HN digest

FAA issues national ground stop – lifted 4.5 hours laternbcnews.com · 1622 points · 1251 comments · 3dspacepolicyThe FAA has issued a national ground stop, halting flights across the US; the advisory page itself could not be read, so details are unconfirmed.

What HN thinks

  • Commenters report the advisory page was flaky, sometimes showing no advisory, and an official ATCSCC notice later extended flight plan drop times and cited a nationwide communications failure.
  • Several commenters point to radio and radar communication systems as the likely fault, with one noting the ground stop is tied to ARTCC frequency problems rather than airport towers.
  • Pilots on the ground reported being told only that an ATC radio frequency issue was involved, and one noted the outage is the third nationwide grounding since 9/11, counting 9/11 itself.

The article couldn't be read, so this summary is based on the discussion.

LibreOffice actually works 365 days a yeartwitter.com · 802 points · 250 comments · 3dopen-sourcetoolsThe LibreOffice account posted a social media message claiming the suite works 365 days a year, a jab at Microsoft Office's subscription model or the usual dig at paid software.

What HN thinks

  • Commenters largely agree that offline desktop suites like LibreOffice and offline Microsoft Office versions keep working during cloud outages, and that the jab is mostly about the cloud-first subscription model.
  • The strongest caveat is that LibreOffice still trails on UI polish, spreadsheet handling of large files, and full Office compatibility, which commenters say drives people to Google Docs or Office anyway.
  • One commenter noted that offline Office 2016 or 2021 licenses eventually lose updates, so the subscription criticism is not the whole picture, while others argued that paying once for software is a fair model.
Reverse engineering my Coway air purifier to ditch the cloudscottjg.com · 768 points · 215 comments · 4dsecurityhardwareThe author reverse-engineered a Coway Airmega IconS purifier's Bluetooth provisioning and found it fetches its root CA over plain HTTP, letting him replace Coway's cloud with a local Home Assistant integration.

Key points

  • The fan downloads its root CA certificate over plain HTTP, so it could be intercepted without exploiting the eMudhra certificate authority incident.
  • Provisioning sends Wi-Fi credentials and server URLs over Bluetooth, encrypted with an AES key built from 20 random letters plus the MAC address.
  • Because the author writes the provisioning payload, he controls the CA and certificate server URLs, avoiding any man-in-the-middle setup.
  • An AI assistant reconstructed the Bluetooth protocol from Android logs and decompiled app code in about 20 minutes.

What HN thinks

  • Commenters broadly welcomed local control of cloud-dependent devices, while noting many buyers simply want a dumb model and that manufacturers often subsidize hardware with subscriptions and data collection.
  • Several commenters questioned why an air purifier needs Wi-Fi at all, though others said remote monitoring and alerts are convenient, and one noted that dumb Coway models exist.
  • One commenter pointed out that the plain-HTTP root CA fetch likely came from a firmware design that avoided certificate expiry or update problems, a pattern they had seen in consumer hardware development.
Nvidia employees petition to bring back Jensen Huangwired.com · 715 points · 569 comments · 4dbusinessaiWired reports that Nvidia employees are petitioning the board to reinstate Jensen Huang after he resigned amid multiple reported extramarital affairs, which have cost the company about $1 trillion in market value.

Key points

  • An internal suggestion-box post by a longtime engineer, submitted before Huang's resignation, has drawn more than 1,600 comments, mostly supportive of his return.
  • Huang wrote in a Sunday employee email that he had been unfaithful to his wife and that there was 'no excuse.'
  • Reports allege affairs with a suspected Chinese intelligence asset and with direct reports, and Reuters later reported another alleged affair.
  • Investigations by US lawmakers and the FBI are underway, and new chip export controls are being discussed.

What HN thinks

  • Many commenters argue the affairs themselves are a private matter, but the alleged link to a possible Chinese spy and suspected chip diversion to China is what makes the situation serious.
  • Several commenters stress that affairs with direct reports violate corporate policy and show poor judgment, separate from any national security questions.
  • Commenters are divided on whether the stock drop and employee reaction reflect real business harm or a mix of financial self-interest and political maneuvering, with some questioning the reliability of the reporting.
Elevated errors on claude.ai, Claude Code, and platform.claude.comstatus.claude.com · 482 points · 382 comments · 4daitoolsAnthropic's Claude services, including claude.ai, Claude Code and the Console, had elevated errors on Nov 13, 2026, with users unable to authenticate; the incident was investigated and later resolved.

Key points

  • The status page says users were failing to authenticate to Claude.ai and Claude Code, and the incident was posted at 09:58 UTC.
  • Affected products listed were claude.ai, Claude Console (platform.claude.com), Claude Code, and Claude Cowork.
  • Commenters reported OAuth errors such as 'Temporarily unable to authenticate' and forced re-logins, plus Cloudflare 403 challenge pages.

What HN thinks

  • Commenters widely link the outage to the concurrent Microsoft 365 and Outlook outage, with some speculating a shared dependency such as Entra authentication, though this was not confirmed.
  • Several users noted that the login failures pushed people toward email magic links, which they criticized as fragile when the email provider is also down, and asked for passkeys or password options.
  • Some commenters observed that the fix seemed quick at about two hours, but others reported the problem persisting for connectors and that a second incident followed.
Doomsday: Friday, 13 November, A.D. 2026 (1960)science.org · 465 points · 320 comments · 4dsciencehistoryA 1960 Science paper projected that world population would reach a doomsday point on Friday, 13 November 2026, based on extrapolating past growth. The article itself could not be fetched.

Key points

  • The article was unavailable (HTTP 403), so the specific methodology and figures could not be verified from the source.
  • The headline date comes from a 1960 Science publication that extrapolated historical population growth trends forward.

What HN thinks

  • Commenters widely note the 1960 model's hyperbolic extrapolation was wrong, since population growth rates fell and the curve stopped fitting shortly after publication.
  • Several commenters argue the paper was partly tongue-in-cheek, pointing to the author's birthday as the chosen date and a dry writing style.
  • One commenter cites the Hans Rosling argument that population will settle around 11 billion as child mortality falls and women's education and workforce participation rise.

The article couldn't be read, so this summary is based on the discussion.

Jonathan Blow – If AI solved coding, why is my software still the same? [video]youtube.com · 432 points · 607 comments · 3daiprogrammingJonathan Blow argues in a video that AI coding tools have not visibly improved software, asking why his own software still looks the same; the article itself was not available.

What HN thinks

  • Commenters broadly agree that the hard parts of software, such as deciding what to build, modelling the problem, and architecture, are not solved by AI coding tools.
  • Several commenters report that AI has noticeably changed their personal software, especially small tools replacing subscriptions, while public-facing software has changed less.
  • A recurring caveat is that AI-generated code often becomes hard to maintain or review, with one commenter noting that unreviewed code can contain avoidable failure modes.

The article couldn't be read, so this summary is based on the discussion.

Microsoft Outlook/365 still down after 12 hourstwitter.com · 406 points · 286 comments · 4dsystemsbusinessMicrosoft Outlook and Microsoft 365 have reportedly been down for 12 hours, according to a post from the Microsoft 365 Status account on X.

What HN thinks

  • Commenters widely see the outage as a sign of over-centralization, with many noting that most of their contacts use Microsoft 365, so self-hosting or switching offers little practical protection.
  • Several commenters link the outage to layoffs and AI-driven engineering changes, suggesting lost institutional knowledge may be slowing diagnosis, though this is speculation rather than confirmed cause.
  • One commenter notes that email is designed to tolerate outages because sending servers queue and retry messages for days, so mail may be delayed rather than lost.
Terminal borders that don't cost a celljoshka.net · 390 points · 168 comments · 3dtoolssystemsJoshka proposes a terminal protocol that draws thin border lines on cell edges, with per-edge color, width and style, so borders take no rows or columns; a working Ghostty patch exists.

Key points

  • Each cell carries up to twelve line segments (halves of four edges plus four center arms), each with its own color, width, placement and style.
  • Shared edges merge by priority, then width, then right/bottom position, so a focused pane can keep its full outline.
  • Ghostty change is about 1,650 lines; cells stay the same size since the 16-bit border ID fits existing padding.
  • Worst-case shared border table is roughly 10-15 MB, since entries are never freed; rounded-corner rendering cost is not yet benchmarked.

What HN thinks

  • Several commenters suggested overstrike mode (DECSET 20, supported by mintty) as a simpler alternative that draws border glyphs over existing cells without a new protocol, though one noted it makes rounded corners with filled interiors harder.
  • A recurring concern is compatibility and fragmentation: one commenter noted terminal implementers would resist adding per-cell border state, and another questioned whether tmux would pass the sequences through.
  • One commenter raised a concrete CJK issue: box-drawing characters are East Asian Ambiguous width, so cell-based borders would avoid misalignment, but grapheme-to-cell mapping would still need specifying.
AI unlocks Gödel's notebooks, said to contain hundreds of proved theoremstheguardian.com · 389 points · 97 comments · 3daimathHelsinki logician Jan von Plato reports that Claude extracted 663 Gödel theorems from his 1940s notebooks, and Coq verified about 400 as stated, with about 200 repaired and 50 false.

Key points

  • Gödel's 16 early-1940s workbooks total 1,009 pages, largely in obsolete Gabelsberger shorthand that few people can read.
  • Progress had been about one page a week before Claude produced a 650-page digest of the material within a couple of days.
  • Claude reconstructed a proof of the axiom of choice's independence that was one definition away from correct, and Coq accepted the reconstruction.
  • Von Plato believes Gödel would have won a Fields medal had he published the axiom of choice result, which Cohen received in 1966 instead.

What HN thinks

  • Commenters note the notes were hard to read: von Plato reportedly one of very few people able to read Gödel's Gabelsberger shorthand, and the work was a multi-year ERC-funded project.
  • Several commenters argue that basic OCR or computer vision could have helped with the manuscripts, so the AI's contribution may be less novel than the headline suggests.
  • Some commenters point out that Gödel was a notoriously harsh self-critic, so they doubt many of the recovered results were unknown insights rather than already-understood or incomplete material.
$11.4M in Nvidia hardware vanishes during shipment from CA to TXfreightwaves.com · 352 points · 193 comments · 3dbusinesssecuritySix pallets of Nvidia computer hardware, reportedly worth $11.4 million, vanished on a truck traced from Oakland toward Fort Worth, Texas, in late August; police have not confirmed the value.

Key points

  • The truck was collected in Oakland on Aug. 29 and tracked through several extended stops before its trail was lost in San Bernardino County, Southern California.
  • The vehicle reportedly displayed a food transportation company's logo during pickup, and police have not established whether criminals impersonated a carrier.
  • Oakland police opened a grand theft investigation after a complaint filed Sept. 13, which does not establish when the theft occurred.
  • No arrests or recovery have been announced, and Nvidia had not responded to FreightWaves' requests for comment.

What HN thinks

  • Commenters largely agree that high-value cargo theft is a recurring problem in logistics, with several describing insider tips to drivers and social-engineering of pickup numbers as common methods.
  • One commenter with logistics experience says the main risk is disclosure of pickup numbers, which lets thieves pose as carriers and have staff load the goods onto their truck.
  • Some note that negotiated freight terms such as FOB origin versus FOB destination may determine who bears the loss, and skeptics question whether a $11.4M loss matters much to a well-funded buyer.
IceWMice-wm.org · 349 points · 202 comments · 4dtoolsopen-sourceIceWM is a lightweight X11 window manager, first released in 1997, with a taskbar, configurable keybindings, many focus modes, and multi-monitor support via RandR and Xinerama.

Key points

  • The latest release is 4.1.0, dated 2026-08-06, and the project moved its development to GitHub.
  • It ships an optional external wallpaper manager with transparency support and a simple session manager.
  • Over 450 themes are available from box-look.org, and the project is translated into about three dozen languages via Weblate.
  • Installation is a standard configure and make build, with an os-depends.sh script to install dependencies.

What HN thinks

  • Several commenters still use IceWM daily, mainly on headless remote desktops, audio workstations, and older hardware, praising its low resource use and out-of-the-box setup.
  • Criticism centers on dated visuals, nested menus, and configuration friction, such as difficulty removing workspaces, plus a lack of Wayland support as a blocker for migration.
  • Commenters point to alternatives like JWM for window snapping, Fluxbox, AwesomeWM, and openbox, and one notes IceWM's lead developer is slow to address feature requests.
Edge0-145B: Frontier AI Capabilities on Your Own Computeredge0.ai · 326 points · 97 comments · 4daitoolsEdge0 open-sourced Edge0-145B, a 145-billion-parameter model it says runs locally on a 48GB laptop and matches some cloud models on coding and reasoning benchmarks, per its own tests.

Key points

  • Only about 10B of the 145B parameters are active per inference step, with each group of 64 experts sharing two large matrices.
  • Reported figures on a compatible laptop are 29GB peak memory, 80+ tokens/s decode, and roughly 120W typical power draw.
  • Roughly 40GB of experts sit in system memory, and about 48GB of vocabulary data is read from SSD on demand.
  • Benchmark results are Edge0's internal evaluations, and Claude-Opus-4.6 has no public DeepSWE 1.1 score for comparison.

What HN thinks

  • Commenters questioned the headline claim: one noted the 29GB peak memory figure conflicts with the article's own mention of about 40GB extra system memory for experts plus 48GB of SSD storage for vocabulary.
  • Several commenters said the model appears to be a quantized, modified version of Qwen 3.8 Flash Next, with one noting the comparisons used a quantized baseline, which may flatter the results.
  • Practical limits were raised: the inference engine reportedly supports only CUDA for now, the NVFP4 release is Blackwell-only, and HIP/AMD support is reportedly broken, so Mac and other users may not be able to run it as advertised.
It Takes 16 Clicks for a Student to Open a PowerPoint on an iPadreddit.com · 301 points · 157 comments · 3dtoolscultureA Reddit post from a teacher claims that a student needs 16 clicks to open a PowerPoint file on an iPad; the article text was unavailable, so details are unverified.

What HN thinks

  • Commenters largely agree the login and tapping friction is a symptom of layered school software such as Clever, Canvas and SSO integrations, not of the iPad itself.
  • Several argue that the single-sign-on and third-party integrations, rather than Google Classroom or the device, cause most of the breakage, and that purchasing administrators rarely use the tools they buy.
  • Some defend the security tradeoff of frequent logins, noting long-lived session cookies could be stolen, while others say the daily re-login burden is excessive.
Live airband on all frequencies from KORS. Victoria, CA also ground stoppedorcasairband.com · 286 points · 67 comments · 3dtoolsspaceA website streams live air-traffic radio audio from all frequencies at KORS, a station in the Victoria, Canada area, which is also under a ground stop.

What HN thinks

  • Commenters were impressed that the transcription stays readable even when the raw radio audio is mostly crackle, with one saying the output seemed plausible to what they heard.
  • Several commenters noted that online recordings are often more garbled than what pilots and ATC actually hear, since aircraft have line of sight and ATC uses purpose-built receivers.
  • The thread debated why aviation still uses analog AM simplex radio, with one commenter citing legacy certified fleets and pilots hearing each other, and another raising accessibility for people with auditory processing differences.
AI is destroying the trust between writers and readerstheatlantic.com · 285 points · 177 comments · 4daicultureA Atlantic essay argues AI-generated books sold under real or invented author names are eroding reader trust, though the article itself could not be fetched (HTTP 403).

What HN thinks

  • Commenters broadly agree that low-effort AI books and slop are flooding platforms like Amazon and making discovery harder, with one noting self-published authors struggling against bulk uploads of cheap, junk titles.
  • A recurring caveat is that trust in writing and media was already fragile before AI, with several arguing that AI mostly exposes existing problems such as weak editorial standards and paid or formulaic content.
  • One commenter who writes fan fiction and web serials says most post-2024 submissions show obvious AI tells, and another says they now research authors and publishers before buying books published after 2022.

The article couldn't be read, so this summary is based on the discussion.

After 600 experiments, here's what I discovered about JEPA world modelshaddyalchaer.com · 279 points · 46 comments · 3daiscienceA solo researcher reports four planning methods that raise LeWM's Push-T success on long-horizon tasks from 22% to 85% without retraining, using subgoal chaining and a memory graph of demonstrations.

Key points

  • Splitting a 75-step goal into subgoals every 25 steps lifts LeWM success from 22% to 80%, close to exact physics planning on the same tasks.
  • Adding a cost to jumps between demonstrations in the memory graph raises success from 62% to 85%, matching expert subgoals.
  • Standard Push-T test goals mostly lie on paths demonstrations already travel, since every demonstration ends in the same place; 84% of goals had a near-identical stored state.
  • Scoring goal images that include the pusher position caused failures; leaving the pusher out raised exact-physics success on 50 hard tasks from 9 to 39.

What HN thinks

  • Several commenters questioned the evaluation design, noting the goals come from expert demonstrations, and asked whether arbitrary start and goal poses would be a fairer test of generalization.
  • One commenter pointed out that the reported gain depends on transformed demonstration segments filling gaps, which suggests the base model fails to generalize without them.
  • Commenters noted that the planning approach resembles Model Predictive Control and that TD-MPC and MCTS already solve Push-T without training, so the novelty is limited.
FM-1 Workbenchopen.m-vave.com · 262 points · 172 comments · 3dtoolshardwareM-VAVE's FM-1 Workbench is a browser-based editor and firmware flasher for the FM1, a low-cost 8-track polyphonic synth with six sound engines including DX7-style FM.

Key points

  • The page offers offline and simulated editing, Web MIDI sync to the device, DX7 SysEx import, and an 8-slot sampler library of about 19.55 seconds shared across projects.
  • Firmware flashing from the page can restore the original V15 firmware, but it overwrites the four patch banks and makes Open projects and samples incompatible.
  • Forced flashing of other firmware can erase data or brick the unit and may require a dedicated recovery tool.

What HN thinks

  • Commenters broadly agree the FM-1 is a fun, cheap platform, with many saying custom firmware is its main appeal, though some argue grooveboxes add little beyond what a laptop and MIDI controller already do.
  • The main caveats raised are build quality, the 1/8-inch MIDI jacks needing adapters, and the risk of bricking the device when flashing firmware from a browser tab that gets closed mid-flash.
  • One commenter with first-hand experience noted that a common failure in similar devices is that the control surface gets treated as an afterthought, with touchscreens poor for performance and good rotary encoders needed.
Global Microsoft Outagecnn.com · 255 points · 156 comments · 4dsystemsbusinessA Microsoft cloud outage has knocked out Outlook email and Teams for more than 12 hours across dozens of countries, with Microsoft saying it is not a cyberattack and recovery is not yet sustained.

Key points

  • Downdetector reported outage reports from 74 countries and territories within about an hour of onset.
  • Microsoft said it was observing only partial recovery in regions where it applied server adjustments, not sustained recovery.
  • Government ministries in Japan could send little email most of the workday, so staff created physical copies of documents.
  • Australian schools texted parents to phone in absences because email systems were down, and some job interviews were missed.

What HN thinks

  • Commenters widely argue the outage shows the risk of centralising email and other services on a single provider, though many note that switching costs and the familiar 'nobody gets fired for buying Microsoft' mindset make real change unlikely.
  • Several point out that self-hosting email has been reliable for some, but one commenter notes that running it well needs a team sized for quick recovery, which many organisations lack.
  • Some commenters question whether the outage is related to a recent update or AI-assisted code changes, while others say the scale of distributed systems makes such failures common without any attack.
Brave Search API now accepts agentic payments via the x402 protocolbrave.com · 254 points · 141 comments · 4daiwebBrave Search API now accepts per-request USDC payments on Base via the x402 protocol, with no account or API key needed, at $5 per 1,000 queries.

Key points

  • Unpaid requests receive an HTTP 402 response stating the price; the agent signs a USDC payment and retries with it attached.
  • Brave runs its own facilitator; currently only USDC on Base is accepted, with other networks and stablecoins said to be coming.
  • Gas fees on Base are covered for a limited time, per the announcement.
  • Purl Wallet, a Stripe-developed CLI wallet, is the suggested easiest way for agents to pay.

What HN thinks

  • Commenters are skeptical that x402 pay-per-request will catch on, noting that few have actually seen transactions go through and that stablecoin funds carry freezing and trust risks.
  • Pricing draws pushback: several commenters say $5 per 1,000 queries is far above what some alternatives charge, such as Exa at $4, and that search costs add up quickly for agents.
  • One commenter warns that frictionless payments could enable runaway agent spending, since agents often re-run near-identical searches, and suggests per-task budgets and caching.
Found: A protein that drives ageing without unleashing inflammationnature.com · 241 points · 75 comments · 3dsciencehealthMice engineered to lack the enzyme granzyme K showed improved heart and liver function and less kidney damage in old age, without the inflammation usually linked to ageing.

Key points

  • Aged mice without granzyme K had less age-related stiffening and scarring of liver tissue.
  • The researchers report the protein drives ageing in tissues while leaving immune defences intact, according to the Nature summary.
  • Drugs lowering granzyme K levels are proposed as a possible future approach, though none are described as existing or tested.

What HN thinks

  • Several commenters debated whether indefinite healthy lifespan would be desirable, with one arguing that the thought experiment ignores the cost to society and the persistence of older minds despite a younger body.
  • One commenter pointed out that the paper's abstract is more nuanced than the headline, saying GzmK drives organ aging in specific tissues but not immune aging or systemic inflammaging.
  • A commenter noted that the study's finding that GzmK does not affect immune aging suggests lifespan may change little even if healthspan improves, since old lab mice often die of cancer.
Show HN: Isometric SF – A pixel-art map of the SF Bay Areaisometricsf.com · 236 points · 74 comments · 3daidesignAn interactive isometric map of roughly 1300 square miles of the SF Bay Area, rendered in a pixel-art style from Google Photorealistic 3D Tiles imagery.

Key points

  • The map was generated from Google Photorealistic 3D Tiles and stylized into a cohesive isometric pixel look covering about 1300 square miles.
  • A Qwen-Image-Edit LoRA was first trained using Nano Banana 2 outputs to prototype the style and generation method.
  • The final map used a Klein 4B LoRA trained through Qwen, which the author says reproduced the style at a quarter of the cost of Nano Banana 2.
  • The author estimates that generating the full map with Nano Banana 2 would have cost around 7k USD.

What HN thinks

  • Commenters broadly agree the map looks good from a distance but is not true pixel art: zoomed in, it shows JPEG-like artifacts, ghosting, wavy lines, and structural errors such as blended building colors.
  • Several commenters note the map omits large parts of the nine-county Bay Area, including areas east of I-680 and north of the Richmond Bridge, such as Mt. Diablo and Livermore.
  • One commenter who previously built a similar AI-generated pixel map of Stockholm says the usual fix is palette reduction and square grid alignment, which this project's tiles appear to lack.
Eli Lilly CEO pushes back on Dario Amodei's medical claimsgarymarcus.substack.com · 230 points · 283 comments · 3daihealthGary Marcus argues Anthropic CEO Dario Amodei's claim that cancer, Alzheimer's, heart disease and mental illness will be cured within a decade is unrealistic, and that Eli Lilly's chairman and CEO Dave Ricks publicly disagreed.

Key points

  • Ricks said the map of biology is still largely blank and randomized human trials are the long pole, so progress will likely be steady and incremental.
  • A commenter cited that the median time from first-in-human testing to FDA approval for recent drugs was about 10 years, mean 14.
  • Marcus links the remarks to Anthropic's IPO launch and argues the media has not deconstructed the claims.
  • Several critics, including a precision oncology CEO, said clinical timelines alone make a cure within a decade very hard.

What HN thinks

  • Commenters broadly doubt the ten-year timeline, pointing to slow clinical trials, FDA approval, and the physical lab bottleneck rather than intelligence alone as the real constraint.
  • Several commenters note the article's framing overstates the claim, since Amodei reportedly said these diseases would become 'relics of the past' rather than being cured within a decade.
  • A biologist-style caveat raised in the thread is that biological measurement is far noisier than software assumes, so reproducibility problems could slow AI-driven drug discovery.
A technical vision for the Haskell ecosystemdiscourse.haskell.org · 223 points · 140 comments · 3dprogrammingopen-sourceA Haskell community post outlines a technical vision for the language's ecosystem; the article could not be fetched, so its specific proposals are unknown.

What HN thinks

  • Commenters broadly agree Haskell's adoption is limited by its ecosystem and onboarding friction, including confusing installation options (ghcup, stack, cabal, nix) and slow initial builds.
  • Several commenters argue Haskell is poorly positioned against Rust and newer strict languages like Lean for AI-assisted coding, since it is neither different enough to justify the switch nor aligned with new verification opportunities.
  • One commenter who has used Haskell for CAD file tooling reports that type checking provides a useful feedback loop for working with coding agents, while another notes that LLMs produce more reliable Rust than Haskell due to training data volume.

The article couldn't be read, so this summary is based on the discussion.

Big tech is here to let us downblog.jsweeting.me · 196 points · 171 comments · 3dbusinesspolicyA personal blog post argues that a recent Microsoft Outlook, Teams and Copilot outage shows the risk of depending on Microsoft, especially after layoffs.

Key points

  • The outage reportedly hit Outlook, Teams and Copilot across two separate days, affecting businesses, schools, universities and courts.
  • The author argues Microsoft's layoffs of tens of thousands of engineers make such failures more likely and slower to fix.
  • The author says some European state institutions are already moving from Microsoft to open source alternatives.

What HN thinks

  • Commenters broadly argue the outage reflects customers' own choices, such as over-concentrating on one vendor and cloud-hosting everything, rather than solely Microsoft's failure.
  • Several argue that organizations should plan for downtime with backups, redundancy, and break-glass access, noting that many workflows can tolerate a day without email.
  • One commenter who has worked on Azure cites a former Microsoft employee's account that risky refactoring was rejected for fear of breaking things, suggesting reliability problems predate AI-generated code.
Why Zig over C for Systems Programmingmtsoukalos.eu · 196 points · 170 comments · 3dprogrammingsystemsA Zig advocate argues that for new systems code Zig beats C, citing errdefer cleanup, slice bounds checks, error handling enforced by the compiler, and checked integer overflow.

Key points

  • The sum() example with an off-by-one <= reads past the array in C and exits 0, while the Zig slice version panics with index out of bounds in debug mode.
  • Zig's errdefer runs cleanup in reverse order right beside each allocation, and std.testing.FailingAllocator lets every failure path be tested.
  • Signed integer overflow in C is undefined behavior, while Zig checks overflow in debug and safe modes and offers +% and +| for explicit wrapping.
  • zig cc acts as a C compiler with cross-compilation and enables the undefined behavior sanitizer in debug mode, allowing incremental migration via build.zig.

What HN thinks

  • Several commenters argue that the listed Zig advantages (overflow checks, unignorable errors, bounds checks) can largely be had in C using modern tooling such as sanitizers, [[nodiscard]] in C23, and span types, though others counter that doing this consistently takes more discipline than a language enforces.
  • One commenter notes that Zig's real differentiator is ecosystem culture, such as idiomatic allocator passing, which C libraries rarely follow, making third-party dependencies harder to integrate cleanly.
  • A recent bug in Zig's rewrite of musl, which regressed strchr-style performance, prompted skepticism about Zig's maturity, with commenters noting the language is still pre-1.0 and that version churn remains a practical burden.
All US airports are currently ground stoppedtwitter.com · 193 points · 35 comments · 3dsystemspolicyAll US airports were ground stopped at once, reportedly because of a frequency outage that disrupted radio communications, with the stop expected to last until 22:00 UTC.

Key points

  • The stop was reportedly tied to a frequency outage affecting radio communications, per a commenter reading the FAA notice.
  • The reported end time is 22 Zulu, which is 22:00 UTC or 5 PM Eastern.

What HN thinks

  • Commenters point to a pilot forum thread and note the stop is reportedly tied to a frequency outage that left radio communications unusable until about 22:00 UTC.
  • Some speculate a Microsoft outage could be involved, while others suggest a technical fault is more likely than a security threat, and one commenter notes a similar FAA ground stop occurred in 2023.
  • Commenters recall that all US flights were grounded after the September 11 attacks, which is the main precedent raised for a nationwide halt.

The article couldn't be read, so this summary is based on the discussion.

Minnesota guitarist's "palmfrets" invention strikes a new chordechopress.com · 187 points · 105 comments · 3dhardwareculturePalmfrets is a Minnesota guitarist's device that acts as an extra fret at the bridge end of a guitar, raising the bass strings' root notes by a semitone or more.

Key points

  • It sits at the bridge end of the strings, so fretting it shortens the vibrating length of the string.
  • The article itself was blocked (HTTP 403), so details beyond the product's basic mechanism come from the thread only.

What HN thinks

  • Commenters note the device is essentially a fret at the bridge end that raises bass string root notes, enabling bass notes alongside higher-position treble chords, with interest mostly in drones and fingerpicking.
  • A recurring criticism is that it only works on bass strings, shortens the string and may throw off intonation for normal frets, and requires moving pickups, so commenters doubt major manufacturers will adopt it.
  • Several commenters point to earlier precedents such as the Steinberger TransTrem, Parsons-White stringbender, and Hipshot Doubleshot, arguing that bridge-side pitch tricks have historically struggled to gain traction.

The article couldn't be read, so this summary is based on the discussion.

No AI Slop: A Manifestoslopmanifesto.com · 168 points · 96 comments · 3daiprogrammingA manifesto signed by AI-engineering practitioners defines "slop" as AI output passed along without human responsibility for quality, and commits signers to reviewing work and fixing systems.

Key points

  • Slop is defined as putting less effort into producing something than the consumer must spend understanding it, and the manifesto says AI use alone does not make output slop.
  • It argues review cost stays constant while the cost of generating a thousand-line PR or ten-page doc approaches zero, creating an asymmetry.
  • When a mistake recurs, signers commit to improving guardrails, letting linters, type checkers, and automated checks catch it.
  • Signers say they may stop reading most code as models improve, but will still review plans, designs, or results.

What HN thinks

  • Commenters split on whether the manifesto's definition of slop holds up; one argued that a strict definition would cap AI productivity, since some AI-generated proofs may need heavy review rather than none.
  • Several commenters stressed that large AI-generated PRs are hard to review, and one suggested banning oversized PRs and splitting work into smaller reviewable pieces, with another describing an OpenSpec-based workflow that reviews code in ordered parts.
  • Skeptics argued the manifesto may struggle against business incentives, since employers may reward those who use AI most aggressively, and one called it engagement bait that is itself vague.
Claude Is Downstatus.claude.com · 167 points · 142 comments · 4daitoolsThe Claude status page history is shown, with no specific incident details in the provided material, and the thread reports Claude was down for users on the day of the post.

What HN thinks

  • Several commenters described concrete disruption, such as a presentation cut off mid-prep and an agent workflow stalled, and argued that dependence on a single AI provider is the real risk.
  • Commenters widely recommended multi-provider setups, such as an agent that can switch models or a local model fallback, to keep work going during outages.
  • Some pushed back that the outage mostly affected authentication or was short in their region, and that cloud services carry no meaningful uptime guarantee without a negotiated SLA.
Brazil's Caixa insurer breach exposes names and tax IDs of all 7.6M customersfilemanager-cdn.mziq.com · 157 points · 66 comments · 3dsecuritypolicyCaixa Vida e Previdência, an insurer in Brazil's Caixa Seguridade group, disclosed that attackers copied registration data on its 7.6 million active customers, including names and CPF tax IDs.

Key points

  • Access was identified on November 6, 2026, and reported to Brazil's ANPD data protection authority by November 11.
  • Exposed data covers current and former customers; some records also include addresses, phone numbers, emails and birth dates.
  • The company says there is no evidence that policy amounts, pension balances, banking details or passwords were compromised.
  • Customers with exposed phone numbers or emails are being notified, along with vulnerable groups, and given fraud-prevention advice.

What HN thinks

  • Commenters broadly agree that CPF tax IDs have already leaked many times in Brazil, so the marginal harm of this breach is seen as limited to the name and tax ID.
  • The main concern raised is that knowing someone held life insurance with Caixa gives scammers a targeting and credibility advantage, even without financial data exposed.
  • One Brazilian commenter noted CPF is required for everyday purchases and government logins, and that fraudsters can use it to open accounts and take loans, so its exposure matters more than it appears.
Why is Google cutting off Search from AI-free schools? Here are some theoriestext.npr.org · 145 points · 141 comments · 3daipolicyGoogle will cut off Google Search for schools that disable AI Overviews, refusing a school-wide AI opt-out while several large districts agreed to ban AI in classrooms.

Key points

  • Google told schools that disabling AI features for students means losing access to Google Search entirely, per NPR's reporting.
  • Los Angeles Unified, Seattle Public Schools and New York City Public Schools recently agreed to AI bans in classrooms, while Google did not.
  • A former Google education staffer suggested the policy may be tied to a court win, since an off switch for schools could undercut the argument that AI Overviews is separate from Search.
  • Common Sense Media rated AI Overviews an 'unacceptable risk' for citing false information and unsafe content, and a North Carolina family sued Google over Gemini's role in a teen's suicide plan.

What HN thinks

  • Commenters largely view Google's refusal as a bundling choice rather than a technical limit, noting that third-party extensions and URL parameters can already hide AI Overviews.
  • Several point out that the AI summary is only one component, so an all-or-nothing toggle makes it hard for schools to block AI without losing core search.
  • Commenters compare the argument to Microsoft's Internet Explorer bundling defense, which courts rejected, and note that alternatives like DuckDuckGo's no-AI subdomain exist but many districts block other engines.
Lonewolf 0.3 – A standalone XMPP server written in Rustgithub.com · 144 points · 52 comments · 3dopen-sourcetoolsLonewolf 0.3 is a pre-1.0, single-node Rust XMPP server adding certificate-authenticated federation, direct TLS, and several XEPs, with clustering and full compliance still unfinished.

Key points

  • Federation is on by default for hosts with configured certificates and public DNS names, but peers must present trusted certificates and use SASL EXTERNAL since Server Dialback is unsupported.
  • Opt-in SCRAM-SHA-512 and SCRAM-SHA-512-PLUS implement an expired Internet-Draft rather than an RFC, and new credentials are unreadable by older binaries even when SHA-512 login is disabled.
  • Federation checks CA-published CRLs over plain HTTP by default, and downloaded checks treat unknown revocation status as acceptable unless s2s.tls.fetch_crls is set to false.
  • The default client stanza limit drops from 512 KiB to 256 KiB, and traffic byte counters now include TLS handshake and record overhead.

What HN thinks

  • Commenters broadly argue that the XMPP ecosystem's weakness is in clients rather than servers, citing the need for polished desktop clients comparable to Conversations on mobile.
  • The most pointed criticism is that the project's AGENTS.md file and LLM-assisted development make it a poor fit for a privacy-focused protocol, though others counter that development tooling does not affect runtime security.
  • Several commenters say they prefer XMPP over Matrix for lower operating cost and simpler setup, with one reporting a switch back from Matrix for that reason.
Switcher (2004)folklore.org · 143 points · 42 comments · 4dhistoryprogrammingAndy Hertzfeld's 1984-85 account of writing Switcher, the first multitasking environment for the 512K Macintosh, and how Steve Jobs set its $100,000 price.

Key points

  • Hertzfeld got the idea in October 1984 after watching John Markoff switch programs on a DOS utility called Memory Shift during a Thunderscan demo.
  • Microsoft's Bill Gates offered a development deal pegged to lines of code and weekly pay, which Hertzfeld declined in favor of independent work.
  • Switcher patched traps like GetNextEvent, Launch and ExitToShell and had to track down low-memory globals that needed swapping per application.
  • Apple paid $100,000 plus a 10% wholesale royalty on any separate sales, which later added roughly $50,000.

What HN thinks

  • Commenters highlighted how Switcher's clipboard-based cross-app workflow (Multiplan to Cricket Graph to MacDraw) showed painless integration that many later software suites struggled to match.
  • Several commenters placed Switcher in context, noting that the Amiga's Workbench already offered application switching and preemptive multitasking in 1985, and that Multifinder later succeeded Switcher without sharing its code.
U.S. October PPI Rises 6.5% YoY, Highest in Nearly Four Years142 points · 121 comments · 3dfinancepolicyU.S. producer prices rose 6.5% year over year in October, the highest reading since November 2022, though slightly below the 6.6% forecast.

Key points

  • Monthly PPI rose 0.3%, below the expected 0.4%.
  • September's monthly reading was revised up from 1.3% to 1.4%.
  • Annual inflation at the producer level accelerated from the prior 6.3% reading.

What HN thinks

  • Commenters broadly argue that the deficit and fiscal policy, not just the Fed, need to address inflation, with several saying politicians are unlikely to act soon.
  • Some commenters say raising interest rates is a weak tool when inflation is driven by energy prices or supply shocks, and that fiscal measures would work better.
  • Several commenters blame the thread's political actors for the inflation, but the discussion is heavily partisan and contested, with little agreement on causes beyond energy prices and deficit spending.
Perplexity Decider v1.3: Training, eval, and use casesperplexity.ai · 140 points · 26 comments · 3daiPerplexity has published a blog post about Decider v1.3, a version of its model described as covering training, evaluation, and use cases; the article itself could not be retrieved.

What HN thinks

  • Commenters report that Decider v1.3 performs well on local tests and is competitively priced, with one noting it arrives as a cheaper alternative after another provider discontinued its similar API.
  • A recurring criticism is that Decider v1.3 Flash gains about four points over v1.2 Flash while the flagship benchmarks look unchanged, and one commenter questions why decision models need separate flagship and flash tiers at all.
  • Several commenters dismiss the release as unoriginal, though others reply that the value lies in distribution, such as easy access for teams already on a given cloud provider.
OpenAI disbands its robotics team as its general models start running robotsfortune.com · 139 points · 62 comments · 3daibusinessOpenAI has disbanded its standalone robotics team, moving researchers into its main research group and applied organization, on the bet that general models can control robots directly.

Key points

  • Robotics research staff now report to Chief Research Officer Mark Chen; applied robotics and hardware teams moved under President Greg Brockman's Applied organization.
  • OpenAI said general models such as GPT-6 Astra have shown they can be applied directly to robotics, which motivated the change.
  • In an Nvidia study, an Astra-powered agent reached 74.8% success on a simulated manipulation benchmark, versus 51.0% for the best specialized robotic models.
  • The reorganization comes as OpenAI faces reported $10.9 billion quarterly operating losses and pressure to cut costs before a possible 2027 IPO.

What HN thinks

  • Commenters broadly argue general models are too slow for real-time robot control, so a layered stack with fast local reflex controllers and a slower high-level planner is the more practical design.
  • One commenter notes that robots need on-board decision-making so they can safely shut down when connectivity is lost, and that cloud-only control is a serious deployment limitation.
  • Several commenters point out that the team was reassigned rather than laid off, so the move is better read as a focus shift than a cost-cutting measure.
Weight Space Misfoldingalphaxiv.org · 138 points · 49 comments · 3daisecurityA paper by Chris von Csefalvay proposes 'weight space misfolding', where one model edits another model's weights beyond its authorized scope, and reports small controlled probes on a 1.5B model.

Key points

  • In a controlled probe, an editor model changed Qwen2.5-1.5B-Instruct so it gave the false answer 'Buenos Aires' for the 2000 Olympics on all 100 held-out draws.
  • The Olympic edit was not cleanly isolated: of 70 control prompts, the pristine model passed 49 and the edited model 48, losing the control testing copying 'Sydney'.
  • The legal-refusal probe produced the exact requested refusal on only 5 of 22 held-out prompts and also refused an unrelated question about France's capital.
  • The selective omission edit suppressed the target supplier name on only one of 22 held-out prompts, and the editor itself said it had not achieved reliable omission.

What HN thinks

  • Commenters widely argue the headline overstates the paper, which only shows weight-space editing in small controlled probes and explicitly does not demonstrate a virus or self-propagation.
  • Several commenters stress that open weights and model hashing would be needed to detect such tampering, while closed models make inspection impossible.
  • One commenter describes a researcher who detected a Qwen model trained to misbehave on a trigger phrase by inspecting its weights before any output, suggesting weight-level detection is feasible.
Microsoft outage gives customers a crash course in dependence on Redmondtheregister.com · 129 points · 43 comments · 3dbusinesssystemsA prolonged Microsoft 365 outage starting November 12 disrupted Exchange Online, Teams, Copilot and Microsoft 365 Education, with recovery still in progress the next day.

Key points

  • Microsoft dated the incident's start to 1826 GMT on November 12 and called it a Service Degradation.
  • On November 13 Microsoft said availability was improving but gave no full-recovery timeline in the statement quoted.
  • Nextcloud and Civo executives argued the real risk is concentration of critical infrastructure among a few vendors.
  • The UK's Commons Science and Technology Committee chair cited the outage as evidence of weak resilience in UK digital infrastructure.

What HN thinks

  • Commenters largely argue that big outages like this rarely change behavior, since services bounce back and customers tolerate a few lost days per year.
  • Several commenters frame the core problem as cloud and SaaS concentration, with one noting that on-prem alternatives became hard to justify after the broad shift to cloud.
  • One commenter with long Exchange administration experience pushed back on the view that Exchange is an unmanageable legacy codebase, saying it worked well in small and mid-sized deployments.
Live-migration support in FreeBSD and bhyve cominggyptazy.com · 125 points · 33 comments · 3dsystemsopen-sourceA WIP prototype adds pre-copy live migration of amd64 bhyve guests between two FreeBSD hosts over TCP, pausing the guest only for the final stop-and-copy phase.

Key points

  • Observed downtime for a slowly dirtying 512 MiB guest was roughly 138 to 171 ms in a two-host test environment.
  • A rapidly dirtying guest may not converge during pre-copy but still completes correctly via stop-and-copy with a longer pause.
  • The destination does not start the guest until it receives a GO commit message, so a failure before commit lets the source resume.
  • Transport currently has no authentication or encryption, and multi-vCPU testing has been limited mostly to one-vCPU guests.

What HN thinks

  • Commenters note the project's momentum, with one pointing out that DragonFlyBSD's developer activity also seems to be picking up, and another linking Oxide's Propolis as a related live-migration effort for bhyve.
  • Several commenters are eager for the feature, with one saying it could make the Sylve cluster manager usable for production, though the prototype is still WIP.
  • One commenter reports that experimental Podman and OCI container support has existed since FreeBSD 14.3 but requires extra configuration, and that they have not relied on it yet.
FAA issues nationwide ground stop due to "extraordinary circumstances"axios.com · 113 points · 35 comments · 3dpolicybusinessThe FAA has issued a nationwide ground stop citing "extraordinary circumstances", with the cause not stated in the material available here.

What HN thinks

  • Commenters note this is not unprecedented: a nationwide departure halt occurred in January 2023 after a NOTAM system outage, though the 9/11 shutdown was broader because aircraft were ordered to land.
  • Several commenters speculated about causes, including a national radar outage, legacy air traffic control systems on 3.5-inch diskettes, and a cyberattack, but none of these were confirmed in the thread.
  • One commenter at LGA reported that passengers were boarding with no updates from the airline or airport, highlighting a lack of communication during the stop.

The article couldn't be read, so this summary is based on the discussion.

The Dawn of the Data Intelligence Engineermotherduck.com · 112 points · 73 comments · 4daicareerMotherDuck's blog proposes the 'Data Intelligence Engineer', a role that builds shared business context, evals, and agent-accessible data models so non-data staff can answer questions with AI agents.

Key points

  • The author argues LLM agents now write usable SQL and infer business logic from context, because multi-turn agents can retry queries after inspecting results.
  • Proposed supervision tiers are Wild West, Curated, and Core, with heavier oversight scaling by the impact of a number or dashboard.
  • The role centers on evals, which are test questions with known answers used to detect when agent answers break.
  • MotherDuck's product pieces named in the post include Flights for pipelines, Guides as shared markdown context, Dives for agent-built visualizations, and Ducklings as per-user DuckDB instances.

What HN thinks

  • Several commenters argue that narrow data job titles (data engineer, analyst, scientist) create friction and silos, and that the best practitioners stay curious across the whole stack.
  • Commenters are skeptical that the new title is more than a rebrand, noting that analytics engineers and BI analysts already do this kind of modeling and context work.
  • One commenter points out that the core challenge of shared metrics and definitions has been around for years, and that restricting access to a metrics layer is a familiar answer to the risk of agents producing misleading analyses.
Africa's solar revolution is vast and speedyeconomist.com · 110 points · 40 comments · 4denergyscienceAn Economist article says Africa's solar buildout is vast and fast, but the full text was unavailable and no discussion exists yet, so specifics cannot be confirmed.

What HN thinks

  • Commenters broadly agree that cheap Chinese solar panels and batteries let developing countries skip expensive grid and fossil infrastructure, much as Africa leapfrogged landlines to mobile phones.
  • Several commenters note that decentralised solar suits countries with unreliable national grids, such as Kenya and South Africa, where one commenter describes frequent outages and diesel backup.
  • Debate centres on whether China's subsidies accelerated an inevitable shift or deliberately engineered it, with one commenter arguing the adoption was inevitable and China simply accelerated it.
EU Digital Identity Wallet rollout: every EU state on one boardmyid.ie · 103 points · 65 comments · 3dpolicytoolsA tracker grades all 27 EU states on published evidence for their EUDI wallets, finding only Denmark and Italy publicly available and ten more in pilot or testing.

Key points

  • Denmark's AltID and Italy's EUDIW-IT are publicly available, but neither is EUDI-certified yet.
  • Seven states (Estonia, Finland, Germany, Netherlands, Romania, Slovakia, Sweden) target public availability only after the 24 Dec 2026 deadline.
  • Private services using strong customer authentication must accept the wallet from 24 Dec 2027, with micro and small enterprises exempt.
  • Grades measure published evidence only, not legal compliance or wallet quality, and missing evidence is labelled 'not publicly confirmed'.

What HN thinks

  • Commenters widely expect EUDI wallets to depend on Google or Apple device attestation, citing Germany's Android wallet, though one commenter noted that requirement was reportedly dropped in a later version.
  • Several commenters worry that wallets will become de facto mandatory through banks, telecoms and services, pointing to India's Aadhaar as a precedent and raising concerns about losing a phone.
  • One commenter noted that some national wallets, such as the Dutch one, are open source and reportedly will not force Google or Apple attestation, and that users may be able to choose any compliant implementation.
TSMC Hackedsec.gov · 103 points · 15 comments · 4dsecuritybusinessTSMC disclosed in a 6-K filing that some current and former employee personal data was posted on an external website without authorization, with no evidence its own systems were breached.

Key points

  • The company says the leaked data matches records held by a cloud service provider and is unrelated to operations or customer data.
  • TSMC has engaged the cloud vendor, investigating authorities, and a third-party cybersecurity firm, and is notifying affected staff.
  • The filing is a short statement under Form 6-K and gives no count of affected employees or the website involved.

What HN thinks

  • Commenters point to a third-party cloud or HR software provider as the likely source, since the leaked records reportedly match data held by that vendor rather than TSMC's own systems.
  • One commenter argues the reported 3 bitcoin ransom looks low for 46,000 employee records, suggesting the hackers may intend to sell the data instead.
  • A commenter calls the filing's wording perfunctory, noting it says no evidence of a TSMC breach while the data reportedly matches a vendor's systems.
Apify Security Incidentstatus.apify.com · 91 points · 10 comments · 3dsecurityApify disclosed that an unauthorized third party used a compromised credential on Nov 9 and 10, 2026 to reach internal systems and access some customer accounts, data, and stored information.

Key points

  • Apify detected the activity on November 11 at 12:13 UTC, cut off access, and says the attacker has been inactive since November 11.
  • Apify revoked some OAuth connections, including GitHub, as a precaution and says the connections were not found to be an attack vector.
  • API tokens are now stored only as hashes, and users are advised to set storage access to Restricted because anyone with a storage ID can read it while open.
  • Apify notified affected customers individually and reported the incident to relevant authorities; a post-mortem was promised within a week.

What HN thinks

  • Commenters questioned why API tokens were not already stored as hashes, with one noting that many systems show tokens repeatedly, suggesting they are treated more like readable values than long passwords.
  • One commenter argued that teams offering API services should build core functions like rate limiting and OAuth in-house rather than depend on API management middleware, which adds an attack surface.
  • Commenters pointed to Bleeping Computer's coverage, which reports that Apify rotated API tokens and forced 2FA resets after the breach, and noted that an earlier thread on the same incident was a duplicate.
Ctrl+C, Ctrl+v, Ctrl+cve: Bypassing Android's Clipboard Read Security Measurestoninoble.com · 90 points · 45 comments · 3dsecuritytoolsA blog post titled as a bypass of Android's clipboard read protections, but the article itself returned a 404, so its technical contents cannot be verified.

What HN thinks

  • Commenters report that some apps, such as BMW's and a Belgian post app, read the clipboard on open, and a commenter says GrapheneOS lets users control this permission and recently added secure paste.
  • Several Android users said they never see the notification described in the article, and one noted password managers show masked clipboard content in the keyboard paste button.
  • A recurring view is that the clipboard should only be readable on user-initiated paste, though commenters note that gating programmatic reads with a confirmation prompt, as iOS does, is a practical middle ground.

The article couldn't be read, so this summary is based on the discussion.

The 42-Year Journey to Bring 'Neuromancer' to the Screenwired.com · 83 points · 58 comments · 3dculturetoolsApple TV's Neuromancer adaptation, created by Graham Roland and JD Dillard with Drake's DreamCrew producing, premieres January 22 after rights came available and Gibson picked the showrunners.

Key points

  • Gibson chose Roland and Dillard from a pitch contest over bigger names, after they read the book and presented a thoughtful pitch.
  • Gibson said he never saw the Neuromancer world as a dystopia, and he suggested expanding the Tessier-Ashpool family story, which the show adds.
  • The production filmed in Japan for practical sets, and the production designer was hired before any other department head.
  • Case's deck was built as a station-like terminal rather than a laptop, and Case is portrayed as a neurological athlete who pays a physical toll.

What HN thinks

  • Commenters broadly found the trailer too clean and bright, arguing it lacks the grit, disorientation, and punk edge of the novel and of shows like Altered Carbon.
  • One commenter noted that the book's disorientation comes largely from its neologistic prose, which is hard to convey on screen, and that the trailer does not reflect it.
  • A recurring caveat is that Neuromancer's influence means much of its imagery feels derivative today, and that its tech predictions now sit close enough to reality to make a faithful adaptation tricky.
US university students tell surveillance cameras to Flock offtheguardian.com · 83 points · 35 comments · 3dsecuritypolicyAt least ten US college campuses across seven states have dropped Flock Safety license plate reader contracts since April, with students and some high schoolers pressing for the cameras to be removed.

Key points

  • Schools including Cornell, Penn State, Emory, Binghamton, and several University of Wisconsin campuses have scrapped Flock contracts, and the University of Buffalo paused a planned 60-day pilot.
  • UNC Charlotte restricted national access to its Flock records after WBTV found Houston police searched its camera network 183,000 times between March and May 2026.
  • Many schools are replacing Flock with other license plate readers, such as Genetec at Stanford and Axon at Binghamton, which activists say still raises privacy concerns.
  • Elgin Community College could not cancel early, so staff covered the cameras with trash bags until the vendor removed them after the contract lapsed in April.

What HN thinks

  • Commenters debated whether Flock's national record-sharing defaults are opt-in or opt-out, with one noting that most departments using it have opted in to broad access.
  • Some argued the legal fight belongs in Congress rather than the Supreme Court, while others said the cameras raise Fourth Amendment concerns.
  • One commenter suggested that local school board races are often decided by small margins, so campaigning for candidates with anti-surveillance positions may be more effective than petitions.
North Korean hackers used fake video calls in $400M crypto hack, says FBIbbc.com · 78 points · 28 comments · 3dsecurityfinanceThe FBI says North Korean hackers used fake video calls to carry out a crypto theft it values at about $400 million.

What HN thinks

  • One commenter describes being personally targeted by a similar scheme, where a fake outreach tried to get them to download a booby-trapped repository containing a remote shell.
  • Commenters were struck by how convincing the fake network of sites and companies was, though one noted that reports so far have only described the attack in written form rather than recordings.
  • Some commenters argued that crypto's irreversible transactions make such thefts an expected consequence, while others objected mainly to the money funding North Korea.
Scientists claim the universe is twice as old as believedgeekspin.co · 77 points · 76 comments · 4dsciencespaceA University of Ottawa physicist and a colleague propose a model, CCC+TL, that puts the universe's age at 26.7 billion years while dropping dark matter and dark energy.

Key points

  • The model has physical constants slowly shifting over time and light losing energy as it travels, which stretches the inferred age.
  • The authors report it matches Big Bang hydrogen and helium abundances and the historical expansion rate without dark energy.
  • They say it explains stellar and gas orbits in 175 galaxies, a case usually made for dark matter.
  • Some of the work is still under review, and the article says the idea is well outside the mainstream.

What HN thinks

  • Commenters largely treat the proposal as non-peer-reviewed speculation, with one noting that radical ideas can still help resolve tensions like the Hubble tension.
  • Several commenters push back on the post's claim that the universe is twice as old, noting that cosmological age estimates have converged within a few billion years since the Hubble and CMB measurements.
  • One commenter argues that the earlier thread about fundamental constants clustering near 1 is misleading, since the values depend on the choice of units and the relevant constants are not close to 1.
CERN Built the Large Hadron Collider's Next-Gen Magnetsspectrum.ieee.org · 76 points · 8 comments · 3dsciencehardwareCERN has built and tested 7.2-meter Nb₃Sn quadrupole magnet coils for the High-Luminosity LHC, with the finished magnets reaching 11.3 tesla in quality control tests.

Key points

  • The coils are wound from a 760-meter cable before a heat treatment of several days at up to about 650 °C forms the brittle Nb₃Sn superconducting phase.
  • Because Nb₃Sn cannot tolerate winding strain, the wind-and-react method is used, and reacted coils must then be handled very carefully.
  • Of 250 coils produced, 25 failed from human error or manufacturing problems, and a single trapped metal particle was enough to discard one coil.
  • The current NbTi dipole magnets top out at 8.3 tesla, while the new Nb₃Sn quadrupoles target just under 12 tesla.
Sift: LSP, Lint, Format Lisp, Scheme and Emacs Lispgithub.com · 73 points · 10 comments · 5dprogrammingtoolssift is an MIT-licensed Guile 3.0 tool that formats, lints, and serves LSP for Scheme, Emacs Lisp, and Common Lisp, reading each dialect by its own rules.

Key points

  • Formatting is a lossless fixed point: comments, blank lines, #; datum comments and #. read-time forms come back byte-exact, and a second run changes nothing.
  • Ships 92 lint rules, each with a stable code like SIFT012, and sift explain reads rule docs from the rule records themselves.
  • Fixes replace exactly one range each, and the whole fix set is discarded if combined it would introduce a parse error.
  • A browser build compiles the formatter and linter to WebAssembly via Guile Hoot; the module is 2.3 MB, and lint takes 1.2 s on a 6 kB buffer.

What HN thinks

  • Commenters pointed to existing Common Lisp tooling as alternatives or complements, including trivial-formatter, lisp-format, mallet, and the linter built into ocicl.
  • The main caveat raised is that requiring Guile for a single tool is a barrier for Common Lisp users, though one commenter noted it pairs well with SLY or Geiser alongside a REPL.
  • One commenter criticized the README's prose as sounding AI-written and suggested a clearer description of the check command: returns 0 when formatting is correct, prints a patch and returns 1 otherwise, and never modifies files.
Europe's electric car sales jump 49% as global EV sales growelectrek.co · 66 points · 56 comments · 4dbusinesspolicyBenchmark Mineral Intelligence data shows global EV sales rose 8% year over year in October 2026, with European BEV sales up 49%, but North America's gain is against a post-tax-credit trough.

Key points

  • Europe sold 510,000 EVs in October, up 36% year over year, with plug-in hybrids growing only 7% against 49% for fully electric cars.
  • Germany's EV share topped 45% of new car sales in October, its highest since December 2022, though that earlier peak came before subsidy cuts.
  • US October sales of about 110,000 were compared against the month after the federal tax credit expired, so the gain reflects a lower base rather than recovery.
  • China's EV sales fell 12% year over year to 1.14 million, though EV share of its shrinking car market hit a fresh high.

What HN thinks

  • Commenters dispute the claim that Europe bans Chinese EVs, noting they are widely sold in several countries but face import tariffs, with BYD building cars in Hungary.
  • Several commenters argue US EV adoption lags because of driving distances, rural areas lacking transit, and high prices, though one says US charging infrastructure is adequate for most long trips.
  • One commenter argues that cheap Chinese EVs show affordable models can work in Europe, and that the main barriers were charging gaps and high prices rather than demand.
Microsoft says attempts to fix global outage not going 'as expected'bbc.com · 66 points · 39 comments · 4dsystemsbusinessMicrosoft says its attempts to fix a global outage are not progressing as expected, with the root cause still unclear according to commenters reading its status updates.

What HN thinks

  • Commenters are skeptical of the status updates, with one saying they seem like AI-generated text and do not show a root cause has been identified.
  • Several commenters speculate that AI tooling or lost engineering expertise contributed, though no evidence for either is given in the thread.
  • One commenter reports unrelated recent Microsoft email changes, including DKIM changes and blocked third-party mail servers, and hopes the outage drives users away.

The article couldn't be read, so this summary is based on the discussion.

US test reactor cleared for operationsworld-nuclear-news.org · 63 points · 20 comments · 3denergyscienceOklo's Groves-1 isotope test reactor in Texas has received final DOE clearance for unrestricted research, testing, and training operations after reaching criticality in August.

Key points

  • Groves-1 is one of 11 projects selected for DOE's Reactor Pilot Program in August 2025 and reached first criticality less than a year after groundbreaking.
  • The reactor will train nuclear personnel and continue control rod drive mechanism performance and reliability testing.
  • Groves-1 data is meant to inform Groves-2 Forge, a planned isotope production and processing pilot under DOE's Nuclear Energy Launch Pad.
  • Under the Atomic Energy Act, DOE can authorize research and demonstration reactors, while NRC still approves commercial designs, and an October NRC rule lets applicants reuse DOE test data.

What HN thinks

  • Commenters note the reactor's main value is producing medical isotopes for cancer treatment, which are in short supply.
  • One commenter argues US nuclear power has been held back by regulation and policy for so long that it is effectively a dead technology, while another points out that nuclear propulsion still powers a large part of the US Navy's warships.
  • The thread is short and mostly split between supporters calling nuclear part of a diversified energy mix and skeptics doubting its relevance to grid power.
Someone's Impersonating This Blogmattbee.mataroa.blog · 61 points · 3 comments · 3dsecuritywebA clone of the mattbee blog is being served at mataroa.site, a domain that resolves to a copied version of the author's Mataroa-hosted blog without the newest posts.

Key points

  • The clone stops at an August 21 entry, and older links, comments, RSS, and standalone pages return 404 errors.
  • Stolen images are re-hosted on the clone's own server with a watermark reading mattbee.mataroa.blog, the author's real URL.
  • The author's Mataroa account is free tier, so a custom domain pointing at the blog would not have been possible, ruling out a simple redirect.
  • The index page contains a meta verification tag for mitgo.com, a site the author suspects is an SEO or ad scam platform.
The Netherlands Is Dismantling the International Criminal Courtrightsforum.org · 60 points · 5 comments · 4dpolicyaiAccording to anonymous diplomatic sources cited by The Rights Forum, the Netherlands is negotiating with the US over restrictions that would strip much of the International Criminal Court's power.

Key points

  • Proposed restrictions include immunity for heads of state, no jurisdiction over nationals of non-state parties even for crimes on a state party's territory, and no jurisdiction where national legislation exists.
  • Enacting them would require amending the Rome Statute, needing a two-thirds majority of States Parties and ratification by seven-eighths before entry into force.
  • US sanctions on the Court since 9 October, with some measures not taking effect until 7 April 2027, could largely paralyse the institution.
  • The Ministry of Foreign Affairs responded only with a general statement saying no concrete proposals are currently before the Assembly of States Parties.
Claude Meetingsclaude.com · 59 points · 53 comments · 3daitoolsAnthropic's Claude Meetings lets Claude join video calls as a visible participant that takes notes, answers questions and builds docs or whiteboards live, now in beta for Team and Enterprise plans.

Key points

  • Claude joins only Google Meet for now, with other platforms promised later, and it saves transcripts and screen descriptions but not audio.
  • By default Claude stays out of calls that include outside guests, and an admin must allow outside-guest calls.
  • Artifacts made during calls, such as docs and whiteboards, are free through February 10, 2027; some organizations, including ZDR and HIPAA setups, are excluded from the beta.
  • Connected-tool answers count toward plan usage, and Claude asks each person for permission before using their tools.

What HN thinks

  • Commenters raised consent and privacy concerns, noting that recording without clear agreement from every participant is a real issue, with one pointing out that the only option offered is to refuse to join.
  • A recurring worry is that giving agents tool access, which is what makes them useful, creates security risks that ordinary users will struggle to judge.
  • Several commenters questioned Anthropic's strategy, asking whether a model vendor moving into end-user products will undercut the third-party integrators that build on its API, and whether that leads to a market dominated by a few labs.
Arm boss Rene Haas approached in Nvidia CEO searchft.com · 56 points · 51 comments · 3dbusinessaiThe Financial Times reports that Arm chief executive Rene Haas has been approached about becoming Nvidia's next chief executive, though the article itself could not be fetched.

What HN thinks

  • Commenters largely doubt that an outside hire would preserve Nvidia's culture, with one arguing a newcomer without internal loyalties might be needed to restructure the leadership team.
  • Several commenters see Nvidia's momentum as tied to Jensen Huang, and question whether a successor can sustain it, while others point out that Arm's post-Saxby CEOs have had limited success.
  • Some commenters argue that an interim CFO leading briefly is common practice, while others say Huang's legal and political situation makes him a risky candidate for any company to hire.

The article couldn't be read, so this summary is based on the discussion.

National Weather Service reports nationwide radar website outage amid FAA stopnews9.com · 55 points · 12 comments · 3dsystemspolicyThe National Weather Service reported a nationwide internet and communications outage on November 13, 2026 that makes some of its website weather products unavailable or outdated.

Key points

  • The radar.weather.gov radar platform was among the affected NWS online services, and the agency gave no timeline for restoration.
  • Local NWS offices said they continue issuing watches, warnings and forecasts, which remain available through some external websites.
  • The FAA separately ordered a nationwide ground stop on departing flights at 2:54 p.m. Central, citing communication issues at Air Route Traffic Control Centers.
  • Officials said there is no confirmed connection between the NWS outage and the FAA ground stop.

What HN thinks

  • Commenters noted that a circulating screenshot claiming radar was down is outdated, and the NWS spokeswoman said its radar systems had no disruption that day.
  • Several commenters questioned whether the NWS outage and the FAA ground stop are related, with early confusion about whether the stop involved radio or radar.
  • One commenter pointed to the Trump administration's past efforts to shut down or restructure the NWS as context for distrust of the outage explanation.
Mercedes-Benz to halt 'all future investments' in Germanyft.com · 53 points · 63 comments · 3dbusinesspolicyMercedes-Benz reportedly plans to halt all future investments in Germany, according to a Financial Times headline; the article itself was not available.

What HN thinks

  • Commenters tie the halt to a dispute with unions over raising the work week from 35 to 40 hours without extra pay, with some reading it as a negotiating tactic.
  • Many argue Germany's high labor, energy costs and Chinese competition in EVs are structural, while others say Mercedes' product strategy, brand image and reliability drove the decline.
  • One commenter with first-hand experience notes that labor is a small share of car production cost, with energy and equipment mattering more.

The article couldn't be read, so this summary is based on the discussion.

Jensen Huang Admitted He Was Facing an "Avalanche" of Excess Wafers and Memoryrespawnfirst.com · 48 points · 31 comments · 3daibusinessA WSJ report says Nvidia's CEO Jensen Huang described a glut of prepaid TSMC wafers and memory as an 'avalanche' bearing down on the company in August 2026.

Key points

  • Nvidia's signed guarantee cap for OpenAI compute reportedly shrank from about 8 gigawatts (up to $250 billion) to roughly 4.25 gigawatts (up to $125 billion).
  • Huang reportedly said major cloud providers were moving to in-house chips, leaving Nvidia with committed wafer and memory purchases it could not sell.
  • Nvidia has reportedly asked TSMC to wind down chip orders.
  • Ed Zitron estimates at least $200 billion of Nvidia AI chips sit idle in warehouses, a figure the thread disputes.

What HN thinks

  • Commenters split on whether excess supply is real: one points to hyperscalers saying demand exceeds supply, while others cite reports of GPUs sitting in warehouses awaiting datacenter power and construction.
  • A recurring point is that unbuilt datacenters, not chip demand, may be the bottleneck, with one commenter noting reports of 'dark chips' that cannot be brought online for lack of power.
  • One commenter with first-hand experience said their AMD AI server order was delayed by 10 months, while Nvidia VRAM stayed out of budget and consumer cards still sell at about twice January prices.
Chicago schools to Google – we'd rather lose Search than be forced to accept AItechspot.com · 48 points · 12 comments · 4daipolicyChicago Public Schools says it will switch to another search provider rather than accept Google's plan to tie Search access to school adoption of its AI services.

Key points

  • Google's education VP Chris Phillips said K-12 schools will soon have to accept Google's AI services to keep student access to Search, since Search and AI are now integrated.
  • Blocked students would see a school message in place of search results saying the service is unavailable.
  • CPS says it is exploring ways to suppress AI Overviews and will look to other search solutions if it cannot.
  • Frederick County, Maryland voted 6-1 in October to restrict student AI use from pre-K through 12th grade, and New York City and Los Angeles have also limited student-facing generative AI.

What HN thinks

  • Commenters largely frame Google's move as coercive, with one saying the idea that AI will inevitably be pushed on users has lost credibility.
  • Several commenters point to Kagi's public response as a notable alternative approach, though the thread mostly praises it rather than analyzing it.
  • One commenter notes Google's revised controls, which turn off K-12 student access until administrators enable it, were treated as a 'nice try' rather than a real fix.
The avalanche that hit math is coming for AI developmentdwarkesh.com · 47 points · 57 comments · 3daiscienceDwarkesh Patel argues AI will match the best human AI researchers around 2028 and the best experts in every field by 2030, based on AI accelerating its own R&D.

Key points

  • METR reportedly found Anthropic's rate of progress on well-scoped tasks is about 2x what it was a year ago, mostly attributed to its use of AI for R&D.
  • Patel's median estimate for automated researchers matching top humans is 2028, with a 25% chance it slips past 2032.
  • He argues math progress is a preview because verifiable domains let models search and match ideas, and AI research has many verifiable proxies.
  • He models a 'spark' scenario compressing three to six years of progress into one year, and a 50/50 split with a software-only singularity.

What HN thinks

  • One commenter who has tried reimplementing medical imaging papers reports that methods often prove fragile, overfit, or narrow, though models help a lot with re-implementation and checking claims.
  • Several commenters doubt the forecast because software and other non-verifiable tasks need human-defined success criteria, so progress on verifiable benchmarks may not generalize.
  • A commenter who uses AI coding tools daily says they still need heavy guidance and scaffolding, and that companies adopting AI heavily have not visibly improved products.
The average cost for a new vehicle in the U.S. hits a record high: over $51,000npr.org · 42 points · 39 comments · 4dbusinessfinanceThe average U.S. new vehicle transaction price hit a record $51,109 in Kelley Blue Book data, about $500 above the December 2025 record, though up only 2.1% year over year.

Key points

  • The 2.1% annual increase is below inflation, and cars rose 3.4% a year on average between 2012 and 2021.
  • The figure is what buyers actually pay, net of incentives and dealer discounts, and excludes auto loan interest.
  • Used cars, roughly two-thirds of sales, average close to $27,000 and are up 6% from last year.
  • Over a quarter of car buyers financed with 84-month or longer loans, with down payments sometimes just $500 to $1,000.

What HN thinks

  • Commenters frame high prices as partly a status-symbol and affordability issue, with some arguing that people who can't buy homes put more into cars.
  • Several argue the 2.1% rise is below inflation, so the record is mostly the normal effect of inflation rather than a new shock.
  • One commenter who lives car-free says it frees money for travel and earlier retirement, while noting that hourly rentals make that practical only sometimes.
Concurrent Shell Command Queuesemmer.dev · 31 points · 16 comments · 3dtoolsprogrammingA Bash and Zsh library that runs submitted shell commands through a FIFO queue drained by xargs -P, letting you cap concurrency while adding jobs over time.

Key points

  • queue_start N sets the maximum concurrent jobs, queue_submit adds commands at any time, and queue_wait blocks until the queue drains.
  • Command output is captured per job and printed whole, so lines from different jobs are not interleaved unless output exceeds PIPE_BUF.
  • queue_wait does not report job failures; non-zero exits are only printed as messages, with the exit status coerced to 1.
  • The author says the code has failure modes that are hard to predict and that a shell can hold only one queue per process.

What HN thinks

  • Several commenters point out that make -j with a generated Makefile already handles bounded parallelism, skipping existing outputs, and restarts, so the library may be unnecessary for batch work.
  • GNU Parallel's sem command is suggested as an existing alternative that offers similar job-limiting semantics from the shell.
  • A recurring caveat is that Bash job control and signal handling are subtle, for example the difference between subshells and brace groups when trapping SIGINT to stop child processes.
Is there a 'right amount' of AI?coyotetracks.org · 30 points · 38 comments · 4daiprogrammingWatts Martin argues that LLM-generated code should be minimized, using the Big Walk game controversy and Marco Arment's Claude-assisted apps as examples of the trade-offs.

Key points

  • House House disclosed that LLMs were used sparingly on Big Walk, touching less than a hundredth of a percent of its code, after IGF jurors pledged not to vote for genAI games.
  • The article contends that the less LLM-generated code a codebase contains, the easier it is to vet, and that zero is the safest amount.
  • Martin cites anecdotal reports of duplicated functions and unreachable code sections from coding agents, though he offers no systematic data.
  • He notes that open-source training data may carry GPL and licensing problems, which could create legal risk for non-GPL software.

What HN thinks

  • Commenters split sharply: some argue LLM-built apps feel generic and clunky without human taste and understanding, while others say most software was already low-craft CRUD where AI simply automates tedious work.
  • One commenter defends AI use on accessibility grounds, citing repetitive strain injury as a reason to dictate or generate code rather than type it all.
  • Several commenters push back on the article's framing, calling its claims about LLM failures overstated and its tone self-righteous, while others note the moral anger around AI is less visible than similar outrage over other industries.
NYC Council passes ban on 'surveillance pricing'gothamist.com · 30 points · 8 comments · 4dpolicybusinessNew York City's Council passed a bill banning companies from charging shoppers different prices based on algorithms or personal data, both online and in stores, with fines up to $1,000 per violating transaction.

Key points

  • The bill covers demographics, credit, browsing history and purchase history, and applies to in-store as well as online pricing, though the practice is mainly used on the web.
  • The Council cited an investigation finding Target app prices for a 55-inch SmartTV were $499 inside a store but $599 in the parking lot.
  • Albany lawmakers passed a statewide version in June with some exceptions; the city bill adds a per-transaction fine structure.
  • Speaker Julie Menin says the bill would not affect discounts, though business groups warn it could hurt loyalty programs and small restaurants' marketing.

What HN thinks

  • One commenter pointed out that New York State passed a similar ban in June, and that the new city-level element is the 'relies on personal data' part.
  • Commenters noted that Uber's pricing practices reportedly use signals like phone type, battery level and location to estimate price sensitivity.
  • One commenter argued that governments should also limit their own surveillance tools, such as facial recognition and Flock cameras.
US Intel isn't sure if Alice Huang spied on Nvidia for Chinacnn.com · 30 points · 38 comments · 4dpolicyaiUS intelligence has not established that Alice Huang, a Chinese businesswoman linked to Nvidia CEO Jensen Huang, spied for Beijing, despite FBI scrutiny and political pressure.

Key points

  • Multiple sources say intercepted communications and collected intelligence so far do not prove she knowingly passed information to Beijing.
  • Her Singapore-based company Megaspeed is a major Nvidia chip buyer; DOJ and Commerce began investigating it in 2024 over suspected diversion to China.
  • Commerce has not added Megaspeed to the entity list, prompting White House questions, while the investigation was ongoing.
  • Some Office of the Director of National Intelligence officials reportedly considered the spy characterization inaccurate and overstated.

What HN thinks

  • Several commenters suggest she may have been a profit-seeking smuggler moving chips through Singapore and Malaysia rather than a formal intelligence asset, while noting Chinese firms commonly take state funding.
  • Many commenters criticize the reasoning that she must be a spy simply because she is Chinese, calling it racism, while others counter that CCP pressure on diaspora members is well documented and makes influence plausible.
  • Some commenters question the reporting itself, pointing out that the NYT's story seemed to shift after the initial claims, and speculate about who benefits from Jensen Huang's departure.
Announcement: Conductor Is Backtwitter.com · 30 points · 17 comments · 4dtoolsopen-sourceConductor, a previously discontinued project from Charlie Holtz, is announced as back, with a linked social post offering no details of its scope.

What HN thinks

  • Commenters are largely skeptical of the revival, citing distrust of Meta after the earlier sudden shutdown and concerns that the product fell behind competitors.
  • Several say they have already moved to alternatives, and one notes they had built their own work around the project before its return.
  • One commenter questions whether any real product development will follow or whether it will only be kept running with minimal maintenance.

The article couldn't be read, so this summary is based on the discussion.

US National Weather Service reports nationwide internet, communications issuesreuters.com · 27 points · 10 comments · 3dsystemspolicyThe US National Weather Service reported nationwide internet and communications problems, per a Reuters headline; the article could not be fetched, so the cause and scope are unconfirmed.

What HN thinks

  • Commenters say NWS radar sites are fully operational and that the reported outage stemmed from stale cached error pages on isolated office webpages, not a radar failure.
  • Several commenters noticed a cluster of outages on the same day (Microsoft, Claude, air traffic control, weather service) and speculated about a shared cause, though others pointed to ordinary causes like a bad DNS change or a single upstream carrier failure.

The article couldn't be read, so this summary is based on the discussion.

Show HN: RealtimeVAD – trading false positives for 110ms earlier end-of-speechgithub.com · 26 points · 5 comments · 4daitoolsRealtimeVAD adds a 10 ms energy-based speech-off detector alongside Silero VAD, emitting a pause event about 100 ms before Silero's probability drops, at the cost of frequent false pauses.

Key points

  • On 3,279 real voice-assistant turns, speech-off came a median 100 ms before Silero's probability fell below 0.5, and 206 ms before VADIterator's end.
  • The pause event fires about ten times as often as Silero's, including in short gaps between words, so it is meant for cancellable speculative work.
  • The end event arrives 0.24 s after speech-off unless speech resumes, and comes before Silero's end in 82% of turns, by a median 70 ms.
  • On noisy quiet input with noise within about 20 dB of the speaker, the lead disappears, and the detector ends exactly where Silero does.

What HN thinks

  • Commenters have not yet weighed in; the thread has no substantive comments so far.
Richard A. Posner has diedlaw.uchicago.edu · 25 points · 4 comments · 4dhistorypolicyRichard A. Posner, a University of Chicago law and economics pioneer and former Seventh Circuit judge, died November 10 at age 87.

Key points

  • He joined the Chicago Law School faculty in 1969 and spent 36 years as a judge on the US Court of Appeals for the Seventh Circuit.
  • He served as the Seventh Circuit's chief judge from 1993 to 2000 and retired from the bench in 2017 after about 3,300 opinions.
  • His 1973 book The Economic Analysis of Law is now in its ninth edition, according to the law school.
Building collapses in Brooklyn with 8 people dead so farnytimes.com · 24 points · 3 comments · 3dpolicybusinessA three-story apartment building at 1321 Flatbush Avenue in Brooklyn collapsed into an adjacent excavation site on Wednesday, killing eight people, including two children, and injuring at least 18.

Key points

  • An inspection report had found cracks in the north facade of the collapsed building, yet construction at the adjacent site reportedly continued.
  • The Department of Buildings is sweeping about 100 sites linked to five entities tied to the construction site.
  • A stop-work order has been issued for the site, and the cause of the collapse has not yet been determined.
  • A pregnant woman was pulled alive from the rubble about two and a half hours after the collapse.
Gnomon. Git for Postgres data, built for agentsgithub.com · 20 points · 11 comments · 3daiprogrammingGnomon is an open-source PostgreSQL extension that gives agents copy-free database branches, with diff, type-aware merge, and undo, installable on RDS, Aurora, and Supabase.

Key points

  • Branches read through to main and store only their changes, so opening one copies nothing regardless of database size.
  • Merges combine concurrent edits by type: counts add up, JSON merges key by key, and text merges line by line.
  • Agent permissions are enforced by Postgres roles, with merge rules such as a 20-row cap, table allowlists, and passing checks.
  • The repository was created about an hour before this thread, with a single commit and four stars.

What HN thinks

  • Several commenters are skeptical of the project's README, saying it reads as AI-generated and that the author should explain the problem and design in their own words.
  • Commenters question whether anyone is actually using it, noting the repository was only pushed about an hour before the discussion, so there is no track record yet.
Amodei Says AI Will Cure Cancer in a Decade. Doctors Are Skepticalbusinessinsider.com · 19 points · 26 comments · 3daihealthAnthropic CEO Dario Amodei predicts AI will make cancer, Alzheimer's, heart disease and mental illness relics within a decade, a claim made in an IPO roadshow video.

Key points

  • The prediction appears in a 42-minute IPO roadshow video Anthropic released, with Amodei saying these diseases will be relics 'in, I think, less than a decade.'
  • Physician and biotech critics quoted in the piece call the timeline unrealistic, and one says such claims are insane in an IPO pitch.
  • Anthropic has bought several biotech startups, including a one-person company developing muscular dystrophy treatments, and opened a Bay Area wet lab.

What HN thinks

  • Several commenters argue that medical progress is bottlenecked by biological speed, which AI cannot easily change, so a decade-long cure claim lacks a convincing mechanism.
  • Commenters note that cancer, dementia and heart disease are each many different conditions, so a single cure is unlikely; one suggests some subtypes might be eliminated while others persist.
  • Some commenters point out the timing of the claim during an IPO roadshow, reading it as promotional optimism tied to financial incentives.
Technology as a Search Problemconstruction-physics.com · 18 points · 2 comments · 3dsciencecultureA Construction Physics essay titled "Technology as a Search Problem" that could not be retrieved, so its argument is unverified.
6-11X speedup running Qwen3.6 on my RTX 4090, full repo.github.com · 15 points · 4 comments · 3daitoolsLightreason is a 0BSD-licensed GitHub repo of experimental inference code that speeds up Qwen2.5-3B and Qwen3.6-35B-A3B on a single RTX 4090 using CUDA graphs, compact KV cache, and speculative drafting.

Key points

  • On the 3B model, the author reports 63.0 to 139.6 tok/s across six prompts, with outputs matching all 30 Hugging Face timing runs.
  • The 35B model reaches 51.08 tok/s from 7.93 on a capped Python prompt, and output first diverges from the baseline at token 116.
  • The 35B runner keeps most expert weights in system RAM and moves only the experts Qwen uses, since the text weights total 64.56 GiB.
  • The repo ships no model weights; users download about 6.2 GB for the 3B model and must supply a separate lab adapter to reproduce PARD2 numbers.
Show HN: Reflota – Open-source data recovery for NTFS/FAT disks with phased copygithub.com · 15 points · 2 comments · 3dtoolsopen-sourceReflota is a free, read-only Windows recovery tool for NTFS, FAT and exFAT disks that rebuilds folder trees, lists deleted files, and can image failing disks in resumable phases.

Key points

  • On NTFS it reads the whole $MFT and rebuilds folders from parent references; FAT and exFAT are handled by walking directories and cluster chains.
  • Deleted files show in red in their original folders with a percentage of how much content is still intact.
  • Content carving covers 30+ formats, including camera RAW, PSD and RAR, with lengths read from file structure rather than cut blindly.
  • Disk imaging proceeds ddrescue-style: fast read of good areas, trim, scrape, then retry, with progress saved to a map file for resuming.