verb — ▲ 56 · 💬 70 · 🔗 adam.farkas.pro
    Even after three years without using google.com, the author still says "to Google" when asking someone to search the web—a habit that persists from Google's long search monopoly. Starting today, they'll consciously switch to generic terms like "look up" or "search" (in German, "nachschauen" or "suchen"). It's not about rejecting eponyms entirely, but about fairness to the search engine they actually use, Kagi.
  59. The Real Reason the Bronze Age Collapsed — ▲ 56 · 💬 36 · 🔗 worksinprogress.news
    The Bronze Age collapse around 1200 BC wiped out the Hittites, Mycenaeans, and other great powers, and no comparable empires returned for centuries. Patrick Fitzsimmons argues that the usual suspects—drought, Sea Peoples, famine—don't explain why recovery was so slow. The real culprit: iron. Once ordinary farmers could make weapons locally, the imperial monopoly on violence vanished, and the political equilibrium shifted forever.
  60. Fake Atari 2600 Box Art That Fooled the Internet — ▲ 54 · 💬 31 · 🔗 mightygodking.com
    In 2008, the blog Mightygodking posted a series of photoshopped Atari 2600 box covers with absurd titles like 'It's Fucking Checkers' and 'Obligatory Educational Game.' The images were so convincing that many readers believed they were real, sparking nostalgia for the console's heyday and its famously sturdy joysticks.
  61. This Wallpaper Pulls Electricity From Thin Air—and Dehumidifies Your Room — ▲ 54 · 💬 33 · 🔗 advanced.onlinelibrary.wiley.com
    Researchers at Binghamton University have created the first wallpaper-type moist-electric generator. A single paper sheet uses a peripheral-absorption, center-evaporation design: glycerol at the edges captures humidity, a polyvinylpyrrolidone layer drives water inward, and a wax-treated center lets vapor escape while blocking liquid. One unit produces about 0.34 V and 2.2 µW/cm² at 80% relative humidity. A 1,596-unit array powers a wireless keyboard while dropping indoor humidity from 38% to 32%.
  62. An American Did a Machine Learning PhD in Japan While Working at Google — ▲ 37 · 💬 9 · 🔗 tokyodev.com
    An American on the MEXT scholarship completed a computer science PhD at Tokyo Institute of Technology while working part-time at Google. He explains why Japan's flexible student visa, three-year graduation clock, and growing AI scene made it attractive, but also warns about name recognition abroad, limited graduate courses, and the isolation of doing a doctorate in a foreign country.
  63. Ken Griffin Gives Carnegie Mellon $3 Billion, the Largest Individual Gift in Higher Education History — ▲ 36 · 💬 40 · 🔗 cmu.edu
    Citadel founder Ken Griffin has given Carnegie Mellon University $3 billion, the largest individual gift ever to a higher education institution. One billion dollars will strengthen the Pittsburgh campus, including $500 million for the School of Computer Science, which will be renamed for him. The remaining $2 billion launches CMU Miami, a 35-acre campus in Wynwood organized around societal challenges instead of academic majors. It aims to enroll students in 2028 and serve over 3,500 students.
  64. Before Pixels: The Modular Industrial Dashboards That Ran the World — ▲ 36 · 💬 6 · 🔗 unsung.aresluna.org
    During museum visits in Germany and Poland, I kept encountering modular industrial dashboards—the physical control panels that once monitored air traffic, subway trains, and power plants. They light up, count, and even have a day/night switch. I don't know much about them, but they're fascinating design and interaction systems, now largely replaced by software. If you know more, please write.
  65. Miller's Monster: The Polyhedron with 1280 Faces to Fold — ▲ 32 · 💬 1 · 🔗 software3d.com
    The Great Dirhombicosidodecahedron, known as Miller's Monster, is a uniform star polyhedron with 124 faces, 240 edges, and 60 vertices. It uniquely has eight faces meeting at each vertex and is the only uniform polyhedron with snub faces that are squares. This page documents the construction of a physical model, requiring 1280 external facelets, 42 A4 pages of nets, and an edge length of 21.6 cm.
  66. Chompi sampler goes open source as it's discontinued — ▲ 30 · 💬 5 · 🔗 chompiclub.com
    Chompi Club has released the complete hardware and software source files for the CHOMPI Mk1 sampler as a discontinuation open-source release. The GitHub repository includes firmware for TAPE, TEMPO, and a new WAVE wavetable synth, plus PCB schematics and enclosure files. No further updates or support will be provided, but the community can continue on the Chase Bliss Discord.
  67. Parrot - Open-Source AI Meeting Recorder with Live Copilot for Mac — ▲ 26 · 💬 10 · 🔗 openparrot.app
    Parrot is a free, open-source meeting recorder and live copilot for Mac that runs entirely on your machine. It auto-detects calls, transcribes and identifies speakers locally, and suggests answers from your own documents while you're still talking. No bots join your meetings, and your audio never leaves your Mac. With support for Ollama, Claude, OpenAI, and more, Parrot adapts to your workflow. It generates post-call reports with action items and coaching, and integrates with Claude Desktop and Cursor. Available for macOS 14+ on Apple Silicon.
  68. Functional ultrasound imaging could be 1000x better than fMRI—and it's just getting started — ▲ 21 · 💬 3 · 🔗 neuroai.science
    Functional ultrasound imaging (fUSI) captures rapid movies of brain tissue, tracking blood flow changes with up to 20% signal change—far exceeding fMRI's few-percent BOLD. It offers ten times higher linear resolution, 1000x more voxels, and no magnet. I break down the physics and signal processing, from plane-wave transmission to Power Doppler and clutter filtering, and share Python tutorials to make fUSI accessible beyond Matlab. The modality is far from its ceiling.
  69. Lathoa - Math app where AI is wrong on purpose — ▲ 21 · 💬 8 · 🔗 lathoa.ai
    Lathoa is a critical-thinking math app for kids aged 10–14. In each challenge, an AI companion named Errol solves a problem step-by-step but makes a deliberate mistake. Kids must spot the bad step, explain why it's wrong in their own words, and earn detective ranks. The app builds the skill of questioning AI-generated answers, with three difficulty tiers (Obvious, Sneaky, Expert) and no timers or pressure. It's free for three cases daily, with subscription plans for families, and every problem is verified before kids see it. No open-ended chat, no profiling, COPPA-respecting.
  70. Two Raleigh Vektars Donated for Restoration, Fulfilling a 30-Year Dream — ▲ 20 · 💬 3 · 🔗 retromash.com
    I finally got my hands on two Raleigh Vektars, the 1985 computerized BMX with a digital speedometer, radio, and sound effects. Donated by Ally the Retrohunter, they're rusty and yellowed but salvageable. I plan to restore them, possibly repainting one frame to the black version I always wanted, and will document each stage—from cleaning to electronics repair—in this blog series.
  71. Coltrane's Tone Circle: Decoding the Geometry Behind 'Giant Steps' — ▲ 20 · 💬 4 · 🔗 jtomschroeder.com
    John Coltrane once sketched a diagram for Yusef Lateef: the chromatic scale arranged in circles, with lines tracing harmonic relationships. This post rebuilds that diagram as an interactive widget and decodes its secrets. It reveals how Coltrane changes, tritone substitution, and the 2-5-1 progression all emerge from the circle's symmetric shapes—squares, triangles, and hexagons—that divide the octave into equal parts.
  72. Corral - Kill every command your agent starts — ▲ 19 · 💬 4 · 🔗 github.com
    Corral is a Linux tool that runs a command with a time limit and guarantees no leftover processes when it returns. It solves a common problem for AI coding agents and CI jobs: processes that daemonize, ignore SIGTERM, or hold output pipes open. Corral controls the full process tree using cgroup v2 or subreaper fallback, ensuring a clean slate. Install via prebuilt binary or build from source. Use with options like --wall, --mem, and --json for audit records. Corral leaves no process alive, as benchmarked against timeout and naive runners.
  73. Strata - Expressive semantic layer that can say no to your LLM — ▲ 18 · 💬 8 · 🔗 strata.do
    Strata is a semantic layer that brings discipline to AI-powered analytics. It runs in Docker with a bundled database, letting you model data from Snowflake, Databricks, PostgreSQL, and more using YAML. Your coding agent reads AGENTS.md to propose fact and dimension tables, fields, and joins, then waits for your approval. Strata enforces naming rules and join constraints, so it can refuse invalid queries instead of hallucinating. Deploy to your own server and explore blended reports. Try it free on your machine.
  74. Quarry: Dumb Servers, Smart Clients for Secure OS Image Distribution — ▲ 17 · 💬 0 · 🔗 amutable.com
    Quarry is a new software delivery toolchain that uses The Update Framework (TUF) to securely distribute immutable OS images. It embraces a 'dumb server, smart client' model where static blobs are served from cheap hosting, and all intelligence lives in the client. Key features include per-machine repositories for granular access control, flexible key escrow, and autonomous updates. Quarry integrates with systemd-sysupdate and adds TUF extensions for repository listing and cross-repository links.
  75. Recursive Make and -j: The Jobserver Trap — ▲ 17 · 💬 5 · 🔗 quuxplusone.github.io
    Recursive Makefiles often break parallel builds because the -j jobserver isn't forwarded to sub-makes. Testing GNU Make 3.81 through 4.3 reveals inconsistent warnings and behavior: without + or $(MAKE), parallelism silently drops to 1; with them, it works but can still trigger 'resetting jobserver mode' warnings when -j flags conflict. No version yet offers the ideal: inherit the parent's jobserver unless explicitly overridden.
  76. I Turned a Free DIRECTV DVR Into a Standalone Jellyfin Box — ▲ 16 · 💬 0 · 🔗 voidnullvalue.github.io
    A friend gave me a DIRECTV HR54-700 Genie DVR. I exploited a boot-plugin bug to get persistent root, removed the satellite blocker, and reused the Broadcom decoder and WebKit UI. The final box boots on its own, joins Wi-Fi, and plays Jellyfin media via the DIRECTV remote. No dish, no helper PC.
  77. The Dual Polytope: A Simple Trick for Simplifying Convex Hulls — ▲ 15 · 💬 1 · 🔗 cairnc.github.io
    Simplifying a convex hull usually means merging or dropping face planes and rebuilding the hull. Cairn Overturf explains the dual polytope approach using the support function's level sets. The key insight: the polar dual is the level set of the support function, its faces correspond to the original vertices, and its vertices correspond to the original faces. This makes editing face normals equivalent to moving dual vertices, and taking their convex hull rebuilds the simplified hull.
  78. EDG C++ Compiler Goes Open Source — ▲ 15 · 💬 1 · 🔗 github.com
    The EDG C/C++ front end, long known for its parsing compatibility and bug emulation that ensures code compiling with Clang, GCC, and MSVC also compiles with EDG, is now open source on GitHub. The repository also includes a C-generating back end, a C++-generating back end, a prelinker for automatic template instantiation, a minimal runtime support library, utilities for intermediate language manipulation, a name demangler, and purpose-built development tools.
  79. The World's First Nuclear Reactor Was a Stack of Graphite in a Squash Court — ▲ 12 · 💬 1 · 🔗 newyorker.com
    On December 2, 1942, the coldest day in Chicago in nearly fifty years, a team at the University of Chicago's Stagg Field started up the Pile — forty thousand graphite blocks holding uranium, with hand-operated cadmium rods as the only radiation protection. Neither the university nor the city was told. It went critical, ran for half an hour, and then became the blueprint for the reactors that produced plutonium for Trinity and Nagasaki.
  80. Wavelets in 3D Graphics: A Memory-Saving Trick for Textures — ▲ 12 · 💬 0 · 🔗 comonad.com
    Wavelets are a hot topic in computer graphics for multiresolution analysis and compression. I focus on the Haar wavelet basis to compress low-contrast textures in a 3D polygonal game engine I'm developing. The problem: a typical 256x256 texture takes 256KB, and a scene with dozens of textures can exceed 200MB. I use nonstandard decomposition to reconstruct mipmaps incrementally, avoiding stutter and maintaining a steady 25 FPS.
  81. From 1% to 40%: How Solar Cells Learned to Harvest the Sun — ▲ 12 · 💬 4 · 🔗 eia.gov
    A timeline from the U.S. Energy Information Administration traces photovoltaic evolution: from early cells converting under 1% of light to today's 40% efficiency. Key milestones include Bell Labs' 6% silicon cell in the 1950s, the first building-integrated PV system in 1980, and Boeing Spectrolab's 40% HEMM cell in 2007. It highlights the roles of space programs, oil crises, and government policies in driving solar innovation.
  82. Groundrun Automates Testing of Every Wi-Fi Setup Method on ESP32 — ▲ 9 · 💬 0 · 🔗 groundrun.io
    Groundrun demonstrates automated hardware-in-the-loop testing for all five Wi-Fi commissioning methods on the ESP32-C6: Bluetooth, Soft AP, captive portal, SmartConfig, and WPS. Using Claude to generate a simple Android app and firmware, they built repeatable CI regression tests. Results show captive portal is trickiest due to its brief sign-in window, while Soft AP's mechanism takes nearly 13 times longer than WPS. The setup also required emulating a router for WPS.
  83. Mechanochemistry of Molecular Motors — ▲ 8 · 💬 0 · 🔗 youtube.com
    A video presentation on the mechanochemistry of molecular motors, exploring how these tiny machines convert chemical energy into mechanical motion at the molecular scale.
  84. SATIS Shield - BGP Blackholing for Single-Router Networks — ▲ 6 · 💬 8 · 🔗 setecastronomyinc.com
    SATIS Shield brings enterprise-grade DDoS protection to single-router networks. Using BGP blackholing (RTBH), it drops malicious traffic upstream before it reaches your network. No need for multihoming or an existing BGP feed—just a router that speaks BGP and a free private ASN. With a GRE or WireGuard tunnel to a SATIS PoP, you get real-time, shared threat intelligence from other peers. Setup takes minutes, no admin approval. $59/month after a 14-day free trial, no credit card required. Ideal for homelabs, small businesses, and enterprises.
  85. AI wrote most of my text editor, and it actually works — ▲ 4 · 💬 0 · 🔗 scattered-thoughts.net
    Jamie Brandon rewrote his text editor, Focus, from scratch with heavy AI assistance, mostly using Opus 5 in Claude Code on a $20 subscription. He now commits AI-generated code after minimal review, relying on end-to-end tests, a fuzzer, and API snapshots to catch regressions. The result is less janky and more featureful than the original, including VCS integration he never got around to before. He still works single-threaded, designing thoroughly before letting the model run.
添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论