HN digest

Qwen4-Maxqwen.ai · 827 points · 437 comments · 22haiAlibaba's Qwen team has released Qwen4-Max, a new flagship model; the article could not be retrieved, so details come only from a short comment thread.

What HN thinks

  • Commenters broadly agree Qwen4-Max lands well below the frontier: Artificial Analysis gives it 43 versus 46 for GPT-6.1 Sol, and it is roughly comparable to Kimi 3.1 and Qwen 3.8 Max.
  • Several commenters note cost comparisons depend heavily on reasoning effort settings, and the Artificial Analysis index is seen as increasingly unreliable due to changing benchmarks and harness choices.
  • One commenter reports that Qwen 3.8 Max performs well apart from speed, with output around 30 tokens per second, which hurts its ranking in agent arena voting.

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

How to make almost anythingpub.cba.mit.edu · 813 points · 137 comments · 18htoolsdesignMIT's HTMAA 2026 course syllabus lays out a weekly fabrication curriculum running September to December, covering CAD, cutting, electronics, 3D printing, machining, and networking.

Key points

  • Topics include computer-aided design, computer-controlled cutting, embedded programming, 3D scanning and printing, and electronics design and production.
  • Later weeks cover machining, input and output devices, motor control, molding and casting, and mesh networks.
  • Two design reviews are scheduled, one in early October and one in mid-November, with a wildcard week in between.
  • The course meets Wednesdays 1:00-4:00 in E14-633, with weekly Wednesday evening recitations and a final presentation on December 15.

What HN thinks

  • Commenters largely defend the breadth-first approach, arguing that knowing what is possible across many fabrication techniques is valuable even if each topic is only skimmed.
  • Former students and TAs describe the course as very time-intensive, around 18 hours a week of stated workload, with the real learning happening hands-on in the lab rather than in the syllabus text.
  • Some critics say the syllabus page alone is hard to use, with few slides or videos, and that the breadth comes at the cost of depth and of clarity about why each skill is needed.
Many of Amazon’s Black Friday and Cyber Monday “deals” offered no savingsgroundworkcollaborative.org · 799 points · 485 comments · 10hbusinesspolicyGroundwork Collaborative found nearly 1 in 3 Amazon Black Friday and Cyber Monday 'deals' in 2024 and 2025 matched or beat prices from the month before, and nearly 1 in 10 cost more.

Key points

  • Analysis covered more than 350 products on Amazon's own holiday deal pages for 2024 and 2025 sale events.
  • One Black Friday 2025 Hunter rain boot listed at $126.99 had been $79.97 nine days earlier, 59% more than the pre-sale price.
  • Among the worse deals, the median sale price was 19% higher than what shoppers could have paid in the weeks before.
  • Groundwork wants regulators to require retailers to show the lowest price a product sold for in the previous 30 days when advertising a sale.

What HN thinks

  • Commenters widely note that the EU and some other jurisdictions require showing the lowest price from the previous 30 days as the reference price, and say this helps expose fake discounts, though some say retailers raise prices just before the window.
  • Several commenters point to third-party price trackers like CamelCamelCamel and Keepa, and to Amazon's own price history graph, as the practical way to check whether a deal is real.
  • One commenter who runs a price tracker explains that Amazon's API limits and affiliate data restrictions make accurate third-party tracking harder than it looks, so tracker data may miss flash or coupon prices.
Technical issue at JPMorgan Chase triggers online banking delays for US banksapnews.com · 677 points · 538 comments · 5hfinancesystemsDowndetector reports Chase Bank is experiencing an outage, according to a social media post that is the story's only source.

What HN thinks

  • Commenters are skeptical of the Downdetector data, noting that visits to unaffected services like Steam, Visa and Google can register as spikes, so the scope of the outage is unclear.
  • Several users report logging into Chase, Bank of America and other banks without problems, with one noting that partial outages by region or service may explain the mixed reports.
  • A recurring theme is whether the cause is AI-generated code, a shared cloud dependency, or an attack, with commenters noting that the Microsoft 365 and FAA outages earlier in the week make the pattern look notable even without a shared cause.

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

Universal XSS in uBlock Origin from website-driven element pickeriter.ca · 585 points · 164 comments · 22hsecurityopen-sourceA malicious website can use an invisible iframe to drive uBlock Origin's Firefox element picker and add trusted filters that inject JavaScript into any site, fixed since.

Key points

  • The attacker learns the random per-install extension UUID by hooking an XHR made by a default no-fetch-if filter rule, then triggering it with a fetch call.
  • The picker trusts messages from any page, and the attacker abuses the unvalidated opt field with an embedded newline to smuggle in a trusted scriptlet filter.
  • Filters added through the picker are treated as trusted until a recompile, so the attack works immediately by opening the victim page.
  • The one-click requirement can be met by overlaying or making the picker iframe transparent over a page the user clicks on.

What HN thinks

  • Commenters agree the root problem is that extensions must inject their element-picker UI into the website's DOM, which a malicious page can exploit; several suggested moving that UI into browser chrome or a sidebar instead.
  • One commenter with extension-development context noted that extensions lack an overlay API comparable to the browser's top layer, so picker UIs end up in page DOM by necessity.
  • Several users reported their uBlock Origin still ran the vulnerable version because auto-update had not fired, and advised manually checking for updates in the add-ons manager and upgrading to 1.75.4 or later.
What's Next for Nvidia?steveblank.com · 431 points · 552 comments · 24hbusinessaiSteve Blank argues that Nvidia, having lost co-founder Jensen Huang as CEO, risks the decline he says often follows visionary founders' departures, since successors tend to be execution-focused.

Key points

  • Blank credits Huang with funding CUDA for years before it earned revenue, then steering Nvidia from graphics boards to a software platform and later to integrated AI datacenter racks after the 2020 Mellanox acquisition.
  • Huang reportedly had about 60 direct reports, discouraged one-on-one meetings, kept no long-term plans, and did not name a successor.
  • Blank names Google, Amazon, Microsoft, Meta, OpenAI and Tesla as customers building their own AI chips, which he calls the clearest threat to Nvidia.

What HN thinks

  • Commenters largely dispute the article's premise, arguing that successful companies like Apple, Microsoft and Google have thrived under non-founder CEOs, and that Nvidia's risks are more about its current situation than leadership style.
  • Several commenters argue the bigger threats are customers building their own chips and the AI data center spending cycle, with one noting Nvidia's role in circular financing could make it a source of instability if it pulls back.
  • Much of the thread shifts to the firing itself, with many commenters focused on the reported affair, spy allegations and export-control smuggling rather than innovation, and some calling the spy claims unsubstantiated.
Carnegie Mellon is putting computer science courses on YouTube for freecs.cmu.edu · 380 points · 74 comments · 3hprogrammingaiCarnegie Mellon is releasing full semesters of its computer science lectures free on YouTube, with one course premiering per day in mid-November and more slated for January 2027.

Key points

  • Premiering courses include Machine Learning Systems, AI Agents, Software Architectures, and Machine Learning in Production, among others.
  • Five further courses, including Probability and Computing and Multimodal Machine Learning, are scheduled for January 2027.
  • Self-paced courseware with instant feedback and badges is listed as next, with AI tutoring planned for it later.

What HN thinks

  • Commenters largely welcome the release, though several note the initial courses are introductory and ask for advanced topics like databases and distributed systems.
  • A recurring debate is whether LLMs make free lectures less necessary, with one commenter arguing fundamentals should not be learned from LLMs since they fabricate details.
  • One commenter cited a history of universities removing free course videos over accessibility and captioning liability, a possible risk for this effort.
Software is breaking down at a systemic levelbreck-mckye.com · 317 points · 196 comments · 5haisecurityA blog opinion piece argues that 2026's outages and exploits (Microsoft Azure, PHP2Shell, eMudhra, Shai Hulud) reflect a systemic decline in software reliability driven partly by AI-generated code.

Key points

  • The author describes the Microsoft outage as likely starting from a network configuration failure that led to a cascading failure, with the full cause kept confidential.
  • PHP2Shell is presented as a memory bookkeeping bug hidden in the PHP runtime for over a decade, exposed by an exploit Wiz shared on October 21.
  • The author uses the term 'decoherence', borrowed from Kristen Maevyn's essay, for locally valid but mutually incompatible models that accumulate in codebases.

What HN thinks

  • Several commenters argue the premise is historically shaky, noting cloud hosting, open source package ecosystems and widespread software failures predate the claimed 2016-era golden age by decades.
  • A recurring view is that the real shift is dependency sprawl and consolidation onto a few cloud vendors, which creates shared failure points and leaves many packages with temporary rather than experienced owners.
  • Some commenters say the bigger cause is business priorities favoring speed and cost over reliability planning, and that AI mainly amplifies existing practices rather than creating new failure modes.
What is w2rt6byl5nf1opmc-a.com?jack.cab · 295 points · 135 comments · 10hsecuritywebA personal blog author finds the domain w2rt6byl5nf1opmc-a.com ranked 8th on Majestic's .com list, with about 240 million nofollow backlinks despite the domain being registered only days earlier.

Key points

  • The domain was first registered, per Certificate Transparency, on November 11, after most of its backlinks had appeared, and it still does not resolve.
  • Majestic lists 5,163,708 referring domains, yet Ahrefs shows only 53 backlinks, and the author found just six subdomains in the downloaded Majestic Million list.
  • The Majestic Million ranks domains by the number of distinct referring subnets, so the author suspects an IPv6 or subnet-counting loophole, a copyright trap, or cloaked links.
  • The domain is also at rank 38 on Tranco, which still uses Majestic data, and the author calls it the fastest-growing domain in the top 100.

What HN thinks

  • Commenters note the domain was unregistered for most of its popularity window and was only bought through Cloudflare on Nov 11, which many find the strangest part.
  • Several commenters suspect the backlinks came from spam or a link-injection campaign, with Chinese-language anchor text and mostly nofollow links on inner pages suggesting a third-party origin rather than the linking sites themselves.
  • One commenter argues this is a loophole where ranking tools trust backlink counts, letting a random domain look authoritative, while another suspects it is malware or botnet-related command infrastructure.
The Secret ThinkPad Powerbutton Code to Bring Dead Laptops Back to Life (2007)masnick.com · 286 points · 113 comments · 12hhardwaretoolsA 2007 Techdirt post by Mike Masnick describes a power-button sequence that revived a dead ThinkPad, which IBM support recommended and which the post's comments report working for many readers.

Key points

  • The procedure: unplug the AC adapter, remove the battery, press the power button ten times at one-second intervals, then hold it for 30 seconds before reinserting the battery.
  • IBM support attributed the fix to static discharge but could not explain the underlying cause further.
  • Masnick's laptop had been dead with no lights and was very hot after being in sleep mode.

What HN thinks

  • Commenters broadly agree the procedure works as a way to discharge residual power, which clears a wedged embedded controller, though several think the exact number of button presses does not matter.
  • One commenter notes the 10-press sequence likely just makes customers hold the button long enough, a pattern they saw with router reset instructions in call-center work.
  • Several commenters report the trick is common across brands, including Dell, Asus, MSI and Lenovo, and one says it sometimes fixes only temporarily or the laptop fails again later.
Jiff 1.0: A datetime library for Rustburntsushi.net · 280 points · 64 comments · 7hprogrammingopen-sourceJiff 1.0 is a Rust datetime library that puts time zone handling at the center of its API, with a commitment to keep the 1.0 API stable indefinitely.

Key points

  • Its core types separate civil (wall-clock) date and time, Timestamp, TimeZone and ZonedDateTime, so ambiguous DST folds and gaps are resolved explicitly.
  • By default, ambiguous times in a fold pick the earlier instant and times in a gap move forward, with earlier, later and reject strategies available.
  • Span keeps each unit distinct, so 2 hours and 120 minutes are different values, and a friendly duration format avoids the lossy ISO 8601 output for 2000 milliseconds.
  • It is modeled on JavaScript's Temporal API, supports only the Gregorian calendar, and limits its datetime range to years +/-9999.

What HN thinks

  • Commenters widely praise the library's author and the API, with several saying Jiff is a pleasure to use and an easy switch from Chrono.
  • Several commenters note that Chrono is now semi-deprecated, which shapes many users' choice of Jiff over it.
  • One commenter with simulation experience found hifitime a better fit for non-Unix time scales, since Jiff supports only the Unix time scale.
Interpreting claims that a theorem is correct because it was checked in Leanxenaproject.wordpress.com · 276 points · 124 comments · 16haiprogrammingXena Project's Kevin Buzzard lays out four checks for judging claims that a theorem is correct because it compiles in Lean, especially for AI-generated code.

Key points

  • The first check is whether the Lean statement matches the paper's theorem, since a statement can compile yet be false, as a flawed Fermat's Last Theorem example shows.
  • Around 38 of the 290 OpenAI papers with Lean code do not state or prove the paper's main theorem, only intermediate results.
  • Anthropic's 13 million line Fermat's Last Theorem formalization needed a 96-core, 500 GB RAM machine before it could be compiled.
  • Lake comparator checks that the trusted statement matches the proved theorem and uses only standard axioms; lake comparator --paranoid adds checks across multiple independent kernels.

What HN thinks

  • Commenters broadly agree that the main risk is a mis-stated theorem rather than a broken proof, so checking that the Lean statement matches the informal claim is the key step.
  • One commenter notes the Lean kernel and the independent Con Leche kernel both had heavy LLM contributions, so the same kind of entity wrote both the checker and the proofs.
  • Several commenters argue the default Lean setup should enforce a comparator check such as lake comparator, since the cheating-via-shenanigans issue is largely a matter of bad defaults.
Frame – 31.5" color ePaper displaytrmnl.com · 264 points · 174 comments · 7hhardwaredesignTRMNL is releasing Frame, a 31.5-inch full color ePaper display in a handmade wooden frame, with a first batch of about 470 units open for pre-order.

Key points

  • Frame adds a clock-synced update schedule to reduce flicker and extend battery life, instead of a fixed every-N-minutes cadence.
  • A microSD port enables fully offline operation so no personal data reaches TRMNL servers.
  • Lazarus, a color conversion engine for Spectra 6 panels, is claimed to offer the best color matching available and will be open source.
  • Batch #1 ships only to USA, Canada, Mexico, and Australia; other countries face potentially high duty and import fees.

What HN thinks

  • Commenters broadly agree the $1,200 to $1,300 price is the main barrier, with many noting that cheaper 13-inch e-ink devices and Spectra 6 panels from other vendors cost far less.
  • Several commenters point out that the display panel itself is sourced from China and is the same Spectra 6 panel used in Samsung's and other vendors' products, so the wooden frame is the main US-made element.
  • One commenter with first-hand use reports that full refreshes take 20-30 seconds, colors are limited to six fixed tones, and photos look poor, making it best suited to paintings and text.
Nvidia was in trouble before the Jensen Huang messwired.com · 238 points · 222 comments · 8haihardwareWired reports Nvidia faces mounting competition in AI chips as rivals catch up, just as CEO Jensen Huang has been ousted and CFO Colette Kress takes over.

Key points

  • Nvidia holds about 80 percent of AI accelerator sales, while AMD's share rose from two percent in 2022 to 7-10 percent in 2026.
  • China fell to eight percent of Nvidia's revenue from 20 percent in 2024, despite Chinese sales of previous-generation chips doubling.
  • Wafer, a startup using AI to port models to AMD hardware, claims it runs new open-weight models at least 80 percent as efficiently within hours.
  • An analyst compares the situation to Intel's 2019 CEO change, arguing Nvidia may shift from engineering focus toward sales.

What HN thinks

  • Commenters broadly argue the stock drop reflects bubble dynamics and profit-taking after a roughly fivefold run-up rather than new technical threats, though some dispute whether competition fears played any role.
  • Several commenters compare the situation to crypto mining, where ASICs displaced GPUs, but others note AI models change too fast for fixed silicon, and that Nvidia's software ecosystem and general availability remain hard to match.
  • One commenter with domain knowledge notes that inference splits into prefill and decode stages with different hardware needs, and that Nvidia's dominance rests partly on R&D-to-production convenience rather than raw price-performance alone.
Tor Project Issues Rapid-Response DoS-Defense Fix Amid Sustained DDoS Attackforum.torproject.org · 236 points · 62 comments · 23hsecuritysystemsTor Project released version 0.4.9.17, a security update with an IPv6 relay DoS-defense fix, issued in response to an ongoing network-wide DDoS attack flooding relays with client circuits.

Key points

  • The release is described as rushed in response to an ongoing DDoS attack that has affected most relay operators.
  • Relays are being overwhelmed by a massive flood of client circuits, which the project says is not an easy problem to address.
  • The project says it is actively evaluating additional defenses and expects further releases in the near future.
  • Relay operators are asked to upgrade as soon as possible and need take no further action after upgrading.

What HN thinks

  • Several commenters say Tor has become much faster over time, with one citing Tor Metrics bandwidth data and saying YouTube plays in high quality, though others still find it slow and unreliable.
  • Commenters note that many sites block Tor exit IPs due to high abuse rates, making Tor hard to use for some services, and one commenter describes relaying through a VPS to reach a site.
  • One operator running a relay at home reported the attack as a sudden drop of about 99% in download speed, with the relay marked overloaded, and said lowering bandwidth limits did not help while disabling guard status did.
BugMeNot has closedbugmenot.com · 204 points · 106 comments · 2hwebpolicyBugMeNot, a site that shared free login credentials for registration-walled websites, has shut down, per a bare page with no explanation.

Key points

  • The article was unavailable, so the closure's reason and timing are not confirmed by the provided material.
  • Commenters recall shared logins often being a generic bugmenot@mailinator.com account with the password bugmenot.

What HN thinks

  • Commenters broadly agree the service stopped working well years ago, as sites added social login, 2FA, disposable-email blocklists and geo-based checks, and shared credentials were abused or expired.
  • Some commenters point to legal risk, saying publishing shared logins may violate terms of service and could raise CFAA concerns, though others think a lawsuit was unlikely.
  • One commenter with recent use says vendors such as Veeam, pfSense and VMware gate free downloads behind logins, so users now rely on Gists or throwaway accounts as substitutes.

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

Grazing sheep in fields with solar arraysnews.cornell.edu · 192 points · 187 comments · 6henergyscienceCornell researchers found that grazing two sheep per hectare under solar arrays raised beneficial insect abundance 38% over mowing, and released a spreadsheet tool for solar-grazing flock economics.

Key points

  • Over three years at Cascadilla solar farm in Dryden, NY, four sheep per hectare matched mowed-plot insect abundance while controlling vegetation better than two.
  • At eight sheep per hectare, vegetation control was maximized but honey bee abundance fell 45%.
  • The 83% honey bee increase drove most of the 38% gain at two sheep per hectare.
  • About half of U.S. lamb and mutton consumption is imported, mainly from New Zealand and Australia, which the authors cite as market opportunity.

What HN thinks

  • Commenters point out that solar fields in grassland, rather than barren desert, are often chosen because they sit closer to population centres and avoid long transmission lines, which face heavy opposition and slow construction.
  • Several commenters note that sheep suit solar sites better than cattle or goats, since cows can rub against and damage racks and goats may climb panels or chew cables.
  • One commenter notes that vegetation under solar arrays must be managed somehow, and that mowing around many support struts is labor-intensive, so grazing offers practical synergy beyond the insect results.
Researchers track whole body movement with just a smartwatchnews.cornell.edu · 192 points · 67 comments · 3daihardwareCornell researchers built EchoMotion, which tracks full-body pose using only the speaker and microphone of a single off-the-shelf smartwatch, emitting ultrasonic sound and reading echoes from the body.

Key points

  • Average error for predicting each of 12 joints was under four inches for new users, falling to about 2.4 inches after a one-time 12-minute calibration.
  • Sensing draws about 12 milliwatts, roughly 40 times less than a wearable camera, and a 30-minute session used 5% to 6% of a Galaxy Watch 6's battery.
  • The same AI model worked without retraining on a Galaxy Watch 7 and a Google Pixel Watch 3.
  • Performance dipped slightly in a narrow hallway because echoes bounced off other people, and the team expects more data to help in crowded settings.

What HN thinks

  • Commenters questioned how the system handles several people nearby, asking whether the ultrasonic signals from different watches would interfere with each other.
  • Some noted the audible-range concern: the 18 to 24.5 kHz sweep is within hearing for many people, especially children, though one commenter said the bands may be inaudible to most adults.
  • Several pointed out that smartwatches already estimate related metrics like respiration rate, rowing strokes, and gait from heart-rate and motion sensors, so this would be an extension of existing features.
Microsoft blames monster email outage on software defect introduced in July fixtheregister.com · 183 points · 128 comments · 10hsystemsprogrammingMicrosoft says a July 2026 gateway-server update contained a lock-contention defect that, under a load spike and capacity loss, exhausted CPU and memory across most of its fleet, causing a multi-day Exchange Online outage.

Key points

  • Load testing ran on weaker hardware with fewer cores, and Microsoft says the contention issue worsens on machines with more cores and newer hardware.
  • At its lowest, more than 80 percent of servers were unresponsive, and restarting them only led to them being overwhelmed again.
  • A rollback was not possible because the defect had been live for months, so engineers shipped a patch that restored the earlier Windows code path.
  • Exchange Online could not fail out of the faulty gateway for scale reasons, unlike Teams, OneDrive and SharePoint, which did fail over.

What HN thinks

  • Commenters compare the outage to a cascading failure or black start, where load balancers shift traffic onto remaining servers until they also fail, and retries keep pressure on the system during recovery.
  • Several commenters argue the core problem was testing: the test environment used low-end hardware, which hid a concurrency bug that appears only on high-core production machines.
  • Skeptics question whether the 15-hour fix time and lack of cell-based isolation were acceptable, and link recent layoffs and AI-driven changes to a rising risk of such outages.
Whistleblower accuses Pentagon over Iran strike on US base that killed 6cnn.com · 177 points · 49 comments · 9hpolicysecurityA surviving Army major filed a whistleblower complaint alleging the Pentagon ignored warnings that a Kuwaiti operations center lacked protection against drones, before a strike killed six reservists.

Key points

  • Maj. Stephen Ramsbottom says he raised lack of aerial protection with superiors and was told "We don't have to worry about protection."
  • He says a promised counter-drone system never arrived at Shuaiba port and it is unclear it was ever requested.
  • The site reportedly had no physician or fixed aid station and only one ambulance when the March 1 strike hit.
  • Pentagon officials had called the building fortified; a source said it had only concrete barriers against car bombs.

What HN thinks

  • Commenters largely agree the institution underestimated drone threats, noting that drone warfare in Ukraine and Russia had been visible for years before the strike.
  • One commenter argued that prior to the conflict it was reasonable not to expect a full-scale air war with Iran, and that the military must be ready for whatever politicians commit it to.
  • Some commenters doubt the complaint will change much, since institutions rarely invest in defenses against threats they have not yet experienced, and a serving officer is unlikely to criticize the war itself.
Coffeehouses and the Rise of Science [pdf]monirbounadi.com · 171 points · 98 comments · 13hhistoryscienceA PDF job-market paper by economist Monir Bounadi argues that coffeehouses helped drive the rise of science, but the article text was not available for review.

What HN thinks

  • Commenters largely agreed that coffeehouses and pubs once fostered open conversation across classes, but that modern cafes, with small tables, laptops and phones, now mostly encourage solitary work.
  • One commenter, a Cambridge resident, reported having more scientific discussions in pubs and college bars than in cafes, and a few suggested pubs or purpose-built conversation spaces might work better today.
  • The strongest caveat was that coffeehouses may be a correlate rather than a cause of scientific activity, since they could simply open in areas already attracting scientists, and commenters asked how this could be tested.
Rust changed me as a developerblog.startifact.com · 147 points · 82 comments · 3hprogrammingA developer with six years of Rust experience argues that the language's sum types, compile-time macros, ownership rules, and arenas changed how he designs software, more than any other language he has used.

Key points

  • Sum types (enums) model optionality and errors directly, and the compiler forces every variant to be handled, which he argues is lower ceremony than interfaces.
  • Serde and clap use compile-time macros to get serialization and CLI parsing without runtime reflection, which is how Rust matches Python's pickle or Pydantic.
  • Ownership rules reject circular references and shared mutation, so he stores tree nodes in arenas and uses indexes for parent and child links.
  • He built Xot, an XML tree library with parent references, after finding no existing Rust crate that offered both mutability and parent links.

What HN thinks

  • Commenters broadly agree that learning a language far enough from one's usual style, whether Rust or another, changes how one thinks about design, and several recommend building a substantial project in different language families.
  • Rust's forced handling of failure cases, such as Option and newtypes for distinct ID types, is praised as a way to guard against one's own mistakes, echoing the purpose of unit tests.
  • One commenter criticizes Rust's enums for conflating the tag with the data, noting that a standalone discriminant type like Zig's or Ada's cannot be named or iterated over without a dummy value, though another replies that discriminant and repr offer partial workarounds.
Debian Becomes Unresponsive after 6.1.0-55 Kernel Updatebugs.debian.org · 128 points · 54 comments · 19hsystemsopen-sourceA Debian bug report says systems become unresponsive after updating to the 6.1.0-55 kernel, though the article text could not be retrieved to confirm details.

What HN thinks

  • Commenters point to a linked kernel mailing list fix for an ext4/ntfs3 null-pointer issue, and say people seem satisfied with the proposed patch.
  • Several commenters note that Debian users can avoid the problem by keeping an older kernel installed at boot, since apt tends to suggest removing old kernel images only after upgrades.
  • One commenter argues that the issue is a reason to move bookworm systems to trixie and its 6.12 kernel, though others counter that stable releases often carry compatibility issues when upgraded.
OOS – an open operating system for the Rigol Scopeseevblog.com · 126 points · 36 comments · 3dtoolsopen-sourceOOS is a community-built replacement operating system for Rigol oscilloscopes, developed by Julian Decker, that runs on existing Rigol hardware.

Key points

  • The article text was unavailable, so details come from the thread and the linked repository rather than the forum post itself.
  • Commenters describe it as a rewrite of Rigol's scope firmware including signal processing and UI code, not a simple add-on.

What HN thinks

  • Commenters praised the project as a serious reverse-engineering effort, noting a scope combines custom analog, digital, FPGA and UI work that a full vendor team normally builds.
  • The FPGA side is the hardest part: its bitstream is effectively opaque and pin mappings are hard to determine without desoldering, so one commenter argued the FPGA work is largely starting from scratch.
  • One commenter noted that the repository suggests more was done than the coverage implies, with the host-side Rigol application reverse-engineered and measurements working, though the FPGA configuration appears to be left untouched.
United Airlines experiencing a widespread system outage, airline sayscnn.com · 123 points · 46 comments · 2hbusinesssystemsUnited Airlines reports a widespread system outage affecting departing flights, with operations not yet restored as of the November 16 report.

Key points

  • The airline said flights already in progress and United Express flights are not affected by the outage.
  • The FAA is aware of the incident and has referred questions to United rather than commenting on the cause.
  • The outage follows a rare FAA ground stop days earlier that halted all US departures for hours.

What HN thinks

  • Several commenters tie the run of recent outages across companies to AI-driven attacks or AI-accelerated code changes, though others call that speculation without evidence.
  • One commenter argues airlines have always had fragile systems and that a 13% delay rate on one carrier is not unusual, suggesting media attention is inflating the pattern.
  • A recurring view is that reliability loses priority to shipping features, and that outages are a predictable result of cost-cutting and layoffs rather than a single cause.
Do we still have to denormalize? Have your schema and search it tooparadedb.com · 113 points · 40 comments · 3dwebsystemsParadeDB 0.26.4 adds search joins that run multi-table OR search queries over normalized tables, with a Rank Join for Top-K and columnar aggregation for facets, without denormalizing.

Key points

  • On the StackOverflow benchmark, a natural 3-table OR Top-K query drops from about 118,000 ms in vanilla Postgres to 100-600 ms on a single core.
  • The Rank Join streams each table in descending score order and uses random lookups by join key to resolve high-scoring partial candidates early.
  • Facet queries across 20M posts and 2.3M users run in 422 ms versus 35.3 s for a standard Postgres join, using tuple tagging and index co-partitioning.
  • The executor is built on Apache DataFusion and intercepts search-join plans via planner hooks, leaving non-search queries on Postgres's standard executor.

What HN thinks

  • Commenters largely agree that search indexes should be rebuilt and iterated on, and that a separate ETL layer between the database and search index helps with enrichment, embeddings and reindexing.
  • The main criticism is that searching across a joined multi-table query can let unrelated child rows, such as posts or comments, drive results; one commenter said running separate queries and merging them is cleaner for their case.
  • Several commenters argued that denormalization brings consistency costs when frequently updated fields change, and that keeping data normalized while indexing a denormalized view is often the better tradeoff.
Squircleslinebender.org · 105 points · 34 comments · 16hdesignwebA reference wiki catalogs squircle curves, from the superellipse to Apple's iOS icon shape and Figma's approximation, and compares their continuity and parameterization.

Key points

  • Apple's iOS 7 icon shape is a flat-sided squircle, not the exponent-5 superellipse early analyses suggested, and has a fixed gauge of 0.809 with no extra parameter.
  • The Figma squircle without a flat side only reaches gauges between 0.707 and 0.854, while the clothoid variant reaches only 0.790.
  • CSS corner-shape standardizes real superellipse corners, and Chromium uses a two-Bézier closed-form approximation exact for the quadrant midpoint.
  • An exponent-3 superellipse with a matching flat side approximates Apple's shape closely and has G2 continuity, unlike the Figma approximation's arc join.

What HN thinks

  • One commenter explains that a squircle differs from a rounded rectangle because its corner curvature changes gradually rather than using a fixed-radius quarter circle, and another notes its sides are never fully straight.
  • Commenters point out that squircles nest more gracefully than ordinary rounded rectangles, which is a practical reason design tools adopt them.
  • Opinion is split: some find squircles ugly or dislike rounded UI corners in general, and one notes squircle corners are only supported as a CSS feature in Chrome, not in Firefox or Safari.
Coles hit by a cybersecurity incidentcolesgroup.com.au · 103 points · 78 comments · 23hsecuritybusinessColes Group, one of Australia's two major supermarket chains, says it is investigating a cyber security incident, with no details of scope or data affected on its updates page.

Key points

  • The company's updates page carries only a one-line notice and a link to 'latest updates', so the article gives no detail on the breach itself.

What HN thinks

  • Commenters say breaches at large Australian firms are frequent and that cybersecurity is often treated as a box-ticking compliance exercise rather than real practice.
  • Several argue large firms have weak incentives to invest in security, since a breach rarely hurts a dominant supermarket's customers or profits, and cyber insurance or self-insurance further blunts pressure.
  • One commenter notes Coles fired about 1000 local back-office staff and offshored the work via Accenture earlier in the year, linking cost-cutting to the risk.
The Active Directory Era Is Finally Overinfisical.com · 96 points · 66 comments · 9hsecuritysystemsInfisical's marketing writer argues Active Directory cannot track machine identities in cloud environments and proposes federation plus a secrets-manager broker as replacements.

Key points

  • AD machine accounts were added by a human admin, but developers now spin up Kubernetes pods, CI runners, and databases continuously, so admin approval does not scale.
  • Many organizations auto-join new machines to AD using an admin token, which the author says defeats the purpose since no human vouches for each device.
  • Each API key, database string, or token forms a separate perimeter, and the author says secrets are often kept in plaintext .env files or hardcoded configs.

What HN thinks

  • Several commenters push back that Active Directory is not a bottleneck for machine onboarding, noting the default machine account quota lets regular users add up to 10 machines without admin approval.
  • Commenters argue gMSA already solves the rotation and credential-sprawl problems the article raises, with automatic high-entropy passwords retrievable only by authorized hosts.
  • Commenters note AD still underpins on-prem environments with many managed PCs, such as clinic devices, where cloud-only management is not viable, though laptops and mobile devices have shifted work toward Entra and cloud services.
Aristotle Expands into Rustharmonic.fun · 92 points · 47 comments · 6haiprogrammingHarmonic's Aristotle formal verification tool now supports Rust, claiming it can prove absence of undefined behavior and conformance to specs, including inside unsafe blocks.

Key points

  • Can prove a Rust C-rewrite behaves the same as the original C code.
  • On the open-source Manifest order book it proved 28 results and found a rounding bug the team has since fixed.
  • On Project 0, a prime broker, it proved books stay consistent through deposits, withdrawals, borrows, repayments, and liquidations in any order.
  • Pricing is self-serve and consumption-based, with volume pricing available by contacting sales; the Solana Foundation is partnering on support.

What HN thinks

  • Commenters asked how the specifications are produced: the company says natural-language intent is auto-formalized into theorems about a Lean embedding of the Rust MIR, but one commenter objected that the embedding is proprietary and wants to inspect the formal spec directly rather than trust auto-formalization.
  • Several commenters pointed to other Rust verification tools such as Verus, VeriFast, RefinedRust and cargo-anneal, noting that proving soundness inside unsafe blocks is an active research area with very different approaches.
  • One commenter noted that Aristotle's cost was steep in practice, with free credits used up quickly on a single check, and another asked whether the tool works by compiling to Lean.
Strengthening the fight against online child sexual abuseopenai.com · 90 points · 190 comments · 6haipolicyOpenAI published a post about its efforts to strengthen the fight against online child sexual abuse, but the article could not be fetched (HTTP 403).

What HN thinks

  • A commenter who says their account was banned after an OpenAI classifier flagged over 100,000 images reports being refused details, so they could not report the uploader to NCMEC, and calls the effort performative.
  • Several commenters argue that OpenAI's CSAM scanning could serve as a pretext for data collection and surveillance, and that any use of such tools risks broader scanning of users' images.
  • Commenters note that major cloud and AI vendors such as AWS, Azure, and OpenRouter already scan zero-data-retention endpoints, and one commenter says the motive is likely legal exposure from having 'actual knowledge' of content.

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

Jolt – Clojure that speaks Cjolt-lang.net · 89 points · 26 comments · 7hprogrammingsystemsJolt is a self-hosted Clojure implementation running on Chez Scheme or Gambit (compiled to JavaScript), with no JVM, a native C FFI, and a jolt build command that emits standalone binaries.

Key points

  • jolt build AOT-compiles a project with its runtime, standard library, and dependencies into one executable, with C dependencies linked statically.
  • defcfn wraps C functions as ordinary Clojure fns, and the db and http-client libraries bind SQLite, Postgres, OpenSSL, and zlib this way.
  • Java interop, reflection, gen-class, and proxy are unavailable; a shimmed subset of java.* is provided instead.
  • BigDecimal is unsupported (decimal? always returns false), and regex uses irregex rather than java.util.regex, so some Java-specific patterns differ.

What HN thinks

  • One commenter contrasts Jolt with Jank: Jank targets tight C++ integration and is not a drop-in Clojure replacement, while Jolt aims to be a superset of JVM Clojure with native FFI and runs many existing Clojure libraries.
  • A substantial sub-thread criticizes the project's LLM-generated README and hard-to-read Makefile, and argues that an AI-contribution policy citing human context limits does not justify making the codebase hard to understand.
  • The author responds that a large specification and conformance test harness is the intended way to keep semantic correctness manageable across the codebase.
Winudf – DVD / BluRay driver for DOS, Windows 1.x – 3x / 95 / 98 / MEgithub.com · 80 points · 13 comments · 16hopen-sourcesystemsWINUDF is a GPL-3.0 driver that reads and writes UDF 1.02 to 2.60 discs and USB sticks under DOS, Windows 1.x to 3.11, and Windows 95/98/ME.

Key points

  • WINUDF.EXE gives DOS and Windows 1.0x to 3.11 drive letters for UDF volumes, while WINUDF.VXD adds long-name UDF support to Windows 95, 98 and ME.
  • Version 1.1 adds the MS-DOS 7 long file name API, so DIR and COPY show long names on UDF disks.
  • Version 1.1 adds writing to Windows 10 UDF 2.60 disks, with the metadata bitmap embedded in its file entry.
  • Version 1.2 fixes a Windows 9x crash (0xC0000005) caused by page-in of a program's own image from the same disc during a read or write.

What HN thinks

  • Commenters questioned why the project reimplements UDF instead of reusing VideoLAN's libudfread, with replies noting that libudfread's modern C targets and the 16-bit segmented DOS/Win16 environment make direct reuse difficult.
  • One commenter argued the effort is of limited practical value, since the driver's size in conventional memory makes it hard to use on DOS and modern toolchains already cover Windows 9x.
  • Others framed the project as retro-computing enthusiasm, noting that hobbyists add hardware support to old operating systems that modern Linux distributions no longer serve well.
Mini.nvim Is Now on Tangledtangled.org · 80 points · 29 comments · 24htoolsopen-sourceThe mini.nvim plugin collection now has an official mirror on Tangled, alongside GitHub as the primary and Codeberg as another mirror.

Key points

  • The repository was moved from echasnovski/mini.nvim to a dedicated nvim-mini organization to improve long-term project stability.
  • The mirror installs with the same URL structure as GitHub, and the main branch is the default for installation.
  • The project contains 45+ independent Lua modules and requires Neovim 0.10 or higher.

What HN thinks

  • Commenters note the Tangled repo is a mirror, with GitHub still the primary host where CI likely runs, and one asked whether Tangled supports CI/CD at all.
  • Codeberg's FAQ says it does not pull-mirror from other forges for resource reasons, so mirroring there is done by pushing manually.
  • Commenters explain the 'mini' name as referring to each module being small, independent and easy to adopt one at a time rather than to the module count.
Michael Mann's 'Heat' Copyright Fight Turns on a Deal with Himselfcopyrightlately.com · 79 points · 34 comments · 6hpolicycultureNew Regency is suing to invalidate three termination notices Michael Mann served to reclaim rights in Heat, arguing he wrote the screenplay as a work for hire through his loan-out company.

Key points

  • Mann's claim rests on an oral 1988 deal with his own companies, which New Regency says nobody can document, and which his story reportedly changed on under questioning.
  • Mann says he wrote early Heat drafts in 1981 to 1985 as an unemployed author, a period not covered by any written agreement New Regency has produced.
  • Even if an oral grant could be terminated, it would likely be only a nonexclusive license, so Mann might win termination and still not recover the exclusive rights he wants.
  • Heat 2 may survive termination under the derivative works exception if it counts as 'prepared' before termination takes effect, which is an unsettled question.

What HN thinks

  • Commenters largely focus on Heat 2 rather than the copyright dispute, with one saying they are hesitantly looking forward to it while noting the source novel is largely a mashup of Heat and Miami Vice.
  • One commenter praises Mann's cinematography and attention to detail as the main draw of his films, and says Miami Vice is their favorite of his work over Heat.
  • A few comments make jokes about copyright and the case, and the thread offers little substantive debate on the work-for-hire question itself.
Tell HN: HN is being scraped for emails to scam76 points · 17 comments · 21hsecuritycultureA user reports receiving a fake AWS payments email at the address they posted on HN, suggesting the site's public emails are being harvested for scams.

Key points

  • The poster used a unique email for each site, which made the fake AWS payments email easy to spot as fraudulent.
  • The poster advises hovering over sender names and links to check the actual domain before opening or clicking anything.
  • The poster says some social media accounts set to private have also started receiving spam at addresses tied to them.

What HN thinks

  • Commenters widely report that email addresses posted on HN, especially in the 'Who wants to be hired?' thread, are receiving scam, spam, and automated cold outreach.
  • Several commenters now use throwaway or per-site email addresses, and one reports the switch noticeably improved the ratio of legitimate to spam mail.
  • One commenter suspects data brokers link posted addresses to personal profiles, which explains scam emails containing personal context that is nonetheless inaccurate.
World's Largest Brands Infected With Malware Via Optimizelyblog.report-uri.com · 72 points · 14 comments · 3hsecuritywebA malicious script injected into Optimizely tag bundles reportedly swaps crypto wallet addresses and skims card details on about 15 major websites, including CNN, Zoom and Fox News.

Key points

  • The injected loader appends a script from 1688360[.]cc that carries a clipper module covering ETH, BTC, SOL, TON and eight other chains with checksum validation.
  • The card skimmer validates Luhn numbers and CVV format, then sends data via sendBeacon to a /card.php endpoint on the attacker's host.
  • Two affected sites blocked the loader with a Content-Security-Policy script-src rule, so the payload never executed.
  • On iPhone and iPad the payload pulls a second-stage boot.js from two .icu hosts; on iOS 18.6 and earlier it also adds a hidden iframe.

What HN thinks

  • Commenters point out that Subresource Integrity alone does not stop a compromised script from loading other scripts, so Content Security Policy domain allowlisting is needed to limit this kind of attack.
  • One commenter notes that in practice CSP is rarely enforced because SaaS vendors and marketing teams frequently add third-party scripts without security review.
  • Several commenters criticize the linked write-up as low-quality, though a more detailed report from report-uri covers how CSP did or did not help in this attack.
My Quest to push native Three.js quality match AAA Game Enginesben3d.ca · 72 points · 20 comments · 3hprogrammingwebA developer's progress report on building AAA-style lighting (lightmaps, light probes, screen-space reflections and GI, TAAU) into Three.js, listing which stages exist, which are in review, and which are still missing.

Key points

  • Octahedral RG16 snorm normals cut mean normal error from 0.17° (RGB8) to 0.0013°, with about 93% of WebGPU devices supporting the format.
  • LightProbeGrid bake time for a resolution-12 grid fell from 5.6 s to 1.2 s by rendering eight probes per pass with an array camera.
  • A Hi-Z SSR prototype matched linear marching bit for bit but was slower overall: on a GTX 1050 one scene fell from 199 to 159 fps.
  • The biggest unfinished piece is a pluggable indirect specular and diffuse path that would let screen-space results feed a material's BRDF, which the author says is the largest missing gap.

What HN thinks

  • Many commenters judged the post to be AI-generated and said that undermines its value, with several saying they only care whether the work functions and not how the write-up reads.
  • The author replied that the series was refined over about two months using AI tooling and that over 20 PRs have been merged, and said they are a long-time three.js contributor.
  • One commenter questioned whether AAA-level graphics are a realistic goal for a JavaScript web rendering engine, while another found the write-up informative for understanding lighting and reflection settings.
Just 400 US Data Centers Use More Electricity Than Michiganwired.com · 72 points · 59 comments · 6haienergyA Harvard study counts 403 US hyperscale data centers that used electricity comparable to all of Michigan over May 2024 to April 2025, mostly from fossil fuels.

Key points

  • Data center emissions are reported as roughly three times their pre-AI-boom level, based on regional grid energy mixes.
  • Virginia, Oregon, Ohio, and Iowa together accounted for more than half of the studied energy use.
  • The average facility in the sample is 40 megawatts, small next to proposed campuses of hundreds or thousands of megawatts.
  • The figures exclude behind-the-meter gas plants; up to 189 gigawatts of such generation is reportedly in the US development pipeline.

What HN thinks

  • Commenters largely dispute the headline framing: the study covers 403 hyperscale facilities specifically, and some argue comparing them to Michigan's total usage is cherry-picked, since many other industries also consume large amounts of power.
  • Several commenters note that the comparison could be checked against other measures, such as the economic output of the companies involved, though one argues that data centers serve far more economic activity than their direct revenue suggests.
  • A recurring caveat is that data center demand strains grid capacity and that much of the power comes from fossil fuels, with one commenter citing renewables supplying only about 22 to 27 percent of AI and data center electricity globally.
Comcast's "fiber-powered Wi-Fi" is cable, AT&T says in false advertising casearstechnica.com · 71 points · 57 comments · 3hbusinesspolicyAT&T sued Comcast in federal court, claiming its 'fiber-powered' Xfinity Wi-Fi marketing is false because the last mile to homes is coaxial cable, and seeks Comcast's profits from the ads.

Key points

  • The lawsuit was filed in the US District Court for the Northern District of California and alleges violations of US and California competition laws.
  • AT&T seeks restitution of profits plus a permanent injunction barring Comcast from describing its hybrid fiber-coaxial service as 'Fiber-Powered'.
  • An industry self-regulator found in June that Comcast's coaxial disclosure was not prominent enough, though it accepted the underlying description as accurate.
  • AT&T has a parallel suit against Charter over similar 'fiber-powered' language, where Charter argues 'powered' means the backbone relies on fiber.

What HN thinks

  • Commenters widely note that AT&T has used similar loose 'fiber' and '5G E' marketing itself, so the lawsuit reads to many as hypocritical rather than principled.
  • Several point out that most of the last mile is coax for Comcast, Spectrum, and Cox, and that 'fiber-powered' and 'gig' labels obscure asymmetric upload speeds like 1000/30 Mbps.
  • One commenter with a Spectrum connection says upstream bandwidth collapsed under load until switching to real fiber, a concrete example of the gap the ads hide.
Artificial Intelligence for the Z80dfsantos.com · 71 points · 8 comments · 3daiprogrammingA 2026 book teaches AI concepts from scratch by implementing them in BASIC on the ZX Spectrum and Spectrum Next, covering search, regression, classification, clustering, perceptrons and Game of Life.

Key points

  • Twelve chapters progress from sorting-algorithm timing and expert systems to ELIZA, minimax with alpha-beta pruning, and k-means clustering.
  • Neural network chapters build up to a digit-recognition project on a 6 × 6 grid using backpropagation.
  • Linear regression uses real Mauna Loa CO₂ measurements, and the Game of Life runs on a toroidal 20 × 20 board with two arrays.
  • The book is available in Spanish and English editions, with code in a public GitHub repository.

What HN thinks

  • One commenter argued the title should say ZX Spectrum rather than Z80, since the programs run in BASIC and the lessons are not specific to the CPU.
  • Another commenter said the book resembles earlier 1980s AI-on-microcomputer work and pointed to a retrospective list of that era's achievements.
"Who is Hiring?" postings stuck near 12-year lows as AI reaches half of all jobshntrends.com · 70 points · 40 comments · 22haicareerHN's monthly Who is Hiring thread for November 2026 has 244 postings, down 37% year over year, with AI mentioned in 51.6% of them.

Key points

  • August's 224 postings was the fewest since January 2014, and every thread since July has landed between 224 and 255.
  • Front End mentions fell from 27% of postings in early 2022 to 6.6%, while Staff roles rose to about 12%.
  • MCP went from 0.3% of postings a year ago to 4.5%, the largest climb in this month's rankings.
  • Rust reached 8.2% of postings, up from 4.9%, and has been ahead of Go in four of the last five months.

What HN thinks

  • Commenters broadly doubt that HN postings reflect the whole software market, noting that many startups now post on applicant tracking boards or use recruiters instead.
  • Several commenters question attributing the drop to AI, pointing to interest rate rises around 2022, Section 174 tax changes, and a return to pre-2015 posting levels.
  • Commenters who recently searched for work report that new graduates and laid-off developers face long searches, while overall unemployment numbers remain relatively low.
Simple 8085 Computerhackster.io · 64 points · 8 comments · 4dhardwaresystemsA hobbyist published a six-chip 8085 computer design with 8kB ROM and RAM, a 4.9152 MHz clock, a 2400-baud serial port, and Tiny BASIC built into ROM.

Key points

  • The design uses an 8085 CPU, a 27C64 EPROM, a 6264 SRAM, and an SN74LS373 latch, with three simple logic chips total.
  • Tiny BASIC, which was written for the 8080, runs unmodified on the 8085 and is used to generate a Mandelbrot set.
  • The author claims undocumented 8085 instructions (ARHL, RDEL) cut a Mandelbrot assembly run from 43.34 to 40 seconds.
  • The project includes a self-modifying program that flips an RAL instruction to RAR by XORing one bit, toggling LED direction.

What HN thinks

  • One commenter asked how hard it would be to port the design to an Intel 8088 or NEC V20 clone, the only comment in the thread.
OpenCode Browseopencode.ai · 62 points · 31 comments · 2haitoolsOpenCode Browse is an MCP server plus Chromium extension that lets AI agents drive your real, logged-in browser, running in background tab groups.

Key points

  • Works with Chrome, Edge, Brave, Arc, Helium, Vivaldi, Opera and Chromium, using existing tabs, logins and extensions.
  • Agents can read pages, click, type, take screenshots, and record tabs, including network traffic.
  • For logins, 2FA or payments, the agent hands the tab back to you and resumes after you click Continue.
  • Users can restrict an agent's tab to a named list of allowed sites.

What HN thinks

  • Commenters point out substantial overlap with existing tools, notably the Playwright MCP server and Chrome's official chrome-devtools-mcp, and ask why this one is needed.
  • One commenter who already uses Playwright MCP with Claude Code and Codex says its value is granting access to pages and sessions already open, and warns to use separate browser profiles given the lethal trifecta risk.
  • One user reports the connection frequently drops and requires rerunning 'browse pair', and suggests it may become unnecessary once OpenCode's built-in browser tool gains persistent sessions.
The Lasting Influence of John Bergernewyorker.com · 56 points · 5 comments · 5dculturehistoryMerve Emre reviews Tom Overton's new biography of John Berger, arguing it is an experimental, uneven portrait that softens Berger's treatment of women and his late turn to pastoral politics.

Key points

  • Overton's biography, the second since Berger's 2017 death, is written in the present tense and addresses 'John' directly, speculating about his thoughts.
  • Emre says the book passes over significant events quickly while magnifying small details, and she finds the 1973 departure from his children hard to excuse.
  • Berger's 1972 fame came from the BBC series Ways of Seeing and the novel G., which won the Booker Prize that year.
  • Emre calls Berger's late Quincy novels a 'Romantic Marxism' closer to what she terms Rousseauian socialism than to class-based politics.
It's always DNS, so why are we running our own?s2.dev · 55 points · 36 comments · 14hsystemswebS2 replaced Route53 with a self-built authoritative DNS server that reads basin records live from its Spanner database, after Route53's API limit of about 5 UPSERTs per second proved too low.

Key points

  • Route53 API throttling (about 5 UPSERTs per second, at least when written) stalled normal DNS updates for hours after a customer created tens of thousands of basins.
  • The server is written in Go using miekg/dns, answering A, AAAA, NS, CNAME and SOA queries, with each query doing a live Spanner SELECT.
  • Cloudflare DNS Firewall sits in front of two GKE clusters in separate regions, and origins accept traffic only from Cloudflare IPs.
  • ACME DNS-01 challenges are offloaded to a separate Route53 zone via CNAME delegation, so the server never serves dynamic TXT records.

What HN thinks

  • Commenters questioned the design choice of explicit per-basin DNS records over a wildcard-per-cell scheme, noting the wildcard approach would avoid putting a database lookup on every query and the firewall risk that comes with it.
  • One commenter argued that DNS is designed to be eventually consistent, and suggested caching or offering longer TTL tiers to reduce database load, while the author clarified they meant atomic creation rather than strong consistency.
  • A commenter with first-hand experience reported that a client relying on instant DNS consistency was asked to add a short delay after creating a basin, and complied.
How Billions in Safe Insurance Money Is Being Bet on A.I. Debtnytimes.com · 54 points · 20 comments · 6haifinanceUS life insurers held $85.2 billion of AI debt as of June 30, up 24% in six months, and Moody's expects that could reach $185 billion by the end of next year.

Key points

  • AI debt is still only 2.1% of life insurers' total debt, but exposure is growing quickly.
  • xAI used a separate financing entity to borrow for chip purchases and lease them back, so xAI was not formally the borrower.
  • Athene bought $2.4 billion of that xAI-linked debt, per a Moody's report, showing insurers funding AI firms directly.

What HN thinks

  • Commenters draw parallels to 2008, arguing that rating agencies may be understating risk on AI debt because they earn fees and face little penalty for miscategorizing it.
  • One commenter notes that prior DOJ and SEC settlements with rating agencies show the regulatory history behind these concerns.
  • Some commenters dispute the crisis comparison, arguing AI debt is far smaller than subprime and that the firms involved are not critical to the US economy, while another says insurers are more directly tied to ordinary people than banks were in 2008, so losses could raise premiums.
Chemistry is Actually Geometrychillphysicsenjoyer.substack.com · 53 points · 11 comments · 3hsciencemathA long explainer shows how chemical reaction networks can be analyzed with linear algebra, using the deficiency zero theorem to count equilibria from network structure alone.

Key points

  • Deficiency is defined as complexes minus linkage classes minus stoichiometric subspace dimension, and a deficiency-zero weakly reversible network has exactly one positive equilibrium per compatibility class under mass action kinetics.
  • Feinberg's toy network has five complexes, two linkage classes and a stoichiometric dimension of three, giving deficiency zero, so its messy kinetic equations have a single positive equilibrium per class.
  • The Michaelis-Menten system has deficiency zero but is not weakly reversible, so the theorem does not guarantee a strictly positive equilibrium for it.
  • The global attractor conjecture, that all positive starting concentrations approach equilibrium, remains unproven even for deficiency-zero networks.

What HN thinks

  • Several commenters push back on the framing, with one arguing that calling chemistry 'actually geometry' is as much a stretch as calling carpentry or orthopedic surgery geometry, even if diagrams can illustrate it.
  • A recurring caveat is that mathematical models are maps rather than territory, and that only experimental results can confirm whether a model of reaction networks matches nature.
  • One commenter found the explainer's material on oscillatory behavior in chemical reactions surprising and interesting, noting how common such phenomena are.
Skydio's asymmetrical F10 Lightrunner solves a problem with fixed-wing dronesspectrum.ieee.org · 53 points · 11 comments · 24hhardwareaiSkydio's F10 Lightrunner is a fixed-wing camera drone with a wingtip-mounted camera, asymmetric wings and offset motors, designed to avoid occluding the camera while turning.

Key points

  • The camera sits on one wingtip so turns are only left, avoiding wing or tail blocking the view during circular tracking.
  • Mounting a counterweight on the other wingtip was rejected because the camera's drag would roughly halve flight time.
  • The asymmetry is offset by extending one wing for extra lift and repositioning the motors to balance thrust against camera drag.
  • A single model predictive control policy covers hover through 100 mph flight, with no separate hover or flight mode.

What HN thinks

  • Commenters note the asymmetry follows from the wingtip camera placement rather than being the occlusion fix itself, and one suggests a chin-mounted sensor like fighter jets use would avoid the problem.
  • One commenter argues the thin design may be driven by compact flatpack storage, and that a nose-mounted drop-down package would be simpler.
  • Several commenters criticize Skydio's law enforcement focus and its drones' surveillance use, with one noting the company has left the consumer market.
How to Break into Security, Ptacek Edition (2012)krebsonsecurity.com · 51 points · 12 comments · 1hcareersecurityKrebs republishes Thomas Ptacek's advice for breaking into infosec: learn to program, focus on application security, and practice on deliberately outdated software in a lab.

Key points

  • Ptacek's concrete path is to run old WordPress releases on EC2, then find bugs with Burp Suite Free Edition or OWASP WebScarab, using The Web Application Hacker's Handbook.
  • He argues appsec roles are the strongest jobs because newer security products and team structures are being shaped by appsec practitioners.
  • For IT staff, he suggests moving into QA or DevOps and instituting basic security testing, since many companies have immature security programs.

What HN thinks

  • Several commenters note the post is from 2012 and should be labeled as such, since it predates widespread AI use and was written when bcrypt advice was not yet standard.
  • Commenters broadly question whether the advice still holds in the age of AI agents and cheap vulnerability discovery, with one arguing that the roles may become less interesting and less well paid.
  • One commenter argues that AI lowers the barrier to entry so much that fundamentals matter less than knowing how to direct AI tools, though others see the document as simply dated.
Stop Filming in Logdigitalfilms.wordpress.com · 51 points · 35 comments · 19hdesignVeteran colorist Oliver Peters argues that log and raw recording are often unnecessary for filmmakers, and that shooting in Rec. 709 with correct exposure usually gives better results.

Key points

  • Log profiles were popularized by ARRI's Alexa in 2010, with Log-C designed so wide dynamic range could be recovered in post.
  • Sony's many profiles, including S-Log, S-Cinetone and s709, vary in how easily they grade, and Peters prefers S-Cinetone.
  • Peters says HDR video reproduces a high-contrast scene in a way that looks painful, unlike how the eye actually adapts.
  • Shooting log often means underexposing, which the author calls a common mistake that undermines the supposed benefit.

What HN thinks

  • Several commenters argue the article does not explain the concrete problems with grading log footage, and note that established camera brands offer conversion paths that reproduce Rec. 709 results.
  • A recurring point is that log's purpose is to preserve highlight and shadow information through quantization in post, not to dictate a look, so shooting log is not inherently wrong.
  • One commenter with color experience says log is an idealized mathematical curve rather than what the sensor actually does, and suggests a wide-gamut in-camera profile may be a reasonable alternative.
How to Intercept HTTP/3 Traffic: QUIC and Qpackinterceptsuite.com · 50 points · 3 comments · 4hsecuritywebInterceptSuite, a raw-traffic MITM proxy, can intercept HTTP/3 by relaying QUIC over UDP, with an open-source extension decoding HTTP/3 and QPACK into editable text.

Key points

  • Standard HTTP proxies such as Burp never see HTTP/3 because it runs over QUIC on UDP, so a UDP-capable proxy and a MITM CA are required.
  • Chromium's SOCKS5 support is TCP-only, so browsers need redirection or flags such as ECH disabling and a forced-QUIC option to be intercepted.
  • One HTTP/3 request produced ten proxy history rows, since each QUIC stream chunk is recorded separately.
  • Requests using the QPACK dynamic table show only placeholders in the decoder and cannot be edited, so tests need a fresh connection or Replay with edits.
Smallest possible nopert polyhedra: the Septopertgithub.com · 49 points · 6 comments · 2dmathscienceThe Septopert is a 7-vertex convex polyhedron with integer coordinates that is proved not to be Rupert, meaning no hole can be cut through it for a copy of itself to pass.

Key points

  • Its seven integer vertices have coordinates up to 7 in absolute value, and it has C3v symmetry with 10 triangular faces.
  • The apex is a cube corner with three mutually orthogonal edges, and the axis projection has an exact Triforce alignment, both of which make the proof harder.
  • The author's Lean proof is checked by the kernel with standard axioms, plus a native program that validates about 732,000 table rows of certificate data.
  • A companion result shows no integer member of the family with smaller height is a nopert, covering 55,651 canonical tuples.
Spamhaus new email sender requirements in Novemberspamhaus.org · 48 points · 41 comments · 2hsecuritySpamhaus will require every sending IP to have a valid HELO, a reverse DNS (PTR) record, and a matching A record, rolled out gradually starting November 2026.

Key points

  • Reverse DNS must resolve to an FQDN that maps back to the same IP (forward-confirmed rDNS), and generic hosting-provider rDNS names are discouraged.
  • HELO/EHLO hostname must resolve to the sending IP and stay consistent across connections; localhost and placeholder values can trigger listing.
  • IPv6 sending allocations must be at least /64 with valid reverse DNS, and SPF must stay within the 10 DNS lookup limit.
  • Confirmed opt-in is a minimum requirement for all bulk mail, and purchased, rented, or harvested lists are named as the top source of spam trap hits.

What HN thinks

  • Commenters running self-hosted mail note that residential or dynamic IPs often lack a controllable PTR, and one suggests using the ISP-assigned reverse DNS name as a workaround.
  • Several commenters argue the HELO, PTR and A-record checks are easy to satisfy and will mainly burden small senders, while spam from large providers like Google, Amazon and Microsoft continues.
  • One commenter with first-hand experience reports Microsoft 365 blocked mail because the HELO lacked a fully qualified hostname, and that delisting took time after the fix, with Microsoft's SNDS offering a way to investigate.
The Miracle Cure: In 1977, a drug rejected by doctors became a sensationslate.com · 48 points · 40 comments · 4dhistoryhealthSlate's long feature recounts how laetrile, an apricot-pit compound rejected by the FDA, spread through 1970s state legislatures and led to a custody fight over leukemia patient Chad Green.

Key points

  • Twenty-seven states eventually legalized laetrile for some prescribing, even though the FDA controlled interstate commerce and refused to approve it.
  • The laetrile movement was organized by John Birch Society members and framed the debate as personal freedom, borrowing rhetoric from Roe v. Wade.
  • Chad's parents stopped his chemotherapy pills in secret, then fled to Tijuana after losing a court ruling; Chad died in October 1979 at age three.
  • A 1982 National Cancer Institute study again found laetrile produced no improvement in patients, and no more states passed laetrile bills afterward.

What HN thinks

  • Commenters widely contrast laetrile with chemotherapy, noting chemo is toxic but targets cancer, while cyanide-releasing laetrile is untargeted and dangerous.
  • Several commenters argue alternative cures persist because people want hope over honest odds, and distrust of established medicine makes them easy targets for sellers.
  • One commenter argues that accelerated approval lets many oncology drugs reach market without proven benefit at high prices, while others counter that delayed approval also has costs.
Doing Show HN generates a ton of automated email pitchespetervandijck.com · 46 points · 42 comments · 13hbusinessstartupsA Show HN author logged 17 unsolicited sales pitches in four days after posting, most generated from the HN title alone, selling backlinks, audits, sponsorships and listings.

Key points

  • Seven of the 17 were launch boards or directories, typically offering dofollow backlinks for prices from free to $49.
  • One directory pasted the whole Show HN title in as the product name, so it pitched a listing for "I built SQLite databases for your agents with an Airtable view for you."
  • One sender emailed the same pitch three times, including a Sunday resend claiming a mail-system glitch.
  • Several pitches included 'no follow-ups' promises that were then broken, and one 'AI editor' was quietly renamed to 'Editor'.

What HN thinks

  • Commenters broadly agree that posting any contact details publicly, including on Show HN, Hacker News comments, GitHub, or freelancer threads, draws a steady stream of unsolicited pitches, scams, and recruiter messages.
  • One commenter reported that the pitches now look targeted and AI-generated, referencing their comment or repository, which makes them harder to filter than older mass spam.
  • Practical mitigations raised include hiding email on GitHub, marking such mail as spam rather than replying with "no" (since replies validate the address), and filtering by patterns like introductions or "confidential" wording.
1900s How do you Wash a Kimono?oldphotosjapan.com · 44 points · 11 comments · 3dhistorycultureA photo-essay collection of Meiji-era (1890s-1900s) Japanese studio and street photographs documents how kimono and household laundry were washed and cared for, using albumen prints.

Key points

  • The article reproduces studio-posed albumen prints showing a washboard and starching, practices a commenter said they had not known about for early Japan.
  • Photos are staged in studios with models, so they show the method rather than documenting everyday practice exactly.
  • A reader's neighbor says a Tokyo-area river beside her home was once where kimono dyers washed fabric, before concrete banks replaced it.

What HN thinks

  • Commenters noted that the article's account of fully unpicking delicate kimono for washing matches how formal silk kimono are still cleaned today, and that they cannot go in a washing machine.
  • Some doubted the article's claim that no soap was used, since mixing lye with plant or food starches effectively produces soap.
  • Commenters pointed out that the labor involved (unpicking, restarching, resewing) and the lack of frequent washing were offset by layered undergarments that protected the outer silk.
A 512 KiB binary message teaching chess without a shared languageartemlegotin.com · 42 points · 12 comments · 9haiscienceGolden Board is a 512 KiB binary message, a 2,048×2,048 bit square, that teaches chess to an unknown recipient by first teaching its own bit, checksum and picture format.

Key points

  • The bootstrap strip teaches binary numbers, bit order, CRC-32C checksums and a record format before any chess content appears.
  • The payload holds five checked documents totaling 506,181 body bytes, including a 64-game anthology of 6,890 plies.
  • Moves are stored as 16-bit packed words (padding, origin, destination, promotion choice) paired with separate display patches.
  • The author tested the design only with LLMs as stand-in recipients, which cannot show an unknown intelligence would understand it.

What HN thinks

  • Commenters largely praised the project as a clever, well-executed take on messages for unknown intelligences, with one noting it evokes the film Contact.
  • Critics argued the design assumes shared conventions: one said an arrow is a culturally shared symbol, and another doubted an extraterrestrial could rederive CRC-32 without more context.
  • One commenter who tested it said a fresh LLM decoded the binary correctly in about ten minutes, though the author noted LLMs are imperfect stand-ins for an unknown recipient.
Exact Gradients Through a BRep Kernelmuni.works · 38 points · 14 comments · 3dprogrammingsystemsvcad, an open-source Rust BRep kernel, computes exact parameter gradients (including fillet radii) by freezing the tessellation topology and differentiating node positions instead of the kernel code.

Key points

  • Derivatives are exact for the frozen mesh; they differ from the continuum derivative by the polygonization error, 6.7% for a filleted cube at 16 segments per circle and 0.11% at 128.
  • A plan stays valid only for parameter steps of at least a fixed millimeter threshold; the optimizer captures a fresh plan at every iterate.
  • A topology change such as a blind hole breaking through returns an error rather than a derivative, and the optimizer shrinks the step.
  • On a five-parameter filleted, drilled box, a seam gradient takes 85 ms versus 860 ms for a rebuild finite-difference gradient.

What HN thinks

  • Commenters compare the approach to profile-guided optimization, framing it as optimizing geometry against physical or engineering objectives rather than machine code against runtime profiles.
  • One commenter explains the method as recording how mesh vertices connect after tessellation so only parameters affecting a target property need changing, avoiding costly rebuild-and-test loops over every parameter.
  • A recurring criticism is that the paper reads as LLM-generated prose, which one commenter feels makes the core idea harder to follow, though the author responded that the ideas themselves are what matter.
Caltech professor no longer grades homework's because everyone uses AIbusinessinsider.com · 37 points · 39 comments · 3haimathCaltech professor Yi Ni says he now grades college math homework only for completion, not accuracy, because students can complete it with AI and he cannot detect that use.

Key points

  • Ni has taught at Caltech for 17 years and says virtually all college math homework can be done with AI.
  • He cannot tell whether an assigned problem set was completed with AI, unlike AI-written essays, which he says are easier to spot.
  • Caltech has no unified guidance for math professors on AI, and some courses now label assignments "AI allowed."

What HN thinks

  • Several commenters argue that dropping homework grades is reasonable when exams are proctored and carry most of the weight, since the real purpose of homework is learning.
  • One commenter notes that removing the fear of a bad homework grade may let good students use assignments to learn, while the pressure of grading pushes students toward AI.
  • Another commenter suggests choosing learning objectives that AI cannot shortcut, such as open-ended project goals, rather than skills like a specific software tool.
Show HN: Coral – Open-source gateway where agents write their own toolsgithub.com · 37 points · 11 comments · 10haitoolsCoral is an open-source, self-hosted gateway where agents call tools by writing short TypeScript programs and save working programs as reusable custom tools with per-tool Allow, Ask, or Block policies.

Key points

  • Each program runs in a fresh V8 isolate with no network access, no credentials, and no imports; Coral checks every provider call against policy before adding credentials.
  • The vendor reports 34 GitHub, Linear, and Slack tasks with Opus 5.5 where custom tools cut tokens 37%, cost 31%, and time 29%.
  • Providers can be set to use each caller's own account, so teammates share tools without sharing API tokens, though each provider needs a one-off OAuth flow.
  • Runs as a desktop app for macOS, Windows, and Linux, or as a shared server needing an OIDC identity provider, a TLS reverse proxy, and SQLite or Postgres.

What HN thinks

  • One commenter noted the name Coral collides with Google's Coral platform and many other products, making the project hard to find in search.
  • Commenters are positive about granular per-tool permissions and see them as important as agents gain more autonomy and write access.
  • A commenter from a deploying team shared advice: start with the local desktop app before the shared server, which needs a bring-your-own identity provider, and let humans review agent-created tools rather than designing them up front.
Duolingo uses screen time permission to block uninstallblog.rhett.info · 36 points · 9 comments · 6htoolsdesignDuolingo's app-blocker feature, when enabled, also prevents uninstalling the Duolingo app unless the user enters their iOS Screen Time passcode, per a developer's report.

Key points

  • The blocker's first option is "All Apps and Categories", which would make the phone largely unusable until the daily lesson is completed.
  • After enabling the blocker, the author was prompted for a Screen Time password when trying to uninstall Duolingo, and a demo video is included.
  • The author says the lock persists even after the "block other apps" feature is disabled, so users without their Screen Time password cannot delete the app.
United Airlines has a widespread system outageapnews.com · 34 points · 6 comments · 1hbusinessUnited Airlines is experiencing a widespread system outage, according to an AP News report whose full text could not be retrieved.
Parcel: A desktop mail client for Linux, built with Zigtangled.org · 31 points · 7 comments · 24htoolsopen-sourceParcel is a Linux desktop mail client written in Zig that works with Gmail, Fastmail, and any JMAP server, showing mail as a single list filtered by stacked labels.

Key points

  • Mail is shown as one combined list filtered by stacked label filters, such as Inbox and Work, rather than per-folder views.
  • Search runs against a local index of your mail and answers most queries without contacting the server once indexing completes.
  • Gmail sign-in uses Parcel's bundled Google client in release builds, but Workspace admins may need to approve it or you can supply your own.
  • Sign-ins made with your own Google client expire every seven days while the Google Cloud project is in testing mode.

What HN thinks

  • One commenter said they would try it but prefer folders to labels, and asked whether the sending-from-an-alias feature is supported.
  • Only one comment was posted, so the thread offers little beyond the project page.
I Can't Believe It Can Sortyugen-tsuki.github.io · 31 points · 37 comments · 3dprogrammingscienceA nested-loop sort that swaps whenever A[i] < A[j] looks like it should fail, yet it correctly sorts an array, as shown in a 2023-era paper by Stanley Fung.

Key points

  • The whole algorithm is two nested loops over the array with a single compare-and-swap, running in O(n^2) time.
  • The pseudocode is taken verbatim from Stanley P. Y. Fung's paper, which proves the method sorts using three theorems.
  • The post includes a Java implementation that sorts a 14-element integer array in place.

What HN thinks

  • One commenter asked whether it is just bubble sort, and a reply said no, pointing to the paper's theorems as the proof of correctness.
Show HN: Fakemetrics.com – "build in public" like a superstarfakemetrics.com · 30 points · 10 comments · 6htoolsdesignFakemetrics.com is a web editor that generates screenshots of fabricated analytics dashboards for Stripe, Instagram, X, LinkedIn, Google Analytics, Substack, YouTube Studio and Shopify.

Key points

  • Users pick a platform, enter invented figures such as revenue, views or followers, and download a screenshot of the resulting fake dashboard.
  • The site's pitch is aimed at fake 'build in public' founder posts that show growth with no real revenue.
  • The author says the UI cloning system was built to reproduce pixel-identical interfaces from screenshots, which general models like Claude did not do well by default.

What HN thinks

  • Commenters largely expect misuse, with one arguing the site enables grifters to fake traction, though another notes faking screenshots was already possible with Photoshop or browser inspect tools.
  • Several commenters doubt anyone will pay for the Pro tier, while one counters that less technical grifters would likely be the paying audience.
  • One commenter is frustrated that legitimate projects struggle to gain visibility while fabricated metrics content reaches the front page.
Psychic: Like fzf but learns from your actionsklauria.tech · 30 points · 9 comments · 8htoolsaipsychic is a terminal file browser, like fzf, that retrains a small LightGBM ranking model on your own open and cd history each time it starts, so frequently used files rise to the top.

Key points

  • The model retrains from scratch on every launch and swaps in the new model only when the user is not looking at results, such as while an editor opens.
  • Early on, a simple heuristic scoring recent clicks, modification time and fuzzy match dominates, then the ML score's weight grows to nearly 1 as clicks accumulate.
  • The author reports 29% of launches needed no typing, 80% of picks were the top result, and the median time to destination was 2 seconds across 1450 opens.
  • LLMs and transformers were ruled out because a network or heavy model call would exceed the 10 ms latency budget, and linear models performed much worse than LightGBM.

What HN thinks

  • Commenters compared it to atuin, which applies similar fuzzy-plus-ML ranking to shell history with ctrl+r, and one said psychic picked up existing zoxide data on first run.
  • One commenter noted it only covers files and directories, not the wider range of fzf uses such as processes or containers, and the author said extending it needs non-file features.
How Should Armies Fight?worksinprogress.co · 30 points · 34 comments · 16hpolicyscienceWatling argues that sensors, weapon range and munition precision now make concentrated, surprise-based attacks nearly impossible, pushing armies toward dispersal, urban fighting, or attacking enemy sensors.

Key points

  • Acoustic sensors carried by a single soldier can identify a tank from ten kilometers away.
  • Heavy artillery missed by about 50 meters at 16 km eighty years ago, but misses by about 2 meters at 70 km today.
  • Ukrainian soldiers often live underground for months, hundreds of meters from the nearest friendly unit.
  • A Ukrainian battalion commander told Watling he needs buildings to fight from, not reinforcements.

What HN thinks

  • Commenters point to the Ukraine war as an attrition stalemate where neither side can achieve air superiority, and argue that cheap drones and missiles mean classical air dominance is largely obsolete.
  • One commenter argues that the shift toward attrition raises the cost of war sharply, since winning now requires enduring heavy suffering on both sides rather than a quick bombing campaign.
  • Another commenter notes that cheap improvised drones, such as small FPV quadcopters carrying explosives, now require very little money to build, lowering the barrier to effective attacks.
Making Atuin sync 32x faster with packfilesblog.atuin.sh · 30 points · 12 comments · 5dtoolssecurityAtuin 18.20.0 makes shell-history sync up to 32x faster for Atuin Hub users by bundling encrypted records into client-built packfiles, with 16x gains for self-hosted servers.

Key points

  • Because the server cannot decrypt data, the client packs records: it zstd-compresses and encrypts batches of about 1000 records into packfiles of roughly 10-100 KB.
  • Pipelining parallel page downloads alone improved sync speed by about 340 percent, and batched SQL inserts were also added.
  • The team reports the 100K-record power-user case improved about 30x, with packfiles contributing roughly 2x of that.
  • The protocol version was not changed, and packfiles are only supported on Atuin Hub so far, with self-hosted servers to follow once stable.

What HN thinks

  • Commenters split on the optimization itself: one argues a 10x-plus speedup suggests the original design was a mistake, while others counter that starting simple and optimizing once sync became a real bottleneck is sound practice.
  • Several users say they have used Atuin for years without noticing slow sync, suggesting the change matters more for scale than for typical personal use.
  • One commenter misread the project as using a blockchain; another clarified that sync runs through a centralized server that users can self-host or disable.
First phase of embedded assessments at OpenAIopenai.com · 29 points · 11 comments · 8haipolicyOpenAI published a post about the first phase of its embedded assessments, but the article could not be fetched (HTTP 403), so its contents are unknown.

What HN thinks

  • Commenters frame OpenAI's move as a response to the competitive pressure from Anthropic, with some arguing it reflects a shift toward safety-focused caution and others reading it as liability management.
  • Several commenters are skeptical that the assessment is more than a hobbled external audit, though one argues it gives auditors more access to internal tooling and resources than a typical external review.
  • One commenter questions whether a safety-first culture at OpenAI is a real differentiator, arguing the company's earlier techno-optimism was its advantage.

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

Ask HN: If AI tools are so good, why isn't your bug backlog empty?27 points · 32 comments · 6haiprogrammingAn Ask HN poster questions why AI coding tools have not emptied engineering bug backlogs, since they would imply teams have run out of work.

Key points

  • The poster argues that if AI truly replaced programmers, teams would be out of work and bottlenecked only on business decisions, which they are not.

What HN thinks

  • Commenters split: some say AI only speeds work when agents run unsupervised, which most teams do not trust yet, while others report their backlogs are now empty.
  • One commenter says AI bug fixes tend to patch symptoms rather than eliminate classes of bugs, because agents lack project-wide context across sessions.
  • Several commenters point to a shift toward rewriting products from scratch with AI, with one reporting this never produced a v1.0 release.
Giant Roadside Attractions Face Mounting Costs and an Uncertain Futurenytimes.com · 27 points · 39 comments · 6dculturebusinessTyra, an 86-foot fiberglass T. rex in Drumheller, Alberta, faces closure when her site's land lease expires in 2029, because the local chamber of commerce says maintenance costs are unsustainable.

Key points

  • The Drumheller and District Chamber of Commerce decided last year to close the site rather than keep paying for upkeep.
  • Tyra has been a draw for about 25 years, with visitors climbing 106 steps to look out through her fiberglass teeth.
  • The article says thousands of roadside sculptures across North America and Australia face similar aging and cost problems.

What HN thinks

  • Commenters questioned the article's framing, noting that two cited attractions roughly broke even, so the stated problem of unsustainability is not clearly demonstrated.
  • Several commenters argued that roadside attractions mainly serve to draw drivers off highways into towns, and one suggested that breaking even while doing so may already be a success.
  • One commenter said the real change is not visual access but that drivers now prefer staying in the car on phones over stopping to stretch and engage locally.
US will say "no" to global ban on killer robotsnewsweek.com · 26 points · 12 comments · 47mpolicyaiThe US is expected to oppose negotiations on a legally binding treaty restricting autonomous weapons at a UN conference in Geneva, instead proposing five more years of nonbinding work.

Key points

  • The Seventh CCW Review Conference runs November 16 to 20 in Geneva and is not expected to produce a treaty that week.
  • The US working paper dated November 4 argues existing international humanitarian law already covers autonomous weapons and that new prohibitions are unnecessary.
  • Stop Killer Robots says 130 countries support a legally binding instrument, while about a dozen, including the US, Israel and Russia, oppose one.
  • The US argues nonbinding rules could be updated faster than treaties and would avoid separate legal regimes for countries that ratify a treaty.

What HN thinks

  • Commenters largely read the US position as diplomatic stalling, noting that existing humanitarian law is already claimed to cover autonomous weapons, which they see as a reason a new treaty is needed.
  • Several commenters expect US opposition to follow from its broader disregard for UN treaties, and one pointed to the Ottawa landmine treaty as an example of a ban that still lacks major powers.
  • One commenter argued that autonomous weapons are an extension of long-standing guided munitions such as cruise and ballistic missiles, so the line between existing and new weapons is blurry.
Linux was created as a rejection of robust software architecture and toolstwitter.com · 26 points · 19 comments · 20hsystemsopen-sourceA GrapheneOS tweet claims Linux was created as a rejection of robust software architecture and tools; the post itself was not retrievable, so its specific argument is unknown.

What HN thinks

  • Several commenters argue Linux's monolithic design was a pragmatic choice in 1991, when microkernel IPC overhead was too costly for available hardware, and that the same tradeoff still shapes the industry.
  • One commenter says security that costs speed and convenience tends to lose in practice, so safer designs like microkernels struggle to win adoption against faster, fully featured kernels.
  • A commenter criticizes the GrapheneOS post as one-dimensional, prioritizing security over all other considerations and not convincingly evaluating the technical tradeoffs involved.

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

US now "comfortable" with children dying of measles as death toll hits 10arstechnica.com · 25 points · 9 comments · 7hhealthpolicyUS loses measles elimination status as PAHO panel finds a continuous transmission chain of 12+ months, with 10 confirmed measles deaths in 2026 and nearly 5,500 cases across 47 states.

Key points

  • Lost status is reported via Wall Street Journal sources; PAHO was set to officially announce its conclusion on November 17.
  • The US earned elimination status in 2000; loss requires a chain of transmission of an identical genotype and lineage lasting 12 months or more.
  • At least four of the 10 deaths were infants, including two in Pennsylvania, one in New York and one in Wisconsin.
  • Pennsylvania alone has over 1,500 cases, and the CDC reportedly removed measles death data and skipped press briefings.

What HN thinks

  • Commenters largely framed the deaths as a consequence of falling vaccination coverage and criticized the handling by health leadership.
  • One commenter noted that measles is small compared with other childhood causes of death, but warned that losing other vaccinations could make things much worse.
  • Some commenters argued that unvaccinated children should be treated as child neglect, and that herd immunity loss exposes vaccinated people too.
Launch HN: Axonnic (YC F26) – Sleeves that track fingers and estimate force24 points · 8 comments · 3hhardwareaiAxonnic sells a 64-channel forearm EMG sleeve that decodes finger movement and estimates fingertip force to capture manual-skill data for training dextrous robots.

Key points

  • Claims surface EMG captures muscle effort that video misses, such as the grip maintenance involved in sewing and delicate assembly.
  • Reports 6.9 degrees mean error across 20 joint angles against Apple Vision Pro, on a 215-second recording taken four days after training with no recalibration.
  • Electrode positions shift each time the sleeve is re-donned, so the company tests across separate wearings.
  • Factory recordings suffered wireless interference and unreliable off-the-shelf hardware, so the team is building a wired PCB design.

What HN thinks

  • One commenter asked how the model handles EMG baseline drift after an hour of sustained gripping, a question the post does not address.
  • A commenter objected that workers would not accept wearing such a sleeve to have their work tracked.
I posted in "Who wants to be hired?" and a recruiter sent me a RAT and a stealergetjump.me · 24 points · 9 comments · 13hsecuritywebA fake recruiter's VS Code take-home extension registers the machine, then runs server-supplied code that finds its command server by reading a public Ethereum wallet's latest transaction.

Key points

  • The extension ships no malware; the server returns a task whose executed field grew from 150 to 8,665 bytes after a device registered, while the displayed field stayed the same.
  • The loader decodes the IPv4 server address from the first eight bytes of a zero-value transaction's recipient, so no registrar can take the pointer down.
  • Three pointer wallets sent 1,157 pings to 11 servers from July to November, with the newest ones at hosting company Virtuo.
  • The payload is a Python stealer for browser data, keychains, Git and cloud credentials and wallet files, plus a remote-access tool that survives uninstalling the extension.

What HN thinks

  • Commenters criticized the post's LLM-generated prose as hard to read, and one questioned whether the author wrote it at all.
  • One commenter asked where the RAT was, since the author described the extension itself as clean, which the post answers only by pointing to a second-stage loader.
Generative AI Usage in the Games Industrygameworkers.co.uk · 23 points · 37 comments · 8haibusinessA coalition of European game worker unions publishes a manifesto calling for workers to avoid generative AI, for employers to ban it from development, and for regulators to forbid training on workers' work without consent.

Key points

  • The manifesto is signed by unions and work councils from the UK, Ireland, Spain and Milan, and covers both shipped assets and internal development materials like concept art and pre-translations.
  • It argues that AI-driven productivity gains tend to be short-lived, citing a study where workers frequently adopted systematically incorrect AI outputs under time pressure.
  • It says workers have been forced to train the models that replaced them, and asks that any use of staff work as training data require explicit consent and fair pay.
  • It also argues that employers blame workers for AI errors they mandated, and that vendors obscure costs and limitations of their tools.

What HN thinks

  • Commenters note the manifesto does not clearly address frame generation tools like DLSS, FSR, or XeSS, and it is unclear how players 'experience' code as an asset.
  • Several commenters argue that player opposition to AI is weak in practice, since many people still buy games that used AI in programming or development, which limits its leverage on publishers.
  • One commenter argues the deeper problem in games is overwork and exploitation of eager entry-level workers, which predates generative AI.
The origins of the placentacaltech.edu · 23 points · 5 comments · 3dsciencehealthCaltech researchers found that a few mural cells from the opposite side of the mouse embryo migrate to help build the placenta, overturning the textbook view that it comes only from one region.

Key points

  • Mapping used single-cell gene activity from more than 22,000 cells across six stages, combined with live filming of implanting mouse embryos.
  • Only two to four border mural cells per embryo migrate upward to join the ectoplacental cone, and their descendants stay on its outer edge.
  • Removing most mural cells stopped the early placental structure from forming in lab-grown embryos, and very few embryos developed in the uterus.
  • Filmed human blastocysts showed similar border cells, and published human datasets contain cells with the same signature among invasive placental cells.

What HN thinks

  • One commenter noted that the stem-cell embryo models failing to implant could be fixed if this finding holds, calling it potentially more significant than current market turbulence.
  • A commenter pointed to a proposed link between invasive viruses infecting germline cells and the evolution of the placenta, a separate hypothesis from this study.
Some OLED MacBook Pro Owners Report Eye Strain and Dizzinessmacrumors.com · 19 points · 1 comment · 5hhardwarehealthSome owners of the new OLED MacBook Pro report eye strain, headaches and dizziness, which they link to the screen's PWM brightness dimming.

Key points

  • Notebookcheck measured 240Hz flickering at brightness levels of 99% or lower on the 14-inch OLED model, with 56% amplitude.
  • The mini-LED MacBook Pro measured at 14,880Hz, a rate generally less likely to bother PWM-sensitive users.
  • Apple offers a Display Pulse Smoothing option on supported iPhones for low-brightness dimming, but macOS has no equivalent control.
  • OLED MacBook Pro models start at $2,899 for 14-inch and $3,399 for 16-inch, while Apple still sells mini-LED models.

What HN thinks

  • Commenters largely dismissed the complaints as overstated or unrelated to the screen, with one joking that the receipt matters more than the display.
  • One commenter with bad eyesight argued mini-LED is a good middle ground between LCD and OLED.