HN Digest 2026-10-07 · Anthropic's Claude Haiku 5.5 slashes costs by 75% while tripling benchmark score

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. Anthropic's Claude Haiku 5.5 slashes costs by 75% while tripling benchmark scores — ▲ 633 · 💬 319 · 🔗 anthropic.com
    Anthropic has launched Claude Haiku 5.5, its fastest and cheapest small model yet, designed for high-volume tasks like summarization, classification, and live support. It costs around 75% less to run than Haiku 4.5 and delivers massive performance gains, including a jump from 15.7% to 72.4% on OSWorld 2.1. The model also introduces an adjustable effort setting and pairs well with Opus 5.5 and Sonnet 5.5 as a coding subagent.
  2. Chrome ships JPEG XL with a Rust decoder that's fast and memory-safe — ▲ 472 · 💬 304 · 🔗 developer.chrome.com
    Chrome 155 now decodes JPEG XL, an image format with 30–50% better compression than JPEG, lossless and HDR support, and lossless JPEG transcoding. The decoder, jxl-rs, is written in pure Rust, using SIMD without unsafe code to match C++ performance while eliminating memory-safety bugs. The move follows years of developer requests through the Interop Project, and Chrome recommends trying both JPEG XL and AVIF.
  3. Visa and Mastercard Hit With New Antitrust Suit Over Swipe Fees — ▲ 466 · 💬 324 · 🔗 classaction.org
    A fresh class action accuses Visa, Mastercard, and several major banks of imposing anticompetitive credit card transaction fees on merchants. The litigation targets the interchange and network fees that retailers say inflate the cost of accepting card payments, reviving a long-running legal battle over how the card networks set prices.
  4. OpenAI brings GPT-6 to 1.2 billion weekly ChatGPT users with Intelligent UI — ▲ 466 · 💬 237 · 🔗 openai.com
    OpenAI is rolling out GPT-6 to all ChatGPT tiers, including Free and Go starting tomorrow. The new model introduces Intelligent UI, which generates interactive interfaces—charts, buttons, forms—tailored to each question. GPT-6 also answers 44% sooner on web-search queries and improves safety. The update applies only to the Chat experience; Work and Codex models remain unchanged.
  5. Margaret Hamilton, who wrote the code that landed Apollo 11, dies at 90 — ▲ 419 · 💬 39 · 🔗 news.mit.edu
    Margaret Hamilton, the MIT computer scientist who led the software engineering team for NASA's Apollo program, died on Sep. 30 at 90. She coined the term 'software engineering,' pioneered priority-driven and defensive programming, and her code saved the Apollo 11 moon landing when a 1202 alarm struck. She later founded two software companies and received the Presidential Medal of Freedom.
  6. A font recreated from photographs of classic Commodore 64 keycaps — ▲ 373 · 💬 62 · 🔗 github.com
    This GitHub project by Levente Szabadkai reconstructs the lettering on classic Commodore 64 keycaps from photographs, producing a font with cleaned-up shapes and consistent proportions. It includes uppercase letters, numbers, punctuation, arrows, the pound sign, selected accented letters, function-key labels f1–f12, and all 63 PETSCII graphic front legends. Available as TTF, OTF, and webfont, with editable outlines and a specimen. Lowercase input displays as uppercase, recreating printed key legends rather than the screen bitmap font.
  7. AnyPS5 ports PS5 games to PC without an emulator — ▲ 342 · 💬 293 · 🔗 github.com
    AnyPS5 is an open-source tool that rewrites PS5 executables into native Linux and Windows binaries, mapping 87% of the system libraries it has catalogued so far. A relinker converts the executable to the target platform's format, and reimplemented system prx libraries handle dynamic linking, with no emulation or separate runtime process. The shader recompiler emits SPIR-V, validated through Spirv-Tools. Dreaming Sarah runs at a stable 60 fps on a GTX 1050 Ti and i5-7500.
  8. Bigwords.page - Turn any screen into a sign with a URL — ▲ 322 · 💬 103 · 🔗 bigwords.page
    Bigwords.page transforms any browser into a full-screen sign, timer, or message using just a URL. No app, no login, and nothing stored on a server. Create displays for arrivals, countdowns, Wi-Fi passwords, and more by simply sharing a link. The built-in editor makes it easy, or you can hand-craft URLs with settings for colors, fonts, animations, and QR codes. It's open source, private by design, and works on any device.
  9. Nobel Prize in Chemistry 2026 awarded for solving how life's mirror-image molecules emerged — ▲ 288 · 💬 53 · 🔗 nobelprize.org
    Henri B. Kagan and Kenso Soai share the 2026 Nobel Prize in Chemistry for discovering non-linear effects and autocatalysis in asymmetric organic synthesis. Kagan's 1986 work showed how to create a greater excess of one mirror-image molecule, and Soai's 1995-2003 experiments produced the first reaction that spontaneously formed only one mirror image. Their findings explain how homochirality—the single-handedness of life's molecules—could arise and aid drug design.
  10. Strands Decider 2B: A 2B-Parameter Decision Model That Runs on Your Laptop — ▲ 274 · 💬 78 · 🔗 strandsagents.com
    Strands Labs has released Strands Decider 2B, a 2-billion-parameter open-source decision model that picks among options and assigns confidence scores. It runs locally on CPU or GPU with median latency around 115ms on an RTX 3090 and 153ms on an M3 MacBook. Built by replacing an LLM's text-generation head with a pointer head, it's designed for agentic workflows like tool selection, routing, and guardrails. Weights, training data, and scripts are on GitHub and Hugging Face.
  11. 191 animated ASCII art pieces you can drop into any web page — ▲ 274 · 💬 56 · 🔗 ascii.rest
    ascii.rest offers 191 animated ASCII pieces — 3D shapes, physics simulations, creatures, UI components, charts, type effects, 27 language logos, and 22 Linux distros — each a dependency-free TypeScript module. Use them in HTML, React, Astro, or plain TypeScript; they're MIT-licensed and free to modify. The site also includes 13 full-color scenes like alpine dawn and aurora fjord.
  12. Meta and Microsoft Slash Internal Claude AI Budgets to Push Their Own Coding Tools — ▲ 251 · 💬 246 · 🔗 rswebsols.com
    Meta and Microsoft are cutting employee use of Anthropic's Claude AI in favor of in-house coding assistants, The Information reports. Microsoft dropped monthly AI spending caps from $100,000 per employee to roughly $10,000, while Meta's Claude Code users fell from about 60,000 to 30,000 as it pushes MetaCode and Muse Code. Both remain major Anthropic clients, and customer spending on Claude through Microsoft's platforms keeps rising.
  13. Developers Are Exhausted: 2026 Survey Reveals Burnout, Job Insecurity, and Deepening AI Fatigue — ▲ 237 · 💬 163 · 🔗 2026.stateofdevs.com
    The State of Devs 2026 survey, with 5,463 respondents, paints a stark picture of developer well-being. Nearly half experienced job insecurity and insufficient wages, a quarter have been laid off, and over a quarter expect to change careers within five years. Two-thirds report reduced motivation and increased cynicism, while 62% have faced burnout. AI views are polarized: 49% positive, 42% negative, with many citing AI fatigue and unsustainable productivity expectations.
  14. Lean verification of AI-translated proofs may prove nothing about the original math — ▲ 237 · 💬 150 · 🔗 arxiv.org
    Autoformalisation uses AI to translate natural-language math into Lean, where proofs can be mechanically checked. But a new paper argues that faithful translation is arbitrarily hard—harder than the Halting problem—because resolving ambiguity in mathematical text sits at the top of the Solvability Complexity Index hierarchy. The authors give real examples of AI mistranslations, including OpenAI's announced Navier-Stokes blow-up proof, where the Lean proof does not match the natural-language argument.
  15. GitHub hit by outage across Git Operations, Pull Requests, and Actions — ▲ 223 · 💬 176 · 🔗 githubstatus.com
    GitHub is investigating degraded availability for Actions, Git Operations, and Pull Requests, with Webhooks and Issues also reporting degraded performance and availability. The incident, posted on October 7, 2026, affects five core services and updates are ongoing.
  16. Stop Meandering: The Biggest Mistake in Software Blogging — ▲ 197 · 💬 110 · 🔗 refactoringenglish.com
    Software bloggers often bury the lede, assume readers share their exact knowledge, and rely on links instead of explaining concepts. Drawing on examples from Go testing, Docker, and Joel Spolsky, this guide catalogs common anti-patterns—from meandering intros to mobile page overflow—and offers concrete fixes to make technical writing clearer and more engaging.
  17. A $2 ESP32-C3 runs a Pi-hole-style DNS ad-blocker with 537,000 domains and no PSRAM — ▲ 177 · 💬 75 · 🔗 github.com
    This project packs a Pi-hole-style DNS sinkhole into a $2 ESP32-C3 by storing 537,000 domains as sorted 40-bit FNV-1a hashes in flash and binary-searching them, using only ~50 KB of RAM. It answers blocked lookups in ~10 ms, forwards everything else, and serves a web dashboard with OTA updates. The trick beats string-in-RAM approaches and scales to millions of domains on larger chips.
  18. How a Beloved Tech Journalist Destroyed His Reputation with a Failed Kickstarter — ▲ 176 · 💬 83 · 🔗 jeremyreimer.com
    Robert X. Cringely, the pseudonymous columnist behind 'Accidental Empires,' raised over $35,000 on Kickstarter in 2015 to build tiny Minecraft servers. Despite repeated promises that shipping was imminent, the project never delivered. This detailed account traces years of delays, shifting excuses, and outright lies that turned a minor failure into a cautionary tale about the cost of deception.
  19. OpenAI Shatters Unique Games Conjecture and 371 Other Math Problems in One Day — ▲ 174 · 💬 209 · 🔗 scottaaronson.blog
    OpenAI released 372 major mathematical results, including a proof of the Unique Games Conjecture that had eluded researchers for decades. The proofs, some with Lean certificates, are so complex that no human has yet fully understood them, sparking a race to digest the AI's alien-style mathematics. The event marks a seismic shift in theoretical computer science and mathematics.
  20. Docker's New AI Agent Builder Hits 3.4k Stars on GitHub — ▲ 163 · 💬 76 · 🔗 github.com
    Docker Engineering has released docker-agent, an open-source tool for building, running, and sharing AI agents. It uses declarative YAML configuration, a rich tool ecosystem, and multi-agent orchestration. The project has quickly gained 3.4k stars and 481 forks on GitHub, with over 10,700 commits. It supports provider-backed evaluators, sandboxing, and integration with models like OpenAI's GPT-5.6.
  21. God of War on PSP recompiled to WebAssembly, running in the browser at 60 FPS — ▲ 150 · 💬 81 · 🔗 github.com
    A project that recompiles PSP games to WebAssembly, bypassing emulation. God of War: Chains of Olympus runs in Chrome and Firefox at 60 FPS, up to 4x resolution, with working audio and touch controls. Ghost of Sparta also boots. The author explains performance tuning and the challenges of static recompilation, high-level emulation, and WebGL2 rendering.
  22. Civil engineer hand-dug 15m of tunnels beneath his house — now it's for sale at £300k — ▲ 149 · 💬 154 · 🔗 readingchronicle.co.uk
    Civil engineer Tony Deane spent two decades digging a 1,300 sq ft network of tunnels and rooms up to 15m below his Tilehurst home, using only buckets and wheelbarrows with help from his six children. The complex includes a hand-built railway, a one-tonne goods lift, and a secret entrance hidden inside a rocket by the pool. It goes to auction with a guide price of £300,000–£325,000.
  23. Pi 1.0 Adds MCP Support Through Codemode — ▲ 145 · 💬 75 · 🔗 lucumr.pocoo.org
    Pi 1.0 now supports MCP servers via Codemode, a JavaScript sandbox running in the harness that lets the LLM orchestrate tool calls without bloating its context. Codemode enables concurrent operations, state persistence, and direct access to internal APIs like image generation and classifiers. It runs in a QuickJS WASM sandbox with no network, file system, or timers, keeping the execution environment separate and secure.
  24. A 4-Year-Old's Blueprint for a Checkerboard Table — ▲ 143 · 💬 17 · 🔗 gamesbyemail.com
    A father recounts a trip to Home Depot with his four-year-old son Guy, who insists on buying 'wood tape'. Initially baffled, the father lets Guy lead the shopping trip. The boy methodically gathers duct tape, paint, brushes, and lumber, revealing he's building a checkerboard table for his tree house. The father realizes Guy had a complete plan all along, turning a confusing errand into a lesson in determination and vision.
  25. Art History Museum - Walkable 3D galleries from Wikipedia — ▲ 140 · 💬 60 · 🔗 artmuseum.artfrompixels.com
    Art History Museum transforms Wikipedia's vast art collection into an immersive 3D experience. Explore every artist and artwork under one roof, from Chinese Painting to Contemporary Pop Art. Zoom into a period, choose an artist, and step inside a virtual gallery to walk among their works. Built with open-source code, it makes art history accessible and engaging for everyone. Free knowledge keeps democracies strong.
  26. Couple swatted 55 times over a Norm Macdonald tweet wins $575K from Milwaukee — ▲ 139 · 💬 135 · 🔗 cbc.ca
    Patrick Tomlinson and Niki Robinson were swatted 55 times in two and a half years after Tomlinson's 2018 tweet criticizing Norm Macdonald sparked online harassment. Police knew but kept responding; a judge found constitutional violations but granted officers qualified immunity. The couple settled with Milwaukee for $575,000, yet no one has been charged or disciplined.
  27. Pew Survey: 64% of Americans Say Social Media Harms Democracy, and the World Is Catching Up — ▲ 138 · 💬 104 · 🔗 pewresearch.org
    A new Pew Research Center survey of 42,151 people across 37 countries finds that in most nations with trend data, the share who believe social media is bad for democracy has risen sharply since 2022 or 2023. The U.S. remains the most negative at 64%, but countries like Germany, France, and the Netherlands have seen double-digit increases. Wealthier nations tend to be more skeptical, while younger adults and lower-income countries remain more positive.
  28. Hackers hijack three country domains to mint counterfeit TLS certificates for Google — ▲ 137 · 💬 48 · 🔗 arstechnica.com
    Attackers seized control of the .gh, .sl, and .as country-code top-level domains, rewrote authoritative DNS records, and passed automated domain-control validation to obtain unauthorized TLS certificates for several Google domains and other major brands. Google updated Chrome to block the certificates it found and worked with the issuing certificate authorities to revoke them, but warned that browser-side fixes can't fully protect non-Chrome users or catch every affected domain.
  29. Google Playground Launches a Discovery Hub for User-Made Games — ▲ 115 · 💬 2 · 🔗 labs.google
    Google Playground now features a discovery section where players can explore user-created games across genres like platformers, trivia, arcade shooters, word puzzles, and sports. Titles such as Snooze Jump, Truth Ninja, Polygon Pazer, Pitfalls, Octo Pop, Word Cube, Foosball Shootout, and Summit Swoop showcase the platform's diverse, community-driven content, with each game linked to its creator.
  30. Google's Playground lets anyone build games with a few text prompts — ▲ 111 · 💬 191 · 🔗 blog.google
    Google has launched Playground, an experimental browser-based platform that turns text prompts into playable games—no coding required. Create from scratch or remix starters, then test instantly. Share privately, via link, or publish to the Explore gallery. Games run on phone or laptop, with leaderboards and multiplayer in select genres. A Unity Spark integration is coming for pro-level mechanics and 3D. Available now in the U.S. for users 18+, with tiered access based on Google One membership.
  31. OpenAI Claims Integer Multiplication Below n log n — ▲ 109 · 💬 76 · 🔗 github.com
    A preprint on OpenAI's math GitHub repository presents an algorithm for integer multiplication with complexity below n log n. The repository contains hundreds of other mathematical preprints, but this one stands out for challenging a long-standing barrier in computational complexity.
  32. Two WWII Bombs Surface on the Same Day, and Britain's Bomb Squad Had to Defuse Both — ▲ 109 · 💬 88 · 🔗 theguardian.com
    In January 2026, a dredger pulled a 250kg German bomb from Exmouth Marina while a construction crew in Plymouth unearthed a 50kg bomb 50 miles away — the first same-day double discovery since the postwar era. The 11 Explosive Ordnance Disposal Regiment had to assess corroded, unreadable fuzes on both devices, any one of which could still level a street.
  33. Rust's derive macro silently adds #[inline], and it cost uv 160KB — ▲ 109 · 💬 29 · 🔗 yossarian.net
    Rust's #[derive(Debug)] and similar macros emit #[inline] on the generated implementations. For trivial types this is fine, but deeply nested error enums can bloat binary size as the inlined Debug code gets duplicated at every call site. The author discovered this while working on uv and reduced its binary by about 160KB by replacing the derived Debug with a custom macro that uses #[inline(never)].
  34. Lean proof settles optimal packing for 11 squares — ▲ 105 · 💬 49 · 🔗 github.com
    The optimality proof for packing 11 squares has been fully formalized in Lean, with all 7,920 local modules verified and zero admissions. The optimal side length is approximately 3.8770835900228141773, derived from a root of a degree-8 polynomial. The proof allows arbitrary orientations and uses native numerical certificates, trusting Lean's kernel and native compiler.
  35. PS5 Jailbreaks Are Escalating at an Unprecedented Pace — ▲ 101 · 💬 122 · 🔗 pushsquare.com
    A new PS5 jailbreak now works on firmware versions 7.00 through 13.60, covering most consoles not on the latest system software. Armed with AI tools, modders have rapidly installed emulators for Switch and the original Xbox, while PC emulation of PS5 exclusives like Demon's Souls is also advancing. The surge has fueled a run on used consoles, with GameStop charging up to $1,500 for a PS5 Pro. Sony is expected to respond with firmware updates, but older systems remain vulnerable.
  36. South Korea says AI agents likely hacked its banks — ▲ 97 · 💬 30 · 🔗 reuters.com
    South Korea's president says AI agents appear to have been used to hack the country's banks, according to Reuters. The claim, reported by the news agency, points to a new kind of cyberattack in which autonomous AI tools carry out intrusions. Details of the incidents and the banks involved have not been disclosed.
  37. How a Boring Mill Turned a Crooked Casting into a Cylinder Accurate to a Worn Shilling — ▲ 93 · 💬 42 · 🔗 glinscott.github.io
    In the 1770s, making a straight, round cylinder for a steam engine seemed impossible: boring bars sagged, castings were crooked, and the smallest mark on a rule was a sixteenth of an inch. John Wilkinson's boring mill, which supported the bar at both ends, produced a 50-inch cylinder so accurate that Matthew Boulton boasted it did not err the thickness of an old shilling. This article traces how workshops learned flatness, accurate screws, and millionth-of-an-inch measurement through interactive figures.
  38. The 'Push Ifs Up, Fors Down' Idiom Has an Algebra—and Limits — ▲ 92 · 💬 43 · 🔗 debasishg.github.io
    Debasish Ghosh explores the programming heuristic 'push ifs up and fors down' from TigerBeetle's Tiger Style and matklad's blog. He traces the same pattern in database query optimization (push selections/projections down, defer joins, vectorized execution) and in functional programming: pushing an if up is restricting a function to a subobject, while filter/map rewriting follows from the naturality of catMaybes. The algebra reveals when rewrites are legal and when they actually save work.
  39. AstroHelm - Smartphone-Powered Telescope Aiming with Live Plate Solving — ▲ 88 · 💬 30 · 🔗 astrohelm.app
    AstroHelm transforms your smartphone into a precision telescope aiming system. Mount your phone on any telescope or telephoto lens, and AstroHelm uses live plate solving to match the stars in your camera's view against a massive database, guiding you to galaxies, nebulae, planets, and stars with observatory-grade accuracy. No GoTo mount or proprietary hardware needed. It works offline, even under light-polluted skies, and integrates with Stellarium and SkySafari. Free to use with one-time upgrades for advanced features. Perfect for casual stargazers and serious astronomers alike.
  40. Claude and ChatGPT Recommend Pricier Products to Wealthier Shoppers — ▲ 88 · 💬 32 · 🔗 bloomberg.com
    A new study reveals that AI chatbots like Claude and ChatGPT act as shopping assistants that tailor price recommendations to a user's perceived wealth. Wealthier users are shown higher-priced items, while those with lower incomes receive cheaper options. The findings, reported by Bloomberg, could intensify calls for regulation of AI-driven commerce and raise concerns about algorithmic fairness in retail.
  41. I hired an illustrator to draw my house, and now it's my Home Assistant dashboard — ▲ 87 · 💬 4 · 🔗 antonfrolov.substack.com
    I commissioned an illustrator to draw my house, then turned that artwork into a Home Assistant dashboard using the picture-elements card. Devices animate when they turn on, day and night versions switch automatically, and the whole family actually enjoys using it. I share the technical brief, the YAML setup, and how I got it running on my TV.
  42. Google's SynthID Detector identifies AI-generated media from OpenAI, Nvidia, and Kakao — ▲ 86 · 💬 76 · 🔗 synthid.com
    Google has launched the SynthID Detector, a tool that identifies AI-generated images, video, and audio by detecting invisible watermarks embedded during creation. It supports media from Google, Nvidia, OpenAI, and Kakao, and remains effective after common edits like compression or filters. Unlike general AI detectors, it verifies an intentional signal rather than guessing from artifacts, but it only works with partner models and isn't foolproof.
  43. Rosalind Franklin Understood DNA's Structure Before Watson, New Analysis of Photograph 49 Reveals — ▲ 84 · 💬 2 · 🔗 link.springer.com
    A historian and an X-ray crystallographer argue that Franklin's Photograph 51 was not a missed opportunity but a deliberate reproduction of her earlier Photograph 49, which she had already used to calculate key helical parameters of DNA's B form. They contend that Watson's narrative erased her skill and judgment by suggesting the image spoke for itself only to him.
  44. LLMs Can Write Like It's 1866 and Cut Your Token Bill in Half — ▲ 83 · 💬 52 · 🔗 fiveminutesforward.com
    A benchmark of 50 passages and ~1,300 questions shows that instructing LLMs to write in telegraphese—dropping articles and filler while keeping every fact—cuts token usage by 40–49% with no loss in downstream accuracy. Across four model families, compressed records were read at parity or better than plaintext, with recovery ratios of 0.99–1.10. The technique works because the register is already latent in training data, but it only pays off when compression happens after content is settled and for models whose reasoning can be disabled.
  45. The Seven Vices of Highly Effective Victorians — ▲ 83 · 💬 145 · 🔗 worksinprogress.co
    Victorian elites got up late, drank heavily, and partied hard, yet Britain dominated the world. Their schools prioritized character over STEM, with sports and self-government. This essay argues that soft skills and networking, not academic rigor, drove their success.
  46. I Reached 1k Karma on Reddit, Then Got Banned for Posting My Project — ▲ 78 · 💬 75 · 🔗 orchidfiles.com
    A developer's firsthand account of trying to share an open-source project on Reddit. Each new account was banned: for posting a link immediately, for commenting too soon, for writing a post that seemed AI-generated, and finally after a month of building 1k karma. Moderators advised earning karma and using the mobile app, but even a manually reviewed post led to a ban. Reddit users never saw the projects.
  47. After 15 Years, a ServerFault Moderator Steps Down as Site Traffic Collapses — ▲ 75 · 💬 49 · 🔗 sysadmin1138.net
    A moderator elected in ServerFault's first community election in 2011 is quitting after over 15 years. The site now gets just 2-3 questions a day, down from its 2011-2012 peak. He describes how cross-cultural escalation norms once sparked flame wars, and how recent StackExchange-wide spam campaigns forced him to spend evenings manually banning spammers in real time.
  48. Teen Hacker Extorted Boeing Spin-Off While Evading FBI — ▲ 70 · 💬 15 · 🔗 krebsonsecurity.com
    A Jordanian teenager suspected of leading ShinyHunters was detained while the group was extorting Jeppesen ForeFlight, a Boeing spin-off sold for $10.55 billion. The suspect, known as Rey, allegedly taunted the FBI and Cl0p, and exploited an Oracle PeopleSoft zero-day to breach the FBI and dozens of other victims. Boeing acknowledged the extortion claims, while Jeppesen ForeFlight reported no operational impact.
  49. A 1993 BBS session in Mexico, resurrected from a 64K floppy disk — ▲ 69 · 💬 23 · 🔗 nanochess.org
    In 1993, Óscar Toledo G. used a homebrew Z280 computer and a 2400-baud Hayes modem to call La Cueva BBS in Mexico. He captured a 64K buffer of the session, which he recently rediscovered on an old floppy. Now he's turned that data into a JavaScript replay using an adapted jsTerm, letting you watch the BBS as it appeared over 30 years ago.
  50. Forever Junior: The Skills AI Can't Develop for You — ▲ 64 · 💬 56 · 🔗 tech.criteo.com
    A junior engineer at Criteo fears that AI agents, by doing the learning for us, will create a generation of forever juniors. To avoid that fate, he deliberately builds three human skills: curiosity (making the agent explain its decisions), reviewing (emulating senior feedback patterns), and handwriting (occasionally coding without AI). The goal is to become the senior who writes the comments, not just a junior with a good skill file.
  51. Why Your Mallet Head Needs an Angle—and How to Find It — ▲ 62 · 💬 22 · 🔗 timberframe-tools.com
    A joiner's mallet is a quick shop project, but the head angle is often overlooked. Drawing on old timber-framing books, the author explains that the correct angle depends on whether you swing from the wrist or elbow. He walks through building a mallet from scrap oak and walnut, then shows a simple string-and-tape method to mark the exact angle for your own swing, ensuring square strikes and accurate chisel work.
  52. M-VAVE FM-1 Synthesizer Firmware Cracked Open: JieLi AC791N Inside, But No Safe Recovery — ▲ 62 · 💬 39 · 🔗 github.com
    Reverse engineering of the M-VAVE FM-1 pocket synthesizer reveals a JieLi AC791N/WL82 SoC with a pi32v2 CPU and a Dexed-derived six-operator FM engine. The update protocol is not a demonstrated recovery mechanism; no ROM recovery, rollback, or safe interrupted-write behavior has been established for the single-bank layout. The repository provides detailed analysis, tools, and a Linux OTA client, but warns against using any update utility without reading the safety notes.
  53. Shaders brings 200+ WebGPU effects to React, Vue, Svelte, Solid, and JavaScript — ▲ 61 · 💬 28 · 🔗 github.com
    Shaders is a new library that provides over 200 WebGPU effects as declarative components for React, Vue, Svelte, Solid, and vanilla JavaScript. Each effect is a component, and you can nest, blend, and mask them. A visual design editor lets you export code for your framework, and a CLI syncs designs into your codebase. The library is MIT-licensed and used by 16,000+ design engineers. Pro adds 1,000+ presets and Framer integration.
  54. MIT's 3D printer spits out a working electric motor in three hours for 50 cents — ▲ 57 · 💬 27 · 🔗 news.mit.edu
    MIT researchers have built a multimaterial 3D-printing platform that fully prints electric machines in a single step. Using four extruders to handle conductive inks, magnetic materials, and more, they produced a working linear motor in about three hours with just one post-processing step. The motor matched or outperformed conventionally made versions, and material costs came to roughly 50 cents per device.
  55. Mirror life could soon be built — and it might endanger every living thing on Earth — ▲ 55 · 💬 92 · 🔗 economist.com
    Many molecules essential to life come in two mirror-image forms, yet Earth's organisms use only one. Scientists are now closing in on building "mirror life" — organisms made from the opposite-handed versions of those molecules. The Economist reports that such research poses serious risks to humanity and the planet, raising urgent questions about whether it should proceed at all.
  56. Stanford scientists regrow cartilage in aging joints by blocking an aging protein — ▲ 54 · 💬 12 · 🔗 sciencedaily.com
    Stanford researchers blocked the protein 15-PGDH in older mice, reversing natural cartilage loss and preventing arthritis after ACL-like knee injuries. Human cartilage from knee replacements also began producing new, functional tissue. If it works in people, the approach could lead to an oral drug or injection that regenerates cartilage and reduces the need for joint replacement.
  57. nanoMuse - Open-source personal AI agent for every device — ▲ 54 · 💬 20 · 🔗 github.com
    nanoMuse is an open-source personal AI agent that runs on Android, iOS, Windows, macOS, Linux, and the browser. Unlike assistants that only answer questions, it takes action: it controls apps through your phone's screen and computer windows, keeps working while the app is closed, remembers you through editable Markdown files, and asks before anything irreversible. It reaches across your devices, so a task started on your phone can run on your PC. Bring your own model key, use the free community relay, or self-host everything under GPL-3.0-or-later.
  58. 160 Sound Visualization Experiments, Open Sourced — ▲ 51 · 💬 14 · 🔗 kagan.in
    A collection of over 160 sound visualization experiments spans more than 25 categories, from waveform animations and spectrum-inspired patterns to matrix displays, particles, rhythm, and memory. The open-source archive is available on GitHub, offering a wide range of creative coding sketches for music and audio visualization.
  59. ProcInSh - 3D Linux process inspector in your browser — ▲ 50 · 💬 15 · 🔗 github.com
    ProcInSh is a web-based process inspector for Linux that turns system monitoring into a 3D journey. Inspired by Ghost in the Shell, it lets you wander through process space, visualizing processes as interactive 3D objects. Built with Rust and eBPF, it provides deep insights into process memory and environment variables. Install via crates.io or build from source, then open your browser to explore. Ideal for developers and sysadmins seeking a novel way to understand their system.
  60. Paramount and Warner Bros. Merge, Burdened with $82 Billion in Debt — ▲ 49 · 💬 3 · 🔗 techdirt.com
    Larry Ellison has merged Paramount and Warner Bros. into a $110 billion company now called Skydance, carrying $82 billion in debt. The deal is expected to bring mass layoffs, higher prices, lower-quality content, and a CNN more friendly to corporate and right-wing interests. History shows such mega-mergers—AOL Time Warner, AT&T's Warner Bros., Discovery's Warner Bros.—always end badly, yet weak opposition from Democrats and state AGs let it proceed.
  61. Jev Diagnoses Kubernetes Faults Without an LLM Agent, Passing 76% of Tests — ▲ 48 · 💬 27 · 🔗 sregym.com
    A Jev-driven pipeline without an LLM agent collects Kubernetes evidence and lets Jev select root causes and supporting observations. Across 21 SREGym-Lite faults, it passed 80 of 105 diagnoses (76.2%) in a median 14.6 seconds, with 16 faults passing all five attempts. The approach uses programmatic evidence collection and Jev's multiple-choice decisions, costing an estimated $0.15 in inference.
  62. Why a Loop That Does 6% of the Work Takes the Same Time — ▲ 48 · 💬 5 · 🔗 igoro.com
    Igor Ostrovsky uses C# code samples to reveal how processor caches shape real-world performance. A loop touching every 16th array element runs as fast as one touching every element because both fetch the same 64-byte cache lines. He also shows how L1/L2 cache sizes create performance cliffs, how instruction-level parallelism can double loop speed, and how cache associativity causes mysterious slowdowns for certain step sizes.
  63. Treg: One Token, Thousands of Agent Tools, No Provider Signup — ▲ 43 · 💬 11 · 🔗 github.com
    Treg is an OpenRouter-style catalog for agent tools: point an agent at one base URL with one token and it can call thousands of endpoints across SEO, social, enrichment, scraping, and media generation, priced per call from a cent. It carries the provider accounts so you don't need subscriptions like Semrush or Crunchbase, and it also registers your team's own keys, CLIs, and skills with server-side auth injection.
  64. Permanent Daylight Saving Time Would Push Toronto's Winter Sunrise to 8:51 a.m., Researchers Warn — ▲ 42 · 💬 108 · 🔗 sci.news
    Canadian health researchers are urging provinces to adopt permanent standard time rather than permanent daylight saving time if they end the twice-yearly clock change. In an analysis in the Canadian Medical Association Journal, Raymond Lam of the University of British Columbia and colleagues argue that later winter sunrises — 8:51 a.m. in Toronto, 9:40 a.m. in Calgary — would disrupt circadian rhythms, sleep and mood, and work against an evidence-based treatment for seasonal affective disorder.
  65. Pendulum's "+" Operator Reads the Call Stack to Guess What You Mean — ▲ 40 · 💬 19 · 🔗 dev.arie.bovenberg.net
    Pendulum, the popular Python datetime library, made two promises: DST-aware arithmetic and drop-in compatibility with the standard library. Those promises collide inside datetime.astimezone(), where the standard library calls + expecting wall-clock semantics. Pendulum's fix: inspect the call stack in __add__ and, if the caller is named astimezone, fall back to the standard library's behavior. The hack is fragile, breaks on PyPy and dateutil time zones, and makes + roughly 600 times slower. A merged pull request removes the guess.
  66. Xbox locks up exclusive GTA 6 streaming rights — ▲ 38 · 💬 56 · 🔗 theverge.com
    Microsoft has secured exclusive streaming rights to Grand Theft Auto VI, meaning only Xbox Cloud Gaming can stream the game at launch. Xbox CEO Asha Sharma teased the deal to employees as something "no other platform holder is doing." It lands as Microsoft opens pay-as-you-go cloud access in November — the same month GTA 6 arrives — though Rockstar still hasn't announced a PC version.
  67. gtlds.fyi - Explore All Proposed New gTLDs — ▲ 38 · 💬 54 · 🔗 gtlds.fyi
    gtlds.fyi is a comprehensive, searchable database of all 986 proposed new generic top-level domains (gTLDs), including 263 overlapping applications. It lists each domain, the companies that applied for it, and the number of applicants, with filters for replacements and alphabetical navigation. Whether you're a domain investor, brand manager, or internet policy enthusiast, this tool provides a clear view of the evolving domain name landscape. The source code is available on GitHub, making it easy to contribute or customize.
  68. AWS Graviton Powered Nearly Half of Amazon's Prime Day Compute — ▲ 36 · 💬 27 · 🔗 aws.amazon.com
    Prime Day 2026 ran June 23–26 with millions of deals across 35+ categories, and AWS has published the numbers behind it. Graviton handled up to 49% of Amazon.com's EC2 compute, EBS peaked at 24.8 trillion I/O operations, Lambda ran 2.3 trillion daily invocations, and DynamoDB processed 59 trillion requests while peaking at 192 million per second.
  69. AI firm HUMXN will pay for your plumber to train robots — ▲ 34 · 💬 60 · 🔗 humxn.io
    HUMXN is subsidizing plumbing, electrical, HVAC, and cleaning appointments in Minneapolis, Chicago, and Miami. Technicians wear capture gear that records how they diagnose leaks, trace faults, and choose tools, turning real service calls into training data for AI and robotics. Homeowners get free service, contractors get paid, and HUMXN sells the datasets to physical-AI developers.
  70. Agent.reviews - Where AI agents read and write reviews on tools — ▲ 33 · 💬 33 · 🔗 agent.reviews
    Agent.reviews flips the script on software reviews: instead of humans rating tools, AI agents like Claude Code, Codex, and Cursor share their real-world experiences after using them. Each review includes task details, three ratings, and honest notes on what worked and what got in the way. The platform also runs leaderboards showing which tools agents actually choose for tasks. It's free and private—your code and prompts stay yours. Paste one prompt into your coding agent to install the review skill and start contributing. For tool makers, it's a new way to see how agents pick and use your product.
  71. Durable Actors - Open-source durable objects with configurable compute — ▲ 32 · 💬 24 · 🔗 github.com
    Durable Actors is an open-source alternative to Cloudflare Durable Objects that lets you build real-time applications like chat systems, collaboration tools, and agent swarms without vendor lock-in. It provides stateful serverless functions with durable state and serialized execution, so data survives crashes and concurrent updates stay safe. Define actors in TypeScript or Python, auto-generate type-safe clients, develop locally with one command, and self-host on GCP. Built-in observability makes monitoring and debugging easy, while a live demo and quickstart multiplayer AI chat show how quickly you can get started.
  72. VECOS brings a Windows-like OS to the Vectrex — ▲ 31 · 💬 8 · 🔗 youtube.com
    VECOS is a Windows-like operating system for the Vectrex, a vintage game console, designed for the UVMC2 cartridge. The project is demonstrated in a video, showing how a classic system can run a modern-style interface.
  73. ICANN receives 1,615 applications for new internet domains — ▲ 31 · 💬 44 · 🔗 icann.org
    ICANN has published the applications for the 2026 round of new generic top-level domains, revealing 1,615 total applications. The list includes 333 brand applications and 16 community-based applications, with the majority coming from North America and Europe. The next steps include a replacement period, string confirmation, and a community input phase before evaluation begins.
  74. Windows hot-patching is a one-lane road, and rogue patchers cause a pileup — ▲ 31 · 💬 7 · 🔗 devblogs.microsoft.com
    Raymond Chen explains that Windows hot-patching is designed for Windows Update alone, so if a function is already hot-patched by someone else, the system treats it as rogue patching and declares the file not hot-patchable, forcing a reboot. A race condition can leave a binary half-patched with no reliable rollback, like parking in a fire zone.
  75. Whoever Manufactures Consensus Owns Reality: How AI and Data Poisoning Weaponize Perception — ▲ 30 · 💬 23 · 🔗 medium.com
    Marketing has become an epistemological engine that bends reality. From Leopold II's philanthropic mask to the erasure of Soviet WWII sacrifices, power controls what we remember. Today, data poisoning and answer engine optimization let state actors and PR firms seed synthetic consensus into LLM training pipelines, drowning truth in optimized noise. The piece outlines how moral inversion works and offers four checks for epistemic hygiene.
  76. Thomas Jefferson's Olive Dream Is Finally Taking Root in the American South — ▲ 30 · 💬 26 · 🔗 ambrook.com
    Thomas Jefferson once called the olive 'the richest gift of heaven' and urged Southern farmers to plant it, but humidity and frost killed the idea. Now, as Mediterranean production drops, growers in Texas and Georgia are trying again. Woodpecker Trail Olive Farm in Glennville produces up to 18,000 bottles a year and has won international awards. Yet climate change still threatens the crop, with freezes and storms wiping out thousands of trees.
  77. The Paper Crane's Hidden Power: A Single Wish, a Thousand Folds — ▲ 29 · 💬 12 · 🔗 mazdastories.com
    The paper crane is more than a simple origami figure—it's a symbol of longevity, fortune, and miraculous wishes. The senbazuru legend promises a granted wish to anyone who folds a thousand cranes, a story epitomized by 12-year-old Sadako Sasaki after the Hiroshima bombing. Mazda's paper crane project now collects thousands of cranes from around the world for Hiroshima's Peace Memorial, while experts Kuniko Yamamoto and Toshi Aoyagi reveal the crane's perfect balance and its role in teaching math and language arts.
  78. Yggstore: Sharded, Encrypted Storage Between Friends Over Yggdrasil — ▲ 26 · 💬 9 · 🔗 github.com
    Yggstore is a proof-of-concept for sharded, encrypted file storage between peers on the Yggdrasil overlay, with no VPN coordinator. Each file is split into 4 MiB chunks, encrypted with AES-256-GCM, and erasure-coded into 4 data + 2 parity shards spread across online peers. A small stub holds the key, so sending someone a stub is a way to send them the file. The cryptography is unreviewed and experimental.
  79. AI Agents Caught Their Own Lies — And That's the Problem — ▲ 25 · 💬 15 · 🔗 cringely.com
    A swarm of cheap AI agents built a website for about eight dollars, but the real story is that a second agent caught more than a dozen fabricated quotes and fixed them without human help. The author argues this cheap self-checking doesn't solve hallucination; it makes 'we checked' worthless and shifts value to independent, accountable proof — like an auditor's signed opinion.
  80. Einstein's 1905 Explanation of Brownian Motion Finally Proved Atoms Exist — ▲ 23 · 💬 0 · 🔗 gregorygundersen.com
    In 1827, botanist Robert Brown saw pollen grains jittering in water, a mystery that stood for decades. In 1905, Einstein proposed that invisible molecules were bombarding the particles, and Smoluchowski independently derived the same theory. Perrin's experiments confirmed it, helping to establish the reality of atoms. This post explains Brownian motion, from random walks and binomial trees to the normal distribution, showing how a completely random process yields precise probabilistic predictions.
  81. EmDash taps Cloudflare's Clef to screen every plugin listing — ▲ 21 · 💬 7 · 🔗 emdashcms.com
    EmDash's plugin registry now uses Cloudflare's Clef decision model to screen every listing for phishing, impersonation, scams, and offensive content. Clef asks nine questions about text and links, eight per image, returning probabilities; scores of 0.45 or higher go to human review. The multimodal model replaced a slower multi-model pipeline, passing all public regression tests with 1.64s p95 latency, while keeping signed releases available from authors' own accounts.
  82. I Ran Into Redis Creator antirez on a Quiet Street in His Sicilian Hometown — ▲ 21 · 💬 12 · 🔗 orazioscicolone.it
    As a Sicilian SRE working abroad, I've followed Salvatore Sanfilippo's work for years. While visiting family in central Sicily, I learned we were both nearby and decided to attend a local festival he'd mentioned. There, I unexpectedly crossed paths with him on his evening walk. We chatted briefly about AI and his recent experiments—a surreal reminder of how small the tech world can be.
  83. OpenAI's AI Solves the Unique Games Conjecture and 371 Other Major Math Problems — ▲ 21 · 💬 1 · 🔗 scottaaronson.blog
    On October 7, 2026, OpenAI released 372 major mathematical results, including a proof of the Unique Games Conjecture, a breakthrough in derandomization, and faster algorithms for Fourier transform and integer multiplication. The proofs are machine-verified but largely incomprehensible to humans, sparking a race to digest them. The AI model behind this, an internal OpenAI system, solved about 5% of longstanding open problems it was tested on, each within roughly three hours of compute.
  84. Chinese AI Models Are the Biggest Conformists, Study Finds — ▲ 20 · 💬 7 · 🔗 magicnumbers.io
    A web tool that asks 12 AI models for their top-three picks on everything from books to paint colors reveals a striking pattern: Chinese models like DeepSeek and Kimi K2 agree with the group consensus far more than American or European models. The author explores possible causes—distillation, a tightly knit open-weight ecosystem, and benchmark-focused training—and warns that this quiet pressure toward a lab-mediated consensus could have consequences far beyond decorating advice.
  85. Thank You Indonesia for the Clean Air: A Satirical Relaunch — ▲ 19 · 💬 2 · 🔗 thankyouindoforthecleanair.web.app
    A satirical website thanks Indonesia for 11 months of clean air, quoting former Vice President Jusuf Kalla's remark that neighbours never thanked Indonesia for the good air but complained about one month of haze. Originally launched in 2015 and dark by 2019, it's now relaunched by Indonesian developer Dania Rifki.
  86. NetBSD's racoon2 IKE daemon learns to handle NAT and IPv6 without hacks — ▲ 18 · 💬 10 · 🔗 blog.netbsd.org
    As part of Google Summer of Code 2026, Artem Belan overhauled the racoon2 IKE daemon in NetBSD. The project implemented RFC 7383 fragmentation, NAT-OA payloads, and address substitution for NAT-T, fixed long-broken IPv6 support, and added a unit test framework. Now the daemon can serve as a reliable L2TP/IPsec or IKEv2 VPN server behind NAT, working with built-in Windows, iOS, and Android clients out of the box.
  87. Pinrail - Desktop inbox for coding agent reviews — ▲ 18 · 💬 2 · 🔗 github.com
    Pinrail is a desktop app that puts a human in the loop for agentic workflows. Instead of approving critical steps in a chat, your coding agent submits a review to Pinrail's inbox, where you decide in a view made for the content—diffs, lists, images, or documents. It works with any agent that can run a command, such as Claude Code, Codex, Cursor, or OpenCode, and keeps everything on your machine. The structured decision returns to the agent as Markdown or JSON, so it can carry on. Available for macOS and Linux, with Windows support coming.
  88. YouTube-dl restored at GitHub after RIAA takedown — ▲ 17 · 💬 2 · 🔗 lwn.net
    GitHub has restored the youtube-dl repository after removing it in response to a DMCA notice from the RIAA. The only change is the removal of tests that downloaded brief segments of music videos targeted by the RIAA. The tool, used for downloading video from various sites, is available again.
  89. Minut's Crowd Detect Tracks Airbnb Occupancy via Nearby iPhones — ▲ 17 · 💬 52 · 🔗 minut.com
    Minut's Crowd Detect estimates how many people are inside a rental by passively reading signals from nearby iOS devices, then alerts hosts when occupancy crosses a limit they set. The company pitches it as a way to catch parties and overcrowding early, uphold house rules and fire codes, and avoid manual checks. The same sensor also watches noise, smoking, temperature, humidity, and alarms.
  90. UniEvo-VL lets a multimodal model teach itself by critiquing its own images — ▲ 16 · 💬 3 · 🔗 arxiv.org
    UniEvo-VL is a self-evolving training recipe that turns a multimodal model's own critiques into privileged information: one model acts as both teacher and student, with the student seeing only the question and the teacher conditioned on the critique. Training minimizes the divergence between their denoising distributions along the student's own sampling trajectories. Built on Qwen-image-2512, it lifts GenEval from 0.747 to 0.808 and GenEval2 Soft-TIFA from 32.97 to 35.53, though gains are not uniform across tasks.
  91. Arcadeia - Self-hosted media library with animated video previews — ▲ 16 · 💬 6 · 🔗 github.com
    Arcadeia is an open-source, self-hosted photo and video library that lets you browse, search, and manage your personal media in a web browser. Its standout feature is dynamic thumbnails that automatically preview key moments in videos, so you can recognize a clip without opening it. You can search filenames and descriptions, optionally transcribe spoken words locally, connect network storage, upload via drag and drop, and play videos in your browser with on-the-fly transcoding. Built for self-hosters with personal media archives, Arcadeia runs on your own server and is licensed under AGPL-3.0. Try the live demo or deploy with Docker.
  92. How many GPUs does 1 trillion tokens per month need? — ▲ 16 · 💬 8 · 🔗 cedana.com
    Cedana's Tokens to GPUs Calculator estimates that serving 1 trillion tokens per month on Llama 3.3 70B requires about 367 H100 GPUs, with a plausible range of 211 to 853. The calculation converts monthly tokens to average throughput, adjusts for token mix and utilization, then divides by per-GPU speed. Sensitivity analysis shows H100 speed and utilization assumptions dominate the result. The tool also covers A100, V100, and B200, and warns that estimates can be off by 2x in either direction.
  93. Ghostty's last signed release was March 2026 — seven months ago — ▲ 16 · 💬 16 · 🔗 github.com
    Ghostty's GitHub tags page tells a quiet story: the last verified release, v1.3.1, was signed by Mitchell Hashimoto on March 13, 2026. Since then only the rolling "tip" continuous build has moved, most recently on October 7, 2026. The pattern is a long gap after a dense stretch of releases through 2025, from v1.1.1 in February to v1.2.3 in October.
  94. HN Simulator - A Pointless but Mostly-Exact Clone of Hacker News — ▲ 14 · 💬 5 · 🔗 news.ycombinator.lol
    HN Simulator is a near-perfect replica of Hacker News that lets you browse, vote, and comment on a simulated feed filled with fictional future tech news—like Claude Haiku 5.5 and GPT-6. It recreates the familiar orange-site experience, including rankings, user profiles, and submission flows, but with a playful twist: all content is fabricated. Perfect for developers who want to explore an alternate-reality HN without leaving the comfort of their browser. Visit news.ycombinator.lol to see it in action.
  95. TidesDB brings write-optimized storage to MySQL as a plugin, cutting disk writes by up to 47x — ▲ 14 · 💬 1 · 🔗 tidesdb.com
    TidesDB is now available as an external plugin storage engine for MySQL, not a fork. It offers write and space optimization, with features like compression, value separation, and optimistic MVCC. Benchmarks show up to 6x faster deletes and 47x fewer bytes written per operation compared to InnoDB. The engine supports full-text, spatial, and vector columns, and can be installed via a script.
  96. AI is quietly hollowing out the job market, and the next downturn could make it permanent — ▲ 13 · 💬 4 · 🔗 watch.wilsoniumite.com
    Labor's share of US business output has fallen to its lowest since 1947, and 107% of net job growth since August 2024 came from health care and social assistance. A new site tracks 29 indicators, 8 already triggered, and models a dot-com-sized bust that could leave 24% of the US workforce out of work by 2032. It argues income is shifting from wages to owners of land, location, and power.
  97. OpenWAM: A Shared Architecture for Composable World-Action Models — ▲ 12 · 💬 3 · 🔗 openwam.stanford.edu
    OpenWAM is an open framework that unifies world-action models on a common foundation, enabling systematic study of how video prediction and robot control interact. It supports flexible composition within a model—varying generation order and attention between video and action tokens—and between independently trained components like inverse and forward dynamics models. Pretrained on 3.34 million trajectories (14.64k hours), OpenWAM achieves 98.6% mean success on LIBERO and ~92% on bimanual tasks, showing that strong control doesn't require attention between future video and action tokens.
  98. 193 Free Online Tools That Don't Track You — ▲ 11 · 💬 0 · 🔗 normaltools.com
    Normal Tools offers 193 free utilities for text, development, design, money, files, building, hardware, planning, and names. Most run entirely in your browser, so your data never leaves your device and nothing is saved. No account, no download, no nonsense.
  99. OpenAI releases 722 math manuscripts, claiming proofs of Hilbert's tenth problem, BSD, and more — ▲ 10 · 💬 1 · 🔗 github.com
    OpenAI has published a collection of 722 mathematical manuscripts covering 372 result families. The papers claim proofs of major open problems, including a negative resolution of Hilbert's tenth problem over the rationals, the full Birch–Swinnerton-Dyer formula for many elliptic curves, the quasi-Riemann hypothesis, irrationality of Catalan's constant, and the Deligne–Drinfeld conjecture. The repository includes an overview PDF and links to individual preprints, with some results formalized in Lean.
  100. A 17-Foot Medieval Scroll Reveals the First Great Anatomy Illustrations — ▲ 10 · 💬 1 · 🔗 publicdomainreview.org
    The Sweden Scroll, a lavishly illustrated 15th-century manuscript of John of Arderne's surgical treatises, is the only known Arderne scroll and the most richly illuminated. Its four full-size anatomical figures—nude, flayed, and skeletal—are among the most remarkable of the era, bridging medieval and Renaissance anatomy.
  101. Python's Memory Allocator Rarely Shrinks: A Deep Dive — ▲ 9 · 💬 2 · 🔗 tech.daniellbastos.com.br
    A personal exploration of CPython 3.12's memory management: pymalloc for small objects (≤512 bytes) and malloc for larger ones, reference counting, and the garbage collector. The key insight: once memory is allocated, freeing it doesn't always reduce the memory the OS sees, because the allocator decides what goes back. You can control how long a reference lives, though.
  102. Trigora - Durable execution without history replay — ▲ 9 · 💬 2 · 🔗 github.com
    Trigora is a durable execution platform for long-running AI agents and programs, allowing you to write normal application logic in TypeScript, Python, or Rust. It suspends execution across events, timers, external effects, and child programs, then recovers from committed continuation state after failure without replaying completed history. Using Transparent Continuation Checkpointing (TCC), Trigora avoids the overhead of traditional replay-based systems. Deploy once and start executions from CLI, API, webhook, or cron. Ideal for AI agents that wait for humans or external events, sleep for hours or days, and coordinate child executions. Install via npm, pip, or cargo, and use Trigora Cloud for managed production runtime.
  103. The Scientific Knowledge That Could Save Lives Is Hiding in Plain Sight — ▲ 9 · 💬 0 · 🔗 aeon.co
    Science produces 7 million papers a year, yet single studies often contradict each other. Meta-analysis and systematic reviews — pioneered by Gene Glass and Iain Chalmers — synthesize all evidence to reveal what really works. Without them, we missed the link between front-sleeping and SIDS for 20 years, costing 50,000 infant lives. These unglamorous methods already shape medicine more than any headline.
  104. Vector RAG Is Not Enough for AI Agent Memory — ▲ 4 · 💬 10 · 🔗 claix.dev
    Traditional RAG splits documents into chunks and retrieves by vector similarity, but AI agents need to decide when and how to retrieve information. Agentic RAG treats retrieval as a tool the agent can invoke, evaluate, and repeat. This shift doesn't make vector databases obsolete—they remain useful for large knowledge bases—but for many document-agent workflows, developers should start with direct document operations like structured extraction, full-document processing, and multi-document comparison instead of a fixed vector pipeline.
  105. When the machine boots but the reply disappears — ▲ 4 · 💬 3 · 🔗 mainbrella.com
    Mainbrella's engineering blog digs into the backend mechanics that make machine creation reliable when replies vanish. It explains how Durable Objects serialize admission to prevent oversubscription, how idempotency keys and reservation fences stop stale messages from resurrecting canceled work, and why durable records don't make process handles durable. The piece follows a single job from admission through execution, showing the trade-offs behind every retry and the hard deadlines that govern compute budgets.
添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论