HN Brief: 2026-08-26
Today's Hacker News was dominated by a mass of Apple hardware news—three separate threads on new Macs and chips—but the real energy was in the backlash, with commenters arguing the M6 era marks the end of affordable Macs and picking apart Apple's AI strategy as a convenient rewrite of history. The Dolly Parton death thread became an unusually warm and substantive appreciation of a celebrity people genuinely admired, while the Nitter/XCancel C&D drama split the room between legal pragmatists and principled open-web advocates. Underneath it all, the Qwen model release and the "how much of HN is AI?" meta-post kept the ambient debate about AI dominance humming.
Threads worth clicking: “Dolly Parton has died” for a rare celebrity death thread that’s actually about craft and character, not hot takes; “Nitter and XCancel receive cease and desist notices” for the sharpest fight on the board about whether open-source tools have a duty to fight or fold; “How much of HN is AI?” for the most pointed self-examination of the community’s own obsession; “Building a backyard office, the build and cost breakdown” for a satisfyingly detailed technical tangent that accidentally became a zoning and insulation debate; “C2PA Cameras Do Not Survive Contact with Reality” for a practical, demo-driven takedown of a crypto-provenance scheme that doesn’t work on rooted phones.
Dolly Parton has died [comments]
1428 points · 217 comments · www.theguardian.com · 14h ago
The Guardian reports that Dolly Parton has died at 80 after a brief battle with cancer, covering her legendary songwriting career, her philanthropy like the Imagination Library that's sent over 300 million books to children, and her rise from poverty to become a global icon. The HN thread largely bypassed the music obituary details and dove straight into personal appreciation, with multiple people who aren't country fans calling her an unequivocal good human and a rare example of a celebrity who actually used her wealth to lift up communities. A substantial portion of the discussion focused on the mechanics of her guitar work—someone pointed out that despite the famous "Jolene" riff being written by Parton, the actual recording was played by session musician Chip Young because her long fingernails made it hard to nail consistently, which sparked a whole tangent on her open tunings and technical skill. The Imagination Library got heavy praise from parents around the world, including someone in London who got books monthly for their daughter, and there was a brief counterpoint where someone posted academic dissent about the program, which another commenter dismissed as people who "just have to find something wrong with everything." The general consensus was that Parton's legacy as a songwriter and performer is secure, but what made the thread feel different from a typical celebrity death was the sheer volume of people saying she was one of the few famous figures they'd never heard a single negative thing about.
Apple introduces M6 and M5 Ultra [comments]
1107 points · 1060 comments · www.apple.com · 19h ago
Apple’s press release announces two new chips: the M6, a 2nm part for the Mac mini with a 12-core CPU and dual neural engines, and the M5 Ultra, a quad-die behemoth for the Mac Studio with up to 512GB of unified memory and 1.2TB/s of bandwidth aimed at running huge local AI models. The thread largely skipped the spec-sheet excitement to argue that Apple’s hardware-first, software-late strategy was the right call, with several people in AI saying local models on your own machine are the real future and that Apple let everyone else burn cash on datacenters before cleaning up. That consensus got pushback from those who pointed out Apple did pour enormous resources into AI software with Apple Intelligence and mostly failed, so “smart sideline strategy” is a convenient rewrite of history. A big secondary argument was about whether these machines are actually practical for AI work: the 1.2TB/s bandwidth is roughly two-thirds of an RTX 5090’s, and while a 512GB memory pool is unmatched, Linux support is still stuck at M2 under Asahi, so the lack of GPU access outside macOS is a real dealbreaker for many. There was also a sharp exchange about RAM pricing and soldered components, with someone noting they had 64GB in a laptop ten years ago, so the “we’re finally getting enough memory” narrative doesn’t hold for power users.
Nitter and XCancel receive cease and desist notices [comments]
863 points · 710 comments · github.com · 14h ago
The Nitter project, which provides a privacy-respecting front-end for browsing X/Twitter without ads or tracking, has received cease and desist letters demanding the permanent takedown of its main instance, GitHub repository, and all stored X data. Hacker News split into two main camps: one side immediately jumped into the legal weeds, arguing that a C&D carries no actual legal weight and that the developer should ignore it or at least wait for proper legal advice before shutting down, while the other side pointed out that fighting this would cost tens of thousands of dollars and the "for the foreseeable future" language from the project maintainer suggests they're complying out of caution rather than conviction. A recurring pushback came from people who couldn't square calling X a "Nazi app" while also being desperate for Nitter to stay up—others countered that journalists, government agencies, and public transit systems still post exclusively to X, making some kind of read-only access a practical necessity rather than an endorsement. The thread also devolved slightly into a familiar HN tangent about whether "free as in speech" or "libre/gratis" is the better terminology, and a few people floated the idea that Elon Musk might be deploying legal threats strategically against small open-source projects he knows can't fight back.
New Mac Studio with M5 Max and M5 Ultra [comments]
758 points · 499 comments · www.apple.com · 19h ago
Apple announced the new Mac Studio with M5 Max and M5 Ultra, positioning it as the ultimate desktop for on-device AI with up to 4.3x faster AI performance, up to 512GB of unified memory, and Thunderbolt 5 clustering for distributed inference. The conversation quickly fragmented into price shock—the top configs push past $5,000 in the US and over €6,600 in Europe, with one person noting a fully loaded IBM PC AT from 1984 would cost about $19k today, so maybe this isn't that crazy. A big split emerged around whether Apple's 512GB ceiling is a strategic blunder or a supply-chain reality, with some pointing out the RAM shortages and Micron's accusation that Apple squeezed margins for years instead of booking fab capacity. The unified memory architecture drew technical debate: it's great for running 70B+ parameter LLMs locally, but commenters caveated that LPDDR5X is slower than dedicated GPU VRAM, and that low-level latency and cache behavior matter more than raw bandwidth numbers when comparing against x86 setups. A surprising tangent was the crowd of people arguing that a Mac Studio plus a cheap laptop (like the Framework Neo) is now a smarter move than a single maxed-out MacBook Pro that lives docked, especially with remote work tools making portability less urgent.
My Friend Aaron [comments]
539 points · 146 comments · rorz.io · 15h ago
This is a long-form short story about the narrator's friend Aaron, a brilliant but directionless guy with an addictive personality who cycles through obsessions—World of Warcraft, prediction markets, hackathons—before building a platform where people can bet on the minutiae of his own daily life, which he then rigs for profit. The HN thread was overwhelmingly positive, with the author himself chiming in to say getting to the front page was better than any writing contest, and many commenters admitting they read more than half the story before realizing it was fiction, not a memoir. The big split in the comments wasn’t about the story’s quality, but about the character of Aaron—half the thread recognized him as a painfully real type they’ve known (the gifted schemer who avoids real work), while others argued he was actually just "INT 18 WIS 3" and not particularly smart at all. A darker tangent emerged about prediction markets inevitably becoming assassination markets, with people pointing out that betting on an individual’s actions is functionally indistinguishable from offering a bounty on them, and that nobody has figured out how to regulate that.
New Mac mini, featuring M6 and M5 Pro [comments]
487 points · 298 comments · www.apple.com · 18h ago
Apple announced the new Mac mini with M6 and M5 Pro chips, claiming huge leaps in AI and graphics performance. The HN thread largely ignored the performance claims and zeroed in on the price hike—the base model jumped from $599 (M4) to $899, which stung worse in Europe at over €1000. A lot of people who grabbed the M4 at its low point feel like they caught the last cheap train, and the consensus is that the era of an affordable little Mac is over, driven by memory and component costs. Several commenters noted Apple’s decision to stick with 16GB of base RAM feels stingy given the sticker shock, though some pushed back that it’s just the new pricing reality. There was also a side tangent that Apple hyped Cyberpunk 2077 performance in the press release, which led to the usual resigned shrug about whether Apple will ever seriously court gaming on the Mac.
OpenAI Jalapeño: Better than Nvidia Blackwell [comments]
454 points · 291 comments · newsletter.semianalysis.com · 17h ago
The article claims OpenAI has built "Jalapeño," a custom inference chip that beats Nvidia's Blackwell and even its successor Rubin on performance-per-watt, using a fast 16-month design cycle and a generalized architecture rather than being locked to OpenAI's own models. The HN thread immediately split on whether this is genuinely impressive or just the obvious result of building a narrow ASIC against a general-purpose GPU—plenty of people pointed out that custom chips always win on efficiency for a fixed workload, and that comparing a new, specialized chip to a mature general-purpose product is an unfair benchmark. A large chunk of the discussion pivoted to whether plummeting token costs are actually coming, with one side arguing Jevons paradox means cheaper tokens just explode total energy use and data center construction, while another side pushed back hard that token prices have actually *risen* due to thinking models burning more compute per query. A few commenters also zeroed in on the memory supply chain, questioning why OpenAI doesn't build its own RAM and getting educated on just how hard and patent-laden DRAM manufacturing is, while others found the FP4 precision figure almost laughable compared to the precision standards of just a decade ago.
FDA authorizes first wearable device that monitors ketone and blood sugar levels [comments]
372 points · 184 comments · www.fda.gov · 12h ago
The FDA just authorized Abbott’s Libre Duo, the first wearable that continuously monitors both ketones and blood glucose in a single sensor, aimed at people with diabetes who risk diabetic ketoacidosis. HN immediately ran with the idea that marathoners, ultrarunners, and biohackers will start using it for performance and metabolic flexibility, but a bunch of commenters pushed back hard: high-level endurance athletes are already cramming as many carbs as they can stomach, so real-time glucose and ketone data won’t give them actionable information, and a fit cyclist can actually appear glucose intolerant on a CGM anyway. The thread also split over whether non-diabetics get any benefit—some argued steady blood sugar aids recovery and longevity, while others pointed to a clinical trial that was aborted after aggressive glucose control caused cardiac events, and a few noted that non-invasive optical sensors have historically been “racist,” failing on darker skin tones. Another tangent dug into lactate: insiders claimed the sensor chemistry can secretly measure lactate, but domain experts corrected that and noted continuous lactate sensors are already banned in cycling competition as a form of doping.
Firefox 157 will include JPEG XL by default on all platforms [comments]
360 points · 92 comments · groups.google.com · 14h ago
Firefox 157 will ship JPEG XL decoding enabled by default across all platforms, marking a major milestone for the next-generation image format. The HN discussion dove deep into what actually makes JXL special, with the strongest consensus forming around its ability to losslessly repackage existing JPEGs into smaller files without re-encoding—a feature neither AVIF nor WebP can match. People went back and forth on real-world performance: some reported noticeably slower decoding on macOS compared to plain JPEG, while others pointed to the Rust library's recent multithreaded improvements closing that gap. A recurring friction point was that while JXL's progressive decoding and massive dimension limits (up to 1M pixels per axis) sound great on paper, the format still faces a chicken-and-egg problem with file explorer support, image editors, and web upload forms—Chrome is finally considering shipping after Apple already did, but nobody expects Firefox alone to move the market.
Don't Wordle [comments]
344 points · 121 comments · dontwordle.com · 20h ago
Don't Wordle is a daily word game that flips the premise: you get six tries to *avoid* guessing the hidden word, using the same green/yellow/gray feedback to narrow possibilities while keeping as many valid words alive as possible. HN mostly found it brutally difficult, with many reporting they accidentally won on the first few tries by playing normally, and a few strategists dug into the math, showing that optimal play involves using weird, uncommon letters and avoiding words that eliminate too many options. The comments also split sharply on the game's aggressive cookie consent popup—some called it a dealbreaker with hundreds of "legitimate interest" toggles and no reject-all button, while others noted that ad blockers handled it fine. A minor debate emerged over whether the rules correctly handle repeated letters (e.g., guessing MAMMA reveals whether there's a second A), and someone pointed out the risk of NYT lawyers coming after the name, though mechanics themselves aren't copyrightable.
Building a backyard office, the build and cost breakdown [comments]
333 points · 211 comments · www.imkylelambert.com · 17h ago
This is a detailed cost breakdown and walkthrough from a Portland homeowner who built an 8x10-foot backyard office using a Tuff Shed frame, coming in under his $20k budget after deciding against a prefab Autonomous pod that would have cost closer to $30k. The HN thread mostly agreed that a physically separate workspace is a game-changer for remote workers with kids, with several people sharing their own successful shed-to-office conversions and praising the mental shift from walking outside to "go to work." The pushback wasn't really about the build itself but about the premise: some argued you could just live near your office instead, or that spending $20k on a shed is wild when co-working spaces exist, and one commenter got roasted for suggesting remote workers are likely to be replaced by AI before the shed pays off. A smaller but sharp technical debate broke out in the weeds—someone pointed out the posted photos showed R-19 insulation crammed into 2x4 studs, which actually performs worse than proper R-13, and others jumped in on whether you need fiber or copper for the network run and how to handle electrical grounding.
Qwen 3.8-Flash-Next releasing tomorrow (125B a6B) [comments]
319 points · 150 comments · modelscope.cn · 20h ago
The linked article wasn't available to this summarizer; from the discussion, Qwen is dropping a new 125B-parameter MoE model with only 6B active per token tomorrow, alongside a claimed 51B n-gram embedding supplement and a redesigned sparse attention mechanism, all framed as a preview for the Qwen4 family. The big immediate split is local hardware: people with 128GB Macs and Strix Halo boxes are stoked, while 5090 owners are disappointed it’s not the 35B-A3B compact model they were hoping for, though some point out the 27B from last week already runs great on that GPU. A few folks with hands-on time are pushing back on the hype, noting the model’s tendency to overthink and burn test-time compute, but arguing that’s by design and you can mitigate it with temperature settings. There’s also a side argument about inference speed claims—specifically 50-70 tok/s on an M5 Max—being misleading without mentioning multi-token prediction tricks like MTP or specific quantizations. The thread drifts into a broader complaint about OpenRouter’s flaky Qwen capacity and poor prompt-cache hit rates, with people recommending self-hosted routers like LiteLLM or just pinning providers directly.
Bomb fishing is wreaking havoc on Indonesia's coral reefs [comments]
308 points · 160 comments · e360.yale.edu · 17h ago
An article from Yale Environment 360 details how fishers off Sulawesi, Indonesia, are detonating over 8,000 underwater explosives annually to stun and kill fish, turning biodiverse coral reefs into rubble within a 90-foot radius. The discussion immediately split on whether poverty drives this—the article explicitly states the cost of boats and explosives makes it a middle-income practice, not a subsistence one, but many in the thread argued that colonialism and global economic exploitation are the root causes, while others pushed back hard, calling that framing a racist denial of the fishers' agency and morality. A few people did the math on the reported damage ("three football fields per year") and found it surprisingly low, questioning whether the bombs are actually aimed at the reef or just near it, though others noted that surviving coral is what attracts the fish. There was also a sharp undercurrent of finger-pointing at Western consumers who eat fish, with the counterargument that no one in global north tech is buying bomb-caught fish, plus a sarcastic side tangent about the absurdity of calling plastic-bottle explosives an "unintended consequence" of anything.
Starbase, LA [comments]
268 points · 485 comments · www.spacex.com · 15h ago
The linked article wasn't available to this summarizer; from the discussion, SpaceX announced plans for a massive Starbase facility in Louisiana, promising thousands of jobs and huge capital investment. The thread immediately split into two camps: one arguing this would be a genuine economic boon for one of the poorest regions in the country, bringing decades of construction work to local tradespeople, and another dismissing it as more corporate exploitation of a state with weak environmental and tax regulations. A significant chunk of the pushback centered on Louisiana's history with the chemical industry—the argument that enormous facilities there enrich a few while leaving residents poor and poisoned, with rising sea levels threatening the very coast the site would sit on. Others took a more cynical view of Musk's history, pointing to unpaid contractor liens and wondering aloud whether Optimus robots (or just cheap, non-union labor) would actually get the work rather than local welders. A side tangent dug into whether SpaceX's methane-fueled rockets could ever be carbon-neutral through direct air capture, countering the obvious point that thousands of launches a year would accelerate the sea-level rise threatening Louisiana.
How much of HN is AI? [comments]
261 points · 317 comments · blog.coredump.cx · 16h ago
The post is a survey of how much of HN's front page is AI-related or AI-written, using the Pangram detector, with the author claiming AI went from 40% of the daily top stories in February 2026 to 50-60% by June. The discussion split hard between people arguing that share is appropriate — AI is the biggest tech story in two decades, so of course it dominates — and people reading it as the crypto hype cycle reborn, same grifters, same messianic promises, same enormous sums of money changing hands. The sharpest pushback targeted Pangram itself: one side cited third-party evals claiming 90%+ accuracy and near-zero false positives, while the other side noted those evals are "collaborations" with Pangram and that AI detectors as a category don't reliably work. That spilled into a broader fight over whether AI is genuinely transformative or just being force-fed into every tool via bundling and free tiers, with one memorable comparison being that removing AI today would barely register while removing Linux would wipe out most of the internet. There was also a nostalgic sidenote about March 11, 2009, the one day the entire front page was Erlang stories.
Black hole singularity is a surface not a point [comments]
231 points · 171 comments · arxiv.org · 15h ago
The linked paper argues that the singularity at the center of a black hole is a surface, not a point, and that different observers falling in along different paths lose contact with each other well before reaching it. The Hacker News thread quickly pivoted to a vigorous debate over common pop-science explanations, particularly the widely-repeated claim that "space and time switch places" inside the event horizon. Several people with GR backgrounds pushed back hard, calling that phrasing a misleading artifact of Schwarzschild coordinates, not an invariant physical fact—though others defended it as useful shorthand for the idea that past the horizon, every direction leads to the singularity and you can no more avoid it than you can avoid next Tuesday. A deeper split emerged over whether the singularity itself is best thought of as a moment in time (a spacelike surface in the future) or still a point in space that you are simply forced to hit, with the more mathematically-minded insisting the invariant description is that it's a spacelike line in spacetime, not a place you can point to.
XCancel also down, at least for time being
229 points · 15 comments · news.ycombinator.com · 10h ago
An X/Twitter alternative front-end service called XCancel has been shut down after receiving a cease-and-desist letter from X Corp., with the site posting a notice that it's stopped until further notice while seeking legal advice. Most of the discussion quickly focused on how this is directly connected to the broader Nitter ecosystem collapse, with people lamenting the loss of a way to view X accounts without creating an account themselves. A sharp split emerged between those who refuse to use X on principle, citing ethical objections to Elon Musk's politics and the platform's influence, and others who dismissed that as "virtue signaling" or excessive emotional attachment to a personality. Several commenters pushed back hard on the "it's just a service" framing, arguing that these platforms are non-fungible—people post content there that doesn't exist anywhere else, and losing read-only access hurts users regardless of their feelings about Musk. The thread also floated LibRedirect as a still-working workaround for redirecting X links to whatever front-ends remain functional.
Why some US restaurants are banning tips [comments]
191 points · 392 comments · www.bbc.com · 21h ago
The BBC article looks at a handful of US restaurants that have banned tipping in favor of higher menu prices and better wages for all staff, including back-of-house, arguing it eliminates bias and gives workers predictable income. The HN discussion largely ignored the article’s examples and instead turned into a full-blown rant about the creep of tip prompts into every transaction—retail stores, airport self-checkout kiosks, even POS terminals in Europe where tipping isn’t customary. There was a clear split between people who see no-tip models as the only honest pricing and servers who make far more with tips than any reasonable hourly wage, with someone pointing out that the waitstaff at South Park’s no-tip restaurant demanded a return to tipping. A lot of commenters pushed back on hidden fees and service charges as being even worse than tipping, arguing that the real issue is that the sticker price should be the final price, full stop, with Europeans chiming in to say they already have that system and it works fine. The consensus was that tipping fatigue is real, but the economic and cultural inertia—plus the fact that tipped workers often fight to keep the system—means it isn't going away anytime soon.
Python's pre-declared constants are kinda weird [comments]
172 points · 171 comments · sebsite.pw · 10h ago
The article digs into the weird inconsistencies among Python’s six pre-declared constants—True, False, and None are lexical tokens (so `x.True` is a syntax error), while `__debug__` is an identifier you can’t assign to, and Ellipsis and NotImplemented are just normal builtins that can be shadowed. The thread immediately zeroed in on `__debug__` being the closest thing Python has to conditional compilation, since `if __debug__:` blocks are stripped from bytecode under `-O`, and the compiler has to forbid assignment to keep that invariant sound. That sparked a broader argument about Python’s age and baggage: some commenters compared it unfavorably to PHP’s recent evolution, claiming the Python community has refused to fix fundamental issues, while others reeled off a list of improvements (typing, GIL removal, pattern matching, packaging via uv) and called the criticism overblown. A separate mini-debate erupted over whether Python is actually beginner-friendly, with one camp arguing that Scheme or Haskell would be simpler to learn, only to get pushback about toolchain pain and confusing error messages.
Tooltips need a delay, and then they need to skip it [comments]
166 points · 44 comments · blog.master.dev · 15h ago
The article describes a tooltip pattern where a 200ms delay is used on first hover, then a 300ms "warm" window allows instant tooltips on subsequent hovers before cooling back down. Hacker News immediately recognized this as a rediscovery of UX patterns from the 1990s — Windows 95’s taskbar did it, Mac System 6 had it, and Amazon’s mega-menu used the same hysteresis trick. Several commenters argued that the no-delay version is actually better, calling the delayed behavior unpredictable and bug-feeling. Others pointed out that modern web frameworks keep reinventing wheels that native desktop UI solved decades ago, and that the real problem is the lack of a mature, ergonomic UI toolkit for the web. A few threads went into the weeds on mouse velocity, touchscreen issues, and the pure CSS alternative linked in the article.
How credit card rewards became a $9.2B wealth transfer [comments]
164 points · 300 comments · www.library.hbs.edu · 20h ago
The article breaks down Harvard Business School research showing that credit card rewards, funded by merchant interchange fees baked into everyone's prices, effectively transfer $9.2 billion a year from lower-income cash and debit users to wealthy cardholders—a regressive system the researchers compare in scale to major government benefits. The HN crowd dove straight into the math and mechanics, with many calling the "26% higher sales tax" comparison intentionally misleading, arguing it inflates the impact by framing it as a percentage of sales tax rather than total cost. A big split emerged between people who see the fee as an unavoidable cost of convenience and fraud protection (especially the liability difference between credit and debit) and those pushing for a behavioral fix, pointing to merchants like T-Mobile and Walmart that already offer real discounts for ACH or FedNow payments to sidestep the card networks. Some commenters pushed back hard on the idea that surcharges or cash-back incentives would change behavior, noting that credit card optimizers are a minority and most people won't chase 1-3% savings. The thread also got into the Durbin Amendment's unintended consequences—how capping debit fees pushed banks to juice premium card rewards, making the poor subsidize the rich even more—and a few people argued the whole system is just another version of "it's expensive to be poor," with means-tested benefits creating similar traps.
Run OpenBSD on DigitalOcean for $4/month [comments]
159 points · 72 comments · nil.wallyjones.com · 14h ago
The article walks through manually installing OpenBSD on a $4/month DigitalOcean droplet by uploading a custom miniroot image and setting up httpd and Let's Encrypt from the base system. The HN thread quickly turned into a referral marketplace for cheaper or more BSD-friendly alternatives—OpenBSD.Amsterdam gets multiple shoutouts for running OpenBSD VMs on OpenBSD's own hypervisor and donating back to the project, while Vultr and netcup.de are cited as well-trodden paths with better pricing or support. Several people warned that DigitalOcean's cheapest tier is actually a bad deal compared to indie hosts like RackNerd or BuyVM, where you can get 2GB RAM for roughly the same money, and a few threads dug into specific gripes: DO won't handle DDoS attacks, drops UDP datagrams needed for IPsec, and lacks a serial console for recovery. There's also a split on whether the $4/month price tag is worth it at all—some argue you can go completely free with Cloudflare Tunnels on a Raspberry Pi, while others push back that electricity isn't free, and the real value here is the DIY tinkering ethos, not the cost savings.
Queryable Executables [comments]
149 points · 25 comments · fzakaria.com · 7h ago
The article proposes a format called SELF where executables are SQLite databases, so a program can store its own code, state, and logs in the same file and query itself with SQL — a proof-of-concept web server runs entirely from a single database file that also serves pages and records visits. HN mostly loved the creativity, calling it “deranged” and “one of the best things this year,” with comparisons to Smalltalk images, OS/400 libraries, and MUMPS globals, but the dominant pushback was security: multiple commenters pointed out that making an internet-facing binary self-writable turns SQL injection into code execution, and that multi-user setups break because each user needs their own copy. Others debated whether the interpreter could use a custom VFS or mount namespaces to protect code pages, and a few noted the practical nightmare of resetting a mutated executable to a known-good state. The thread split between those who want to see the idea pushed further and those who would never let a production service modify its own binary.
Show HN: LatticeDB – Like SQLite but for graph databases [comments]
148 points · 38 comments · github.com · 15h ago
The submission is LatticeDB, an embedded single-file graph database that combines graph traversal, vector similarity, and full-text search in one query layer, aiming to be for graph data what SQLite is for relational. The author was refreshingly candid about having built the entire thing with heavy LLM assistance, describing how tests would pass but the feature would be clearly broken when actually used — a sentiment that resonated with many commenters. The thread split into two camps: one impressed by the engineering and benchmarks (especially the 70x speedup over SQLite on variable-length path queries), and another pointing to existing alternatives like LadybugDB (the open-source successor to Kuzu) and duckpgq, with the LadybugDB maintainer chiming in to correct some benchmark comparisons. Several people pushed for practical features like a litestream-style continuous backup mechanism and RDF/SPARQL support, which the author promised to add based on the feedback.
C2PA Cameras Do Not Survive Contact with Reality [comments]
141 points · 81 comments · www.da.vidbuchanan.co.uk · 12h ago
David Buchanan’s post is a practical demolition of C2PA’s claim to authenticate photos on Android, showing exactly how he forged signed images from a Pixel camera app using both software exploits and cheap hardware glitching. HN largely agrees that C2PA on Android is fundamentally broken, but the conversation splits on what that means: some argue the whole provenance project is snake oil that will make people *more* trusting of fakes, while others point out that defeating 99.9% of casual slop mills is still a win, even if state actors can bypass it. A running thread questions whether Apple’s vertical integration and Secure Enclave could do better, with skeptics warning that Apple’s eventual system will just create a PR disaster when a signed fake goes viral. The underlying frustration across the comments is that Google shipped a system whose security hinges on devices never being rooted—a premise that, as the author demonstrates with a cigarette lighter and a CVE-2026 exploit, doesn’t survive a single afternoon of tinkering.
AI is hitting entry-level jobs hardest, Stanford study finds [comments]
140 points · 165 comments · arstechnica.com · 17h ago
The Stanford study shows AI is now causing a 19 percent employment gap for workers aged 22–25 in exposed fields, up from 13 percent last year, as entry-level white-collar jobs take the hit while older workers are largely untouched. The thread immediately split between those arguing the jobs simply aren’t coming back—comparing it to horse-and-buggy transitions—and a more skeptical camp insisting the layoffs are driven by executives chasing AI optics and short-term bonuses, not actual productivity gains, with one line of reasoning that hiring managers now get asked "why aren't you using AI?" instead of praised for training a junior. Several people pointed out that the real bottleneck isn't AI capability but the collapse of the training pipeline: companies won't invest in juniors because they can poach later, and the "lost generation" of 2022–2030 will leave a senior drought that eventually forces rehiring, though others countered that by then AI will be even more capable. A persistent undercurrent pushed back on the article's framing, noting that entry-level tech has always been brutal for non-traditional backgrounds and that the study only covers white-collar roles—kitchen potwashers and apprentice welders aren't affected, but nobody on HN talks about them.
Show HN: I made a Raspberry with Qwen my local car AI [comments]
132 points · 37 comments · github.com · 16h ago
A Raspberry Pi 5 running a 35B Qwen model locally, plugged into the car's OBD port and manufacturer cloud, lets you ask questions from the owner's manual and control climate or locks via voice or chat. The HN crowd immediately zeroed in on the model choice: while the 35B is impressively dense, the Pi 5's memory bandwidth (~17 GB/s) makes even 12B models painful, so a smaller reasoning-oriented model like Ornith or Qwen 4B would be more practical. Several commenters were deeply skeptical of the "MITM the manufacturer's API" claim, and the README itself admits some cloud integrations are untested against the real car. A chunk of the discussion questioned the entire premise—why use an LLM to look up oil type when the manual is already on the SD card, and why build an offline system that still routes through the author's paid GroupMind service? The pushback was sharp, with some calling the project aspirational marketing rather than a working product.
Anthropic tells staff to work from home due to possible security team strike [comments]
121 points · 126 comments · www.businessinsider.com · 18h ago
Anthropic told San Francisco staff to work from home this week over a possible strike by security contractor Allied Universal’s unionized workers, though the union says it hasn’t authorized any strike. The HN discussion mostly ditched the details of the labor dispute to dive into a sprawling debate about whether billionaire bunker security teams would stay loyal after civilization collapses — with one side arguing strongmen have always kept goons in line, and the other pointing out that palace coups and assassinated bodyguards are also historically common. A few commenters with local knowledge pushed back hard, noting that Anthropic is just a subletter in a building where they can’t control the security contract, so this is less a PR blunder and more a non-story blown up by Business Insider. The prevailing sentiment among the more grounded takes was that Allied Universal’s security guards are underpaid and need unionizing, and that Anthropic’s real mistake is not using its leverage to pressure the building owner, not the remote-work mandate itself.
France's tax agency got hacked (in French) [comments]
116 points · 53 comments · www.cybernetica.fr · 19h ago
France's tax agency breach leaked 678,000 records including names, income details, and tax rates, with the same attacker also hitting the land registry for over two million people and the Education Ministry quietly suffering its own data theft. The HN thread zeroed in on the detection gap — the intrusion was caught in late June but the actual data theft went unnoticed until the stolen files hit the dark web in August — as the real indictment of French government cybersecurity, not the breach itself. People familiar with public-sector IT immediately recognized the pattern: legacy "trust everything inside" architectures that got VPN access slapped on during COVID without rethinking the underlying trust model, a problem they said is endemic everywhere. A 2023 EU court ruling from Bulgaria's tax agency hack got flagged repeatedly, establishing that fear of misuse alone counts as damage and shifting the burden of proof onto agencies to show their security was adequate, which several considered the most actionable takeaway. The thread also devolved into a running joke about the phrase "got hacked in French" and a few sprawling anecdotes about French bureaucracy's reflexive refusal to help anyone, though the core consensus was that France's stalled NIS2 transposition law and political preference for comms over engineering are the deeper problems.
OpenAI restores 5-hour Codex and Work limits for ChatGPT Plus users [comments]
113 points · 120 comments · 9to5mac.com · 19h ago
OpenAI is bringing back a five-hour daily cap on Codex and ChatGPT Work for Plus subscribers, after a brief period where only a weekly limit applied. The HN thread largely pushes back on OpenAI’s stated justification—that the limit smooths compute load and protects casual users from accidentally burning through their week in one sitting—with many arguing that a mandatory cap is a poor substitute for an opt-out guardrail, especially for people who want to binge on weekends. Several commenters liken the reset timers and "surprise" resets to casino or mobile-game mechanics, calling out the deliberate design of variable rewards and habit formation. Others point out that the $20 Plus plan suddenly looks less competitive against Cursor’s team subscription, and that corporate users can’t even buy the $100/$200 Pro plans to bypass the limit. There’s also a split between those who find the cap helps with work-life balance and those who see it as a frustrating throttle that punishes efficient, concentrated use.
Generated 2026-08-26 08:12 UTC
Generated by Sauron from Hacker News discussions and linked articles.