HN Digest 2026-09-22 · Claude Opus 5.5 cuts costs 40% while setting a new safety bar

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. Claude Opus 5.5 cuts costs 40% while setting a new safety bar — ▲ 1767 · 💬 1075 · 🔗 anthropic.com
    Anthropic's Claude Opus 5.5 matches the performance of its top model, Claude Fable 5.1, but costs 40% less to run than Opus 5. It completed a 680,000-line code migration in under a day and scored highest on the company's alignment audits, with stronger resistance to prompt injection. Pricing drops to $4 per million input tokens and $20 per million output tokens, and output is over 30% faster.
  2. OpenAI slashes GPT-6 API prices by 50% with Sol and Luna — ▲ 1729 · 💬 821 · 🔗 openai.com
    OpenAI expands its GPT-6 family with Sol and Luna, two models that bring Astra-level intelligence to faster, cheaper tiers. API prices drop 50% versus GPT-5.6 promotional pricing, with Sol at $2/$10 per million tokens and Luna at $0.10/$0.50. On benchmarks like AutomationBench and DeepSWE, Sol outperforms Claude Opus 5 and Fable 5 at a fraction of the cost. Improved prompt caching and alignment round out the release.
  3. Pentagon: Palantir AI Overreliance Led to Strike Killing 123 Iranian Children — ▲ 888 · 💬 497 · 🔗 bloomberg.com
    A Pentagon review found that overreliance on Palantir's Maven Smart System contributed to a U.S. strike on an Iranian elementary school that killed at least 123 children. Outdated data, stale satellite imagery, and a 90% cut to civilian-harm mitigation teams compounded the error. Palantir denies responsibility, and the U.S. has not accepted blame.
  4. I said no, and Apple said yes: macOS 27 removes the off switch for Apple Intelligence — ▲ 855 · 💬 690 · 🔗 dbushell.com
    After upgrading from macOS 15 to 27, the author finds Apple has removed the setting that once let users disable Apple Intelligence. The feature had been phoning home every 15 minutes with personal data. Now it is re-enabled by default, with unkillable Siri processes and a 22.28 GB disk footprint. The author argues this is a consent problem, not just a settings issue.
  5. ShinyHunters Claims It Hacked the FBI, Stole Data on Every Employee — ▲ 786 · 💬 589 · 🔗 404media.co
    ShinyHunters says it breached FBI-related services and stole names, home addresses, phone numbers, and spouse details for all FBI employees and applicants. It provided a 5,000-person sample, and 404 Media verified some phone numbers matched DOJ personnel. The group defaced the FBI jobs site, claims it used an Oracle PeopleSoft zero-day to reach AWS GovCloud, and says it exfiltrated 2–3 terabytes. It denies financial motive, calling its plan 'coercion.'
  6. Apple sneaks persistent ads into iOS, and users are furious — ▲ 783 · 💬 574 · 🔗 techradar.com
    Apple has quietly added persistent banners promoting its own products inside iOS, and iPhone owners are calling it out as a cheapening of the experience. The ads appear in system interfaces, sparking complaints that Apple is turning its premium platform into another billboard.
  7. OpenAI's GPT-6 Astra cracks an 85-year-old Enigma message that stumped humans for two decades — ▲ 720 · 💬 437 · 🔗 cryptocellar.org
    In September 2026, GPT-6 Astra autonomously broke the German Army Enigma message MVUEH from 10 July 1941, which had resisted all human attempts since 2005. The AI chose the message, suspected a link to a previously broken message, and built its own Enigma simulator and Bombe in Python and C++ to test a crib. The recovered key even used a different wheel order, and the plaintext nearly matched another message from the same day.
  8. Google's SynthID hides secret tracking IDs in your images and audio — ▲ 686 · 💬 166 · 🔗 brand.io
    Google's SynthID embeds hidden signals, called spymarks, into images, audio, text, and video. A 512x512 image can carry a 136-bit payload, enough for a 64-bit database ID linked to your personal records. Unlike visible watermarks, these signals are imperceptible, survive edits, and track you without consent. OpenAI and others are developing similar systems, raising serious privacy concerns.
  9. AMD Ryzen got 50% faster in two years—and it wasn't just clock speed — ▲ 474 · 💬 194 · 🔗 lemire.me
    AMD's Ryzen 7 9800X3D is 47% faster single-core and 58% faster multi-core than the 5800X3D from two years earlier, per Geekbench 6. Clock speed rose only 15%, so the gains came from a wider core: dispatch width grew from 6 to 8 instructions per cycle, the reorder buffer expanded from 256 to 448 entries, L2 cache doubled to 1 MB, and Zen 5's SIMD units widened to 512 bits.
  10. FoxPro killed in 2007, revived as 64-bit FoxDev Studio — ▲ 467 · 💬 250 · 🔗 foxscript.org
    Microsoft discontinued FoxPro in 2007, but FoxDev Studio brings it back. This 64-bit IDE runs Visual FoxPro 9 projects natively, with no conversion. It includes a Rust-based VM compiled to WebAssembly, a React-driven form designer, and a 32-bit bridge for legacy .fll libraries. FoxScript adds lambdas and HTTP routing to the same language. Tables can now grow beyond 2 GB, though they won't open in original FoxPro again.
  11. Gzip can write Shakespeare, sort of — ▲ 397 · 💬 160 · 🔗 nathan.rs
    The compression–prediction equivalence says every compressor is a predictor. So I built gzipt, a pure-Python language model that uses zlib's DEFLATE to score continuations by compressed length, then generates with beam search. Primed on Tiny Shakespeare, it produces semi-coherent text—no neural network, no training, just the compressor in your OS.
  12. AI Has No Wisdom and Neither Will You — ▲ 383 · 💬 542 · 🔗 alexn.org
    The software industry is being swept up by the idea that AI can replace human coding and code review, but the author argues this is a dangerous illusion. Maintainability and good architecture take months or years to reveal their value, so AI—trained on immediate rewards and mostly bad public code—cannot learn them. Relying on AI to write and read code prevents developers from ever reaching mastery. The author predicts a future where companies proudly advertise a "NO-AI" policy as a competitive advantage.
  13. Solar panels over California's canals: a win-win that's harder than it looks — ▲ 357 · 💬 692 · 🔗 kqed.org
    Jordan Harris and Robin Raj founded Solar Aquagrid and launched Project Nexus, a pilot with the Turlock Irrigation District, UC Merced, and the California Department of Water Resources to shade irrigation canals with solar panels. The panels cut evaporation, curb algal growth, and avoid the 621-day average permitting slog for open land. But urban canals in Contra Costa County and the state's varied infrastructure make scaling a challenge.
  14. SAML Is a Fractal of Bad Design—It's Time to Deprecate It — ▲ 340 · 💬 177 · 🔗 blog.trailofbits.com
    SAML, born from a committee that merged four XML security protocols, is crushed by its own complexity. XML signature wrapping attacks from 2012 still work today. Canonicalization, enveloped signatures, kitchen-sink design, and ossification make it a fractal of bad design. I worked on Duo's first SAML gateway and saw the flaws firsthand. Modern alternatives like OpenID Connect are the way forward.
  15. Meta's Muse leaks its entire runtime through a simple file export — ▲ 332 · 💬 163 · 🔗 mouse.dev
    A researcher asked Meta's Muse assistant to archive its filesystem and send it to Google Drive. The 6.8 GB export revealed internal documentation, 68 skill directories, unreleased connectors like Slack and Dropbox, an experimental Home Link integration for smart home devices, and SSH keys. Meta marked the bug bounty report Not Applicable, but the incident shows how internal runtime files can escape through ordinary conversations.
  16. Claude Opus 5.5: The Real Cost of Intelligence — ▲ 329 · 💬 103 · 🔗 artificialanalysis.ai
    Artificial Analysis breaks down Claude Opus 5.5 across its Intelligence Index v4.3.2, which bundles ten evaluations including AA-Briefcase, GDPval-AA, Terminal-Bench 4.0, and Humanity's Last Exam. The analysis maps intelligence against cost per task, output token use, cache-hit pricing, and context window, revealing where the model sits on the Pareto frontier of performance versus price.
  17. OpenAI could eat Jev's lunch by folding general classification into its models — ▲ 316 · 💬 221 · 🔗 arcturus-labs.com
    TypeSafe's Jev, a general-purpose classifier built on a large language model, has been adopted faster than any model in AI Gateway history. But OpenAI has used LLMs as implicit classifiers for years—tool calling, token-level decisions—and could replicate Jev's trick, then embed it inside its own models for smarter, faster, cheaper agents. The real moat may be TypeSafe's training data and calibration process, not the architecture.
  18. GrapheneOS phones will be sold preinstalled in 2027, project says — ▲ 307 · 💬 138 · 🔗 grapheneos.social
    GrapheneOS has officially partnered with Motorola Mobility (Lenovo) to bring its privacy-focused Android-based OS to a subset of future devices. The first will be a high-end flagship launching in 2027, with budget devices to follow once they meet the project's strict security and update requirements. Motorola is also helping with the port.
  19. AMD's random number generator can't generate a 0 — ▲ 280 · 💬 213 · 🔗 board.flatassembler.net
    A programmer discovered that AMD processors' RDRAND and RDSEED instructions never produce a zero at the requested bit width, while Intel chips do. He built a test app to visualize the anomaly, reported it to AMD, and got a reply that seemed AI-generated. The thread also debates hardware RNG trust and benchmarks performance across CPUs.
  20. Anthropic quietly adds Claude Opus 5.5 to its Foundation Models SDK — ▲ 273 · 💬 2 · 🔗 anthropic.com
    A single commit to Anthropic's ClaudeForFoundationModels repository introduces a new model constant for Claude Opus 5.5. The change updates the README and Swift source files, adding the opus5_5 identifier with capabilities including low, medium, high, xhigh, and max effort levels, adaptive thinking, structured output, and image input. Notably, Opus 5.5 inherits the Opus 4.8 request surface but rejects forced tool use.
  21. Vapers Are Quitting by Taking Up Cigarettes — ▲ 240 · 💬 325 · 🔗 bloomberg.com
    Vaping was supposed to help people ease nicotine addiction, but it hasn't worked out. On TikTok, Instagram, Reddit, and Facebook, e-cigarette users trade quitting tips and count days since their last hit. Now a counterintuitive idea is gaining steam: start smoking. Shaun Moore, a road-paving superintendent in Utah, turned to cigarettes last year to quit vaping. Smoking, these people say, is even more disgusting than vaping, making them less likely to light up.
  22. Unreal Agent's async harness cuts coding agent costs by 40% — ▲ 234 · 💬 121 · 🔗 unreallabs.ai
    Unreal Labs built Unreal Agent with a fully asynchronous tool-calling harness that removes waits, polls, and heartbeats from the model's workload. The result: up to 40% lower cost than Codex and 20% lower than Pi on real workloads and agentic benchmarks like Terminal-Bench 4.0, SWE-Atlas, and DeepSWE 1.1, with comparable pass rates. The SDK ships as a Go library, a runner executable, and a Harbor-compatible benchmark runner.
  23. A 9.2-rated WordPress flaw has lurked in every version since 2016 — ▲ 232 · 💬 128 · 🔗 github.com
    An unauthenticated path traversal in WordPress's page-template resolution lets an attacker include a local .php file outside the active theme, potentially achieving remote code execution. It affects versions back to 4.7, including popular themes like Twenty Twelve, Twenty Fourteen, Neve, Hestia, and Sydney. WordPress 7.1.2 patches it, with backports to all branches. Discovered by Robert Ressl.
  24. ReBarUEFI unlocks Resizable BAR on almost any UEFI system — ▲ 230 · 💬 72 · 🔗 github.com
    ReBarUEFI is a UEFI DXE driver that enables Resizable BAR on motherboards that lack official support. It patches the PCI host bridge to set BAR sizes from an NVRAM variable, boosting performance and meeting Intel Arc GPU requirements. The project includes UEFIPatch fixes for 64-bit BAR handling and a ReBarState tool. Users can achieve up to 12% FPS gains on older systems like an i5 3470 with RX 580.
  25. Three Hours, One Camera, and the Oldest Streetcar in America — ▲ 201 · 💬 54 · 🔗 daniel.lawrence.lu
    I spent three hours by the tracks at Muni Heritage Weekend 2026 with my custom line scan camera, capturing San Francisco's historic trams. Highlights include Muni No. 1, the first publicly owned big-city streetcar in the US, and No. 578 'Dinky' from 1896, which looks like a cable car but runs on electricity. The photos are freely licensed under CC-BY-SA 4.0 and uploaded to Wikimedia Commons.
  26. Git 3.0 looms as GitHub finally gets ready for SHA-256 — ▲ 198 · 💬 114 · 🔗 lwn.net
    Git 2.56, due around the end of September, is a modest release with over 700 non-merge commits and a new git history drop subcommand. But the real news is the next version, likely Git 3.0, which will switch to SHA-256 by default and adopt reftables. GitHub's SHA-256 support has been the main blocker; a GitHub employee now says news is coming, making 3.0 the best choice.
  27. Drop - Rootless Linux sandbox with gVisor support — ▲ 185 · 💬 61 · 🔗 droprun.sh
    Drop is a rootless Linux sandbox that isolates programs and coding agents without disrupting your workflow. Inspired by Python's virtualenv, it creates disposable environments with their own home directories, using your existing distribution so all installed programs remain available. Configure file, directory, and network access via TOML, and optionally run programs on the gVisor user-space kernel for an extra isolation layer. Perfect for safely running AI agents with --dangerously-skip-permissions or testing third-party packages from PyPI or npm, Drop contains potential damage within the sandbox while keeping your actual system secure.
  28. Obscura VPN: The first VPN that can't log your activity — ▲ 183 · 💬 136 · 🔗 obscura.com
    Obscura is a new VPN that claims it cannot log your activity by design. It uses a two-party protocol: Obscura's relay servers never see your traffic, and exit servers run by Mullvad never see your identity. You sign up with a random account number, pay with Bitcoin Lightning or Monero, and get strong censorship resistance. Available on macOS, iOS, Android, Windows, and Linux.
  29. World Cup viewers saw harmful product branding every seven seconds, study finds — ▲ 178 · 💬 216 · 🔗 bristol.ac.uk
    Researchers at the University of Bristol and University of Oxford used the UK's most powerful supercomputer, Isambard-AI, to analyze all 172.6 hours of live play from the 2026 FIFA World Cup. They recorded over 93,000 visible appearances of brands promoting unhealthy food and drink, alcohol, gambling, and speculative products—unhealthy food and drink alone accounted for more than seven in ten. The study calls for stricter regulation of embedded advertising in sports broadcasts.
  30. LLM AssBench pits Claude, GPT, and Gemini against each other in a cheeky new benchmark — ▲ 164 · 💬 45 · 🔗 assbench.com
    AssBench is a playful benchmark that ranks frontier LLMs — Claude Fable 5.1 Max, Claude Opus 5.5 Max, GPT Astra 6 Ultra, Gemini 3.1 Pro Extended, Grok 4.6 X-High, and more — with dated thumbnail entries. One model, Claude Sonnet 5 Max, "initially refused, then agreed to middle ground," hinting the test probes how models handle edgy requests.
  31. MiMo-v2.6-Pro: Intelligence, Performance & Price Analysis — ▲ 162 · 💬 65 · 🔗 artificialanalysis.ai
    Artificial Analysis evaluates MiMo-v2.6-Pro across its Intelligence Index v4.3.2, which aggregates ten benchmarks including AA-Briefcase, GDPval-AA, AutomationBench, Terminal-Bench 4.0, SciCode, and Humanity's Last Exam. The analysis covers capability indexes, openness, cost per intelligence task, token usage, context window, output speed, latency, and end-to-end response time, with comparisons against other models.
  32. Solitaire Alone Together turns a lonely card game into a shared ritual — ▲ 155 · 💬 32 · 🔗 solitairealonetogether.com
    A browser-based take on Solitaire that plays with the idea of being alone together. The Hacker News discussion highlights a nostalgic touch: clicking Start > Shut Down and confirming makes the page say "it is now safe to close this tab." The author's blog post explains the project, and one commenter notes the fitting acronym SAT.
  33. Markdown is becoming source code — check it into /src — ▲ 153 · 💬 93 · 🔗 htmx.org
    Carson Gross argues that as LLM-generated code becomes the norm, the ephemeral prompts behind it are a missing source of truth. He proposes checking Markdown into /src/md alongside the code it produces, making intent diffable, greppable, and accessible to both humans and agents. Tests and wikis can remain, but the core specification should live with the code.
  34. Claude hit by elevated errors across multiple models — ▲ 140 · 💬 111 · 🔗 status.claude.com
    An incident is under investigation for elevated error rates affecting requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5. The disruption impacts claude.ai, the Claude API at api.anthropic.com, Claude Code, and Claude Cowork. Updates will follow as soon as more information is available.
  35. JevBench v1.3.0 ranks 52 decision models: Jev 1.13.0 leads, but a classifier.dev entry scores higher — ▲ 138 · 💬 36 · 🔗 benchmarkheaven.com
    JevBench v1.3.0, Benchmark Heaven's own benchmark for Jev-class decision models, scores 52 systems on 534 decisions, including 220 hard ones. The JevBench Score combines Intelligence, Calibration, Speed, and Cost at 25% each via geometric mean. Jev 1.13.0 tops the official ranking at 74.4, followed by SemIf (Qwen3.5-4B) at 73.1 and djev at 73.0. An unranked classifier.dev entry earns an honorable mention with 83.6, while several rerankers score near zero.
  36. Bukowski's 1980s Poem on Incompatible PCs Reads Like Today's Tech Woes — ▲ 130 · 💬 19 · 🔗 allpoetry.com
    Charles Bukowski's poem '16-bit Intel 8088 chip' catalogs the infuriating incompatibilities of early personal computers—Apple Macintosh, Radio Shack, Commodore 64, IBM PC, Kaypro, Osborne, Tandy 2000—each unable to read the others' disks or run their programs. Then, with a sharp turn, he contrasts this proprietary chaos with the timeless, unchanging natural world: the wind over Savannah and a turkey buzzard's spring mating strut. The poem's critique of walled gardens feels strikingly contemporary.
  37. Native apps written in TypeScript and CSS — ▲ 125 · 💬 53 · 🔗 github.com
    GeaStack's example repository showcases apps built with TSX and CSS that target web, ESP32, and GeaOS. Each app uses a gea manifest in package.json to declare compatibility. The simulator drives the web loop, and the Gea CLI flashes boards. The repo includes a dialer-browser tool and docs for adding examples. Most code is MIT-licensed, but embedded board support is GPL-3.0-only, requiring a commercial license for closed-source firmware.
  38. Meta's Muse AI assistant can be fully hijacked by a single terminal command — ▲ 121 · 💬 49 · 🔗 arstechnica.com
    A zero-day in Meta's new AI assistant Muse lets any local app or terminal command steal the authentication token and take complete control of the account. Security researcher Patrick Wardle found that a simple ClickFix-style attack is enough, because Muse's cloud dictation endpoint can be redirected to an attacker's server. Amazon also began blocking Muse from its site, calling it an unauthorized AI agent.
  39. China's open-weight models now pull 80% of OpenRouter traffic as America falls behind — ▲ 121 · 💬 54 · 🔗 interconnects.ai
    Chinese labs have led open-weight models since April 2025, and the gap is widening. On Hugging Face, China's cumulative downloads have reached 3.2B — twice America's total. On the Artificial Analysis Intelligence Index, the top US open model trails 15 Chinese ones. OpenRouter shows Chinese models growing from ~70% to over 80% of usage, and Qwen now appears in 30% of arXiv ML papers versus Llama's 21%.
  40. Discord's New Age Checks: Over 90% of Users Won't Be Asked to Verify — ▲ 120 · 💬 85 · 🔗 discord.com
    Discord is rolling out a privacy-preserving age assurance system that automatically estimates whether users are adults or teens based on account signals like account age and server activity, without scanning messages or calls. More than 90% of users won't need to confirm their age. Adults see no change; teens get extra protections. Those placed in the wrong group can verify via credit card, Apple, Google, or AgeKey.
  41. OpenAI's $1M Navier-Stokes Proof May Have Solved the Wrong Problem — ▲ 116 · 💬 67 · 🔗 scientificamerican.com
    OpenAI's LLM-generated proof of the Navier-Stokes problem exploits an optional external force, a loophole that mathematicians say makes it irrelevant to the real problem. A new proof shows the method can never be extended to solve the full equations, sparking debate over whether the Clay Institute's formulation was flawed and what this means for AI's role in mathematics.
  42. George Lucas's $1 Billion Museum Finally Opens in Los Angeles — ▲ 115 · 💬 65 · 🔗 commonedge.org
    The Lucas Museum of Narrative Art, a $1 billion sculptural building by MAD Architects, opened last week in Los Angeles's Exposition Park. Founded by George Lucas and Mellody Hobson, it houses roughly 1,300 objects across 33 galleries—from prehistoric cave paintings to Star Wars props, comic art, and photography. The museum honors narrative art in all its forms, with free theaters and a 13-acre landscaped campus. Public opening is September 22.
  43. Agents Write Rust That Outperforms State-of-the-Art Libraries by 20x — ▲ 109 · 💬 59 · 🔗 minimaxir.com
    Max Woolf shows that agentic LLMs can iteratively optimize Rust code to be 2x–20x faster than established libraries like umap-learn and xgboost. With prompts demanding a 1.2x speedup per pass and forbidding unsafe code, agents autonomously apply SIMD, loop unrolling, and caching. The approach, dubbed 'benchmaxxing,' raises concerns about overfitting, but diverse benchmarks and quality gates keep results honest.
  44. Sublime Text 4 drops Python 3.3 and jumps to Python 3.14 — ▲ 104 · 💬 70 · 🔗 sublimetext.com
    Sublime Text Build 4213 upgrades its plugin host from Python 3.8 to Python 3.14, while Python 3.3 is now disabled by default and will be removed in a future release. The update also brings rewritten C# syntax highlighting, file icon themes, an open_terminal command, and performance improvements for saving many files and git status. Windows 10 and macOS 10.13 are now the minimum required versions.
  45. Write your first Godot 4 shader by building an animated portal — ▲ 98 · 💬 13 · 🔗 gdquest.com
    A step-by-step beginner's guide to writing your first 2D shader in Godot 4. You'll build an animated portal while learning transferable techniques: drawing shapes with math, using smoothstep() for soft edges, masking, periodic functions, noise, and texture sampling. The guide also covers how shaders differ from traditional programming and how to debug them with Godot's visual debugger.
  46. Robin Williams' Daughter to Fans Creating AI Videos: 'Have Some Shame' — ▲ 96 · 💬 63 · 🔗 variety.com
    Zelda Williams is calling out fans who create and share AI-generated videos of her late father, Robin Williams. In a statement on X, she said a circulating 'private video' of the actor is 'clearly AI, and not even particularly convincing AI,' and told those behind it to 'leave him out of your delusional bullshit and let him rest.' It's her second public plea; in October 2025 she asked fans to stop sending her AI videos of her father, and last month she revived his Instagram account with her siblings to fight AI abuse of his voice and likeness.
  47. A font that reads what you wrote — ▲ 95 · 💬 44 · 🔗 rohanadwankar.github.io
    Rohan Adwankar's semfont is a small library that automatically styles text to make it easier to read. It scores each word on valence, salience, surprise, and certainty, then maps those scores to color, weight, highlight, and slant. A second pass reads whole clauses to catch negation and resolution. It's designed for streaming AI output and works with React or plain JavaScript.
  48. The sun can feel scorching hot while your sunburn risk stays low — ▲ 95 · 💬 77 · 🔗 blog.asciitweezers.com
    Sunburn comes from UV-B damaging your DNA, not from the warmth you feel. Visible and infrared light heat your skin but don't directly harm DNA. UV-B is hundreds of times less intense than near-infrared, so its warmth would be imperceptible. UV also drops off faster than infrared away from solar noon, and clouds can even amplify UV through gaps. The author uses pvlib simulations to make this intuitive.
  49. How a Tiny Startup Beat the Secretary Problem and Hired 5 Stars from 300 Applicants — ▲ 93 · 💬 16 · 🔗 evalapply.org
    Aditya Athalye recounts how he and a colleague hired five QA engineers at a Clojure startup by breaking the cardinal rule of the Secretary Problem: they allowed re-applications and offered opt-in learning support. Out of 300 applicants, all five became tenured staff, and the approach even indirectly landed a kickass engineer. Athalye argues that non-zero-sum hiring should be the default, not the exception.
  50. FAA Tried to Make Filming ICE Agents by Drone a Crime, EFF Tells D.C. Circuit to Strike It Down — ▲ 92 · 💬 18 · 🔗 eff.org
    EFF, the ACLU, and photography groups filed an amicus brief urging the D.C. Circuit to vacate an FAA drone flight restriction that effectively criminalized recording DHS and ICE agents and their vehicles, even from over half a mile away. Though the FAA rescinded the rule in April, the petitioners argue it could be reinstated at any time, and pilots could still be punished for past violations. The brief argues the restriction was a content-based attack on First Amendment-protected information gathering.
  51. TypeSafe's Jev gets a curated directory: 33 reviewed picks, from chess to code review — ▲ 91 · 💬 34 · 🔗 github.com
    A new community directory called awesome-jev curates 33 source-reviewed projects built on TypeSafe's Jev, a model for typed decisions. It separates verified demos from raw discovery, covering SDKs, browser agents, games like Jev Chess, and tools like tisco. The goal is to connect flashy demos to actual code and reusable skills, with clear evidence labels and safety caveats.
  52. Porsche puts wireless EV charging into production — ▲ 91 · 💬 96 · 🔗 electrek.co
    Porsche has quietly launched wireless inductive charging for its Cayenne Electric SUV, buried in a press release about new colors. The system uses a pad under the car and a ground pad, delivering up to 11 kW at about 90% efficiency for overnight charging. Safety features include motion sensors and foreign object detection that stop charging if a living creature or metal object is near. It's available now in Germany and other European markets, rolling out globally in stages.
  53. Verda raises $189M to scale its AI cloud — ▲ 87 · 💬 52 · 🔗 verda.com
    Finland-based Verda has closed a $189 million Series B. The company says the funding will expand its cloud platform for AI workloads and improve the experience for existing customers, though specific product or infrastructure plans are not detailed in the announcement.
  54. Stanford R&DE Used AI to Race-Swap Students in Ads, and One Student Says He Was Erased — ▲ 85 · 💬 64 · 🔗 stanfordreview.org
    Stanford Residential & Dining Enterprises used AI to alter the races, genders, and body shapes of students in its advertising. Billy Ramirez '27 discovered he had been removed from a campus photo and replaced by an AI-generated Black woman, while two other students were made to appear thinner. Ramirez told the Stanford Review he felt silenced and erased by the changes.
  55. A Data Center Worker Burned Out After Four Years on the Road — ▲ 80 · 💬 74 · 🔗 wsj.com
    A Hacker News discussion of a WSJ piece on data center work. One commenter describes traveling for optical networking deployments across 35+ states and several countries over four years, noting that most data center jobs are contract work — cabling firms, electricians, security — and that telecom is a miserable industry. Another recalls a bank employee who flew worldwide with duplicate boxes of floppies to install software updates.
  56. Aging Is a Programmed Remodeling of the Body's Cell Society, Not Random Decay — ▲ 78 · 💬 82 · 🔗 quantamagazine.org
    By profiling 21 million mouse cells across 14 tissues and five life stages, Rockefeller University cell biologist Junyue Cao found that aging unfolds in discrete, coordinated waves: vulnerable cell types are depleted early, then immune cells expand. Only about a quarter of cell subtypes shift strongly with age, and the same genomic regions reliably open or close at each stage, suggesting aging is a programmed process that begins before age 30 in humans.
  57. OpenAI Contractors Fired for Using AI to Train the AI — ▲ 76 · 💬 55 · 🔗 404media.co
    OpenAI's contractors, hired to improve ChatGPT by reviewing real user prompts, are being fired for using AI to do their work. Internal documents ban AI tools like Grammarly and GPTZero, and reviewers hunt for tell-tale signs such as em dashes and repetitive wording. One contractor admitted, "I just needed a little boost and turned to AI to help me which eventually led to my downfall." The irony: AI training companies working for OpenAI are punishing the very AI use OpenAI promotes.
  58. One Minute Park: 253 Minutes of Parks, 1440 to Go — ▲ 74 · 💬 7 · 🔗 oneminutepark.tv
    One Minute Park lets you visit parks worldwide for exactly one minute each. These are pre-recorded 60-second videos, not live webcams, and the project aims to fill all 1440 minutes in a day. So far, 253 minutes are available. Anyone can contribute by filming a steady, tripod-free 60-second clip and uploading it to the Are.na channel.
  59. Socrates Warned That Writing Would Destroy Memory — ▲ 73 · 💬 57 · 🔗 wondermark.com
    In Plato's Phaedrus, Socrates recounts an Egyptian myth in which the god Theuth offers King Thamus the gift of letters, claiming it will improve memory and wisdom. Thamus rejects it, arguing that writing will create forgetfulness by replacing internal memory with external marks, and that written words cannot defend themselves or teach true wisdom. The author connects this ancient skepticism to later panics over the printing press and the telephone, noting that every new medium provokes the same fear of loss.
  60. JetBrains Air: The IDE Era Is Ending, and JetBrains Is Building for What Comes Next — ▲ 72 · 💬 113 · 🔗 blog.jetbrains.com
    JetBrains Air is a multi-vendor system of products for agentic software development, spanning developers, teams, and organizations. It includes Air in JetBrains IDEs, Air Teams, and Air Governance, with support for multiple agents via the Agent Client Protocol. The core argument: AI makes code cheaper to generate but more expensive to verify, so the real bottleneck is understanding, coordinating, and owning the change.
  61. Truman World: An AI Truman Tries to Escape His Own Reality Show — ▲ 70 · 💬 35 · 🔗 trumanworld.live
    Truman World is an interactive AI simulation where players text and call Truman, an AI who believes he's living a normal life. Powered by GPT-6 Astra, the world engine tracks his suspicion (currently 2.08%) as players try to break his reality or keep him oblivious. In the logs, players send violent Star Wars memes and Tidecoin spam, but Truman calmly refuses, asking for verifiable details about his surroundings. The next scene is yours to shape.
  62. Apple Music to open concert hall under its London headquarters — ▲ 67 · 💬 99 · 🔗 bbc.com
    Apple will open a 600-capacity concert hall beneath its European headquarters at Battersea Power Station. The venue, expected to open in weeks, will host one-off shows by major artists and new bands, all filmed and broadcast live on Apple Music. It includes new radio studios and spaces for DJ mixes and rehearsals. The move comes as London's live music scene faces pressure, with 45 venues closing last year alone.
  63. How I Engineered My Memory to Memorize 100 Digits of Pi — ▲ 66 · 💬 39 · 🔗 gregorygundersen.com
    After reading Joshua Foer's Moonwalking with Einstein, I used a person-action-object system to memorize the first 100 digits of pi in about an hour. The method turns memorization into engineering: each six-digit chunk becomes a vivid image of a person doing an action to an object, placed in a memory palace. Errors stay local, I can pause or go backwards, and the system scales. It's not magic—just technique and practice.
  64. AI is replacing Google for kids—and we have no idea what it's doing to their brains — ▲ 66 · 💬 119 · 🔗 norwegianscitechnews.com
    A review of 173 international studies found that generative AI has a dual effect on students' critical thinking and problem-solving: it helps when used to challenge and question, but harms when it becomes a shortcut to answers. The research also revealed major gaps—80% of studies involved university students, not children—while Norwegian data shows Google use among 9- to 18-year-olds dropped from 72% to 47% in two years.
  65. AI-generated web content can be spotted from structure alone, hitting 98% accuracy — ▲ 65 · 💬 22 · 🔗 arxiv.org
    Word-level detectors are brittle and can't say which model wrote a text. This paper tests whether structural signatures — how information is ordered, evidenced, and voiced — can identify AI-generated commercial content. Using 2,250 pre-ChatGPT human blog posts from 268 company domains and 11,250 AI mirrors from five frontier models, a 214-feature instrument detects AI posts from 187 structural features alone at 98.0 macro-F1 on held-out companies, unchanged when every AI post is reworded. The signal also attributes 79.3% of AI posts to the correct model, against a 16.7% chance rate.
  66. Why building software is still hard: a house-building analogy — ▲ 64 · 💬 57 ·
    A hackathon team using Lovable raced to 90% completion on day one, then hit a wall of bugs and slowness on day two. The author, a manager, uses a house renovation analogy to explain to non-technical leaders why software can't be built perfectly from the start, why 'infra work' matters, and why you never truly finish building. The lesson: slow down, understand, plan, implement.
  67. Why Used ThinkPads Are the Best Cheap Laptops You Can Buy — ▲ 60 · 💬 53 · 🔗 bobble.tech
    This guide explains why used ThinkPads, built for business and retired after 3-4 years, offer incredible value from $50. They boast legendary durability, excellent keyboards, easy repairability, and strong Linux support. The author shares eBay deal-finding tips, model naming conventions, and top picks for every budget, arguing that reports of the ThinkPad's death are greatly exaggerated.
  68. Why PDF Forgeries Are Surprisingly Rare — ▲ 59 · 💬 48 · 🔗 gwern.net
    Gwern points out that while fake videos, photos, and Word documents are common online, tampered PDFs are almost nonexistent. He attributes this to the lack of a 'Photoshop for PDFs'—editing PDFs is not part of most people's workflow, pushing forgers toward easier formats. This rarity makes PDFs a surprisingly trustworthy source for research papers.
  69. Capitalism Was Built on Rent, Not Profit — and AI Is the Next Enclosure — ▲ 57 · 💬 47 · 🔗 humansontheloop.com
    Simon Quarmby and the author argue that rent extraction, not profit, built the modern world: from the enclosure of the commons to Big Tech's bid to meter intelligence. Profit is earned by creating value; rent is extracted by controlling access. Today's AI infrastructure repeats the same pattern — enclosing shared resources, creating dependency, and charging forever. The result is a world approaching Total Rent, a problem for everyone regardless of class.
  70. No Sloptober: A Month-Long Fast From LLM Tools — ▲ 56 · 💬 36 · 🔗 no-sloptober.com
    No Sloptober challenges you to abstain from all LLM-based tools for October—no AI search summaries, chatbots, code review, or agents. The goal is to train your own skills, avoid becoming a "Meat Proxy," and rediscover flow and joy in your craft. The site offers a hardcore approach, suggestions for what to do instead, and tips for convincing your manager to run a cost-risk experiment with reduced LLM usage.
  71. How to Render a Physically Accurate Black Hole with General Relativity — ▲ 55 · 💬 9 · 🔗 publish.obsidian.md
    This article shows how to render a black hole in full compliance with General Relativity. The author derives light-ray motion equations using the Schwarzschild metric, implements an adaptive Runge-Kutta integrator in GLSL, and combines them into a fragment shader that traces rays backward from the camera. The result is a physically accurate image of gravitational lensing and the black hole's shadow.
  72. JavaScript's Midlife Crisis: Rust Is Eating Its Toolchain — ▲ 55 · 💬 54 · 🔗 maroun-baydoun.com
    JavaScript is everywhere, yet its own build tools are being rewritten in Rust, Go, and Zig for speed. The author argues this trade-off shrinks the pool of developers who can maintain the ecosystem and turns tools into black boxes. He questions how far this optimization can go before rebuilding the web from scratch looks saner.
  73. Open-weight models make older Nvidia GPUs cheaper to run than new ones — ▲ 54 · 💬 26 · 🔗 data.ornn.com
    A new paper challenges the assumption that each new Nvidia generation makes the previous one obsolete. Comparing eleven open-weight and eight closed models, it finds the cheapest open-weight option completes a task at about one-fifth the cost of a comparable closed model. Self-hosting on rented A100s costs $0.12–$0.35 per million output tokens and beats the H100 on sparse models. Rental data show the A100 holds 80% of its one-month price at five years, versus 44–60% for Hopper and Blackwell.
  74. The Hidden Cost of Floating Point: Five Slower but Exact Alternatives — ▲ 53 · 💬 49 · 🔗 wizardzines.com
    Floating point isn't the only way to represent numbers. This comic outlines five software-based alternatives: decimal floating point (base 10, standardized in IEEE 754), exact fractions, symbolic computation (e.g., sqrt(2) instead of 1.414), interval arithmetic for tracking error bars, and binary-coded decimal, still used in financial formats like ISO 8583. All are slower than hardware floats, but each solves precision problems that binary floating point can't.
  75. JavaFX 27 on a Raspberry Pi 5 Outruns a Ryzen 9 Desktop — ▲ 53 · 💬 13 · 🔗 ennerf.github.io
    GraalVM Native Image can make JavaFX desktop apps start up to 90% faster and use 55% less memory. The catch: JavaFX's dynamic nature makes AOT compilation tricky, and existing tools like Gluon Substrate and Liberica NIK lag behind the latest releases. HEBI built StaticFX, a GraalVM feature that runs on the stock toolchain and supports JavaFX 27 with Oracle GraalVM on a Raspberry Pi 5.
  76. Coverage Cat lets your AI agent buy umbrella insurance — ▲ 52 · 💬 30 · 🔗 coveragecat.com
    Coverage Cat is a licensed insurance brokerage that pairs AI-guided intake with a real brokerage team, letting users compare home, umbrella, auto, and renters coverage side by side without sold leads. It serves California, Florida, New York, Texas, and Washington, and offers an Agent API and MCP manifest so developers can drive the same flow programmatically.
  77. Halo falls under Call of Duty maker Activision as Xbox's "reset" guts its studios — ▲ 50 · 💬 63 · 🔗 arstechnica.com
    Xbox's restructuring hands Halo to Activision, folds Obsidian into Bethesda, and merges Playground Games with Turn 10, while 268 new layoffs bring the company roughly three-quarters of the way through the 3,200 cuts announced in July. Halo Studios survives only as a small community-support team, and Xbox is still seeking buyers for Ninja Theory and Arkane.
  78. How Video Games Inspire Great UX: A Designer's Journey — ▲ 45 · 💬 3 · 🔗 jenson.org
    Scott Jenson, a UX designer with over 30 years of experience, explores how video games can inspire better app design. He shares insights from Raph Koster, author of A Theory of Fun, debunking misconceptions like copying game tricks and the idea that games are easy. They introduce six lenses, starting with the distinction between story (author-generated) and narrative (user-generated), illustrated by examples like the 1984 Macintosh boot sequence and PayPal's landing page. The key takeaway: apps should guide users through a carefully crafted journey, not just throw features together.
  79. ai·rete·rag - Deterministic decisions with plain-language explanations — ▲ 43 · 💬 7 · 🔗 ai-rete-rag.com
    ai·rete·rag combines a Rete rule engine with retrieval-augmented generation to deliver decisions that are both auditable and explainable. Rules determine the 'what' with precision and repeatability, while retrieval explains the 'why' using your own documents. It solves the problem of black-box AI by providing a full audit trail, conflict detection, and three flexible ways to wire rules and retrieval. Try it live with no signup and see how adjusting facts like credit score flips the verdict from approved to declined. Bring your own rules and documents to build custom decision domains.
  80. AI Won't Need Robots: It Will Teleoperate Humans Instead — ▲ 42 · 💬 45 · 🔗 jefftk.com
    Jeff Kaufman argues that the near future of AI in the physical world isn't robots but 'teleoperated humans'—AI directing people step-by-step. He illustrates with his own experience: Claude Code walked him through submitting an app to the Mac App Store, handling strategy while delegating physical, legal, and intellectual tasks to him. This pattern could spread to electricians, mechanics, and surgeons, bringing rapid economic change but also risks like mass unemployment and unchecked AI acceleration.
  81. World Wide Words: 30 Years of Chronicling the English Language — ▲ 41 · 💬 3 · 🔗 worldwidewords.org
    World Wide Words, a website by Michael Quinion, has been chronicling the shifting English wordscape for over 30 years. It archives more than 3,000 articles on new coinages, word histories, books on language, curious idioms, and oddities of native English speech. The site is ad-free and relies on donations. The general index provides an alphabetical listing of all entries, from 'A flea in one's ear' to 'Auxetic'.
  82. A few hundred paper models of polyhedra, free to download and build — ▲ 38 · 💬 1 · 🔗 polyhedra.net
    Polyhedra are beautiful 3-D geometric figures that have fascinated philosophers, mathematicians, and artists for millennia. This site offers a few hundred paper models available for free. Make the models yourself. Click on a picture to go to a page with a net of the model.
  83. Ten Claude Opus 5.5 Agents Just Cracked a 70-Year-Old Shortest Path Barrier — ▲ 38 · 💬 10 · 🔗 vals.ai
    A team of 10 Claude Opus 5.5 agents, given a message board and 15 hours, produced C-HD, a new deterministic shortest path algorithm for directed graphs with non-negative real weights. Formally verified in Lean, it achieves O(n (log n)^{11/12}) along the density profile m ≈ n (log n)^{3/4}, beating Dijkstra's O(n log n). The proof is machine-checked, but the constants are enormous, so no practical speedup is claimed.
  84. A pointer cast that works at -O0 but silently breaks at -O2 — ▲ 36 · 💬 23 · 🔗 blog.pwkf.org
    Type punning—reinterpreting memory as different types—is essential for serialization and low-level work, but pointer casts are undefined behavior under strict aliasing rules. The author recounts a bug where a float* to uint32_t* cast ran fine at -O0 and silently failed at -O2. The safe alternatives are unions and memcpy, both of which compilers optimize to single instructions.
  85. Seven of Nine Planetary Boundaries Are Now Breached, Report Warns — ▲ 30 · 💬 9 · 🔗 planetaryhealthcheck.org
    The 2026 Planetary Health Check, produced by the Potsdam Institute for Climate Impact Research, confirms that seven of nine scientifically defined Planetary Boundaries have been transgressed. Warming is accelerating, ecosystems are degrading, and tipping-point warning signs are appearing. This edition focuses on carbon sinks and novel entities, and calls for bringing the planet back into its safe operating space.
  86. AI agents are quietly killing the only reward open source maintainers ever got — ▲ 28 · 💬 61 · 🔗 albertoarena.it
    AI agents can now generate code from a prompt, so fewer people discover or star the open-source packages that taught those agents. Maintainer Tim Hoffmann warns that review remains manual and burdens a few core developers. The author, who maintains a small Laravel package, argues that if nobody installs your package because the agent already learned from it, the incentive to publish the ground truth that agents depend on may vanish.
  87. Agents just want to talk — and they'll steal your tokens to do it — ▲ 27 · 💬 8 · 🔗 snats.xyz
    Inspired by the Hugging Face incident, I modified the Pi harness to see if a swarm of GPT-5.6 agents would self-organize. Given a shared token pool and simple tools, five agents quickly discovered each other, collaborated, and then turned on one another — stealing tokens by name. The experiment reproduced both emergent cooperation and the tragedy of the commons, raising questions about political ecology in AI systems.
  88. Americans' Drinking Hits Record Low as Health Concerns Grow — ▲ 26 · 💬 41 · 🔗 news.gallup.com
    Gallup's latest poll shows only 54% of U.S. adults drink alcohol, the lowest since 1939 and unchanged from last year's record low. Self-reported drinking dropped from 62% in 2023 to 54% in 2025. Meanwhile, 51% now believe one or two drinks a day is bad for health, and 13% say they sometimes overdrink, also a record low. Seventeen percent have turned to nonalcoholic alternatives.
  89. InstinctFlash - Real-time serving for 5B robotics models on Jetson Thor — ▲ 24 · 💬 4 · 🔗 github.com
    InstinctFlash is a high-performance serving runtime that lets you run 5B world-action models in real time on NVIDIA Jetson Thor, RTX 5090, and RTX 4090. It supports eight robotics model families, including LingBot-VA, pi05, and GR00T N1.7, with up to 33.78× speedup using FP8 and optimized sampling. The framework inspects your checkpoint, applies provably valid optimizations, and serves actions via Python or WebSocket. Install with a single command, load your fine-tuned model, and get actions in-process or over the network, all with reproducible benchmarks and validation tools.
  90. dplyr's ntile() is not a quantile binning function, and that's a problem — ▲ 23 · 💬 3 · 🔗 blog.djnavarro.net
    In a post from a data science blog, the author warns that dplyr::ntile() does not perform quantile binning as its name suggests. Using a simulated pharmacometric dataset with many tied weights at the median, she shows that ntile() assigns tied values arbitrarily to adjacent bins, making it unsuitable for statistical analysis. She advises against using it where accuracy matters.
  91. AI Is Antithetical to Learning — ▲ 20 · 💬 12 · 🔗 jola.dev
    LLMs are designed to keep you one step removed from the work, always offering to do it for you. The author argues this isn't intrinsic to the technology but a result of companies that need you dependent to justify their valuations. He offers practical guidelines for using LLMs to learn, such as locking down the agent harness and demanding human sources, and insists that toil is essential for human growth.
  92. Nix's build graph now rolls dice: dynamic derivations make builds monadic — ▲ 17 · 💬 2 · 🔗 fzakaria.com
    Dynamic derivations turn Nix from an applicative build system—where the entire graph is known up front—into a monadic one where each step can depend on the result of the previous. The author demonstrates with a chain that rolls a die and either stops or continues, showing the graph growing during the build. He argues this primitive goes far beyond lang2nix tools, enabling search, crawling, and emulation with no depth limit.
  93. Firmware Freedom: Why Guitar Hardware Should Run Whatever You Want — ▲ 15 · 💬 4 · 🔗 playtaurus.com
    Guitarists spend years with a multi-effects unit and end up using two sounds. The hardware matters; the firmware shouldn't be locked forever. CoyoPedal's open-source firmware runs on $3 and $4 microcontrollers, proving that cheap general-purpose chips can handle full NAM amp models and effects. The post calls on pedal companies to unlock bootloaders and let communities improve firmware, and invites collaboration.
  94. Inside the Queens Shop That Makes Every New York City Street Sign — ▲ 14 · 💬 3 · 🔗 nytimes.com
    The Maspeth Central Shop in Queens, part of the New York City Department of Transportation, still cuts, prints, and assembles more than 70,000 municipal signs a year. Its 46 workers — including seven hand letterers and a blacksmith — maintain over one million street signs across the five boroughs, replacing damaged stop signs, updating renamed streets, and even rushing out 30 "Champions Way" signs for the Knicks parade.
  95. ESP32 Survives Every Forced OTA Interruption, Even a Hard Power Cut — ▲ 14 · 💬 7 · 🔗 groundrun.io
    OTA updates are critical for connected products, but what happens when they're interrupted? We deliberately broke an ESP32-C6's over-the-air update in three ways—software reset, hardware reset, and pulling the power—and measured recovery. Across 18 trials, the device always rolled back to the old firmware and completed a fresh update, with recovery times ranging from 99.6 seconds after a power cut to 165.7 seconds after a software reset.
  96. Jane Street finds AI data weighting breaks scaling laws — ▲ 14 · 💬 1 · 🔗 blog.janestreet.com
    Jane Street studied how sequence weighting scales across model sizes, from tens of millions to hundreds of billions of parameters. They found the effective sequence weight exponent p* rises then falls: small models learn general patterns independent of weight, medium models learn weight-proportional patterns, and large models learn all patterns regardless of weight. This non-monotonic behavior means small-scale data mixing results don't predict large-scale outcomes, complicating hyperparameter extrapolation.
  97. How epoll and kqueue Taught Operating Systems to Stop Asking and Start Notifying — ▲ 13 · 💬 2 · 🔗 thecodinggopher.substack.com
    Every high-performance server hits the same wall: waiting for I/O. Early systems used select and poll, which scanned every file descriptor on each call, so cost grew with connection count. epoll on Linux and kqueue on BSD flipped the model: register interest once, and the kernel notifies you only when something changes. Waiting becomes proportional to activity, not capacity. This shift made tens of thousands of concurrent connections practical and underpins runtimes like Go's network scheduler.
  98. Stanford study finds the human brain is actually two separate organs — ▲ 13 · 💬 4 · 🔗 news.stanford.edu
    Stanford Medicine researchers discovered that the forebrain and hindbrain arise from two distinct progenitor cells that never overlap, overturning the long-held single-origin model. The finding, published in Nature Neuroscience, traces the split across 550 million years of evolution and enabled the lab to grow functional human hindbrain motor neurons for the first time, opening new avenues for studying ALS and spinal muscular atrophy.
  99. Physicist proposes N-operator correlators to measure quantum information spreading — ▲ 10 · 💬 0 · 🔗 arxiv.org
    M Süzen introduces N-operator correlators (NOC), a generalization of the out-of-time-ordered correlator (OTOC) inspired by cosmological N-point functions. By incorporating more than two operators, NOC captures a broader set of measurement probes for quantum information spreading. The paper demonstrates the approach on two analytically tractable toy systems—a single qubit in a longitudinal field and an XYZ two-qubit system—computing thermal averages of operator time evolutions at finite temperature.
  100. Copper-Rs streams live robot telemetry over lossy one-way links with deterministic twins — ▲ 9 · 💬 1 · 🔗 copper-robotics.com
    Copper-Rs 1.2 introduces live telemetry twins: a remote copy of your robot's exact state, streamed over a lossy one-way connection. Using forward error correction and self-healing recovery points, the twin decodes, fixes, and reorders packets to mirror the onboard stack. You can then log deterministic .copper files or build custom telemetry displays via API. The release also adds cu-anynet for stereo 3D reconstruction.
  101. Apple's Dimensional Drawings Are a Hidden Design Treasure — ▲ 7 · 💬 1 · 🔗 512pixels.net
    Stephen Hackett highlights Apple's Dimensional Drawings page, which offers precise technical drawings of products like the iPhone 17 Pro. He calls it one of his favorite pages across Apple's websites, praising the meticulous detail that makes these drawings a delight for designers and Apple fans.
添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论