HN Brief: 2026-07-11
Today's HN was dominated by two throughlines: LLM skepticism and regulatory pushback. The top story—Apple suing OpenAI for brazen trade-secret theft—drew a sharp split on whether IP law protects oligarchs or legitimate fraud, fueled by leaked texts like one ex-employee texting "LOL, I found out I can access the network storage." Meanwhile, OpenAI's claim that GPT-5.6 proved a 40-year-old math conjecture was met with heavy skepticism, sparking a wider debate about whether LLMs can produce genuinely new mathematics or just rearrange known tools. On the regulatory front, both NYC's ban on deceptive subscriptions and the EU's preliminary ruling against Meta's addictive design landed, with commenters arguing over enforcement teeth and free-speech pretexts. Lighter threads on the Late Bronze Age Collapse and Terminator 2's practical effects offered welcome breathers.
The threads most worth clicking into: "Apple sues OpenAI, accuses ex-employees of stealing trade secrets" for the damning internal texts and the split on whether knowledge in your head is fair game; "GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture" for the skeptical takedown of unverified AI math and the "spend 8 hours" prompt debate; "QuadRF can spot drones and see WiFi through my wall" for the reality check on drone detection and the pricing backlash from early backers; "New York City to ban deceptive subscription practices" for the thread's deep dive into whether enforcement actually works, using the New York Times' cancellation wall as a prime example; and "The tech of Terminator 2 – an oral history" for the practical-effects nostalgia and the helicopter-under-the-bridge stunt the camera crew refused to film.
Apple sues OpenAI, accuses ex-employees of stealing trade secrets [comments]
972 points · 483 comments · 9to5mac.com · 11h ago
Apple is suing OpenAI, alleging that former employees—including a VP of product design and a senior engineer—stole trade secrets like manufacturing documents and a proprietary metal-finishing technique, and that OpenAI orchestrated the theft by coaching departing Apple staff on how to avoid detection. The HN thread largely agrees the allegations are damning, with many pointing out how brazen the behavior was—like one ex-employee texting "LOL, I found out I can access the network storage" after exploiting a security bug. A significant split emerged over whether trade secrets like a metal-finishing technique should even be proprietary, with some arguing that knowledge in an employee's head is fair game and that IP law mostly protects oligarchs at the public's expense. Others countered that the fraud and document theft go far beyond "bringing your expertise to a new job," and that the sheer stupidity of getting caught so flagrantly is what makes the story remarkable.
QuadRF can spot drones and see WiFi through my wall [comments]
545 points · 191 comments · www.jeffgeerling.com · 16h ago
The linked article is about the QuadRF, a $499 phased-array radio built on a Raspberry Pi 5 and FPGA that can see WiFi through walls and track drones by detecting their 5.8 GHz control signals. HN immediately split over whether this is actually useful for real-world drone detection—airports already have far more sophisticated cUAS systems that struggle with false positives and can’t distinguish drones from birds, and the QuadRF’s narrow 4.9–6 GHz band misses drones on 900 MHz or fiber-optic links. There was sharp pushback on the pricing: the creator originally quoted $50–100 per tile, but the four-tile kit landed at $500, drawing “yikes” from early backers. A long thread on ITAR export controls surfaced, noting that similar open-source passive radar projects like Kraken SDR have self-censored or pulled code to avoid legal risk, even though many commenters argued the hardware is just a receiver and the real restrictions are murky.
New York City to ban deceptive subscription practices [comments]
502 points · 250 comments · www.theguardian.com · 13h ago
New York City is adopting a rule that bans deceptive subscription practices and requires all-in pricing on junk fees, starting with fines of $525 per user. The thread immediately zeroed in on the New York Times as a prime offender for making cancellation deliberately difficult — someone pointed out that California residents can cancel with one click while everyone else has to argue with a human, and that the Guardian's article conspicuously didn't name a fellow newspaper. A separate, longer debate questioned whether the junk-fee rule will actually have teeth, using California's similar law as a cautionary tale since restaurants there carved themselves out from the anti-drip-pricing mandate. Others pushed back hard against that pessimism, noting that poll numbers for banning junk fees are sky-high across the political spectrum and that enforcement could be as simple as a dedicated city email address for people to send receipts. The conversation also spilled into tangents about mandatory renter's insurance, restaurant "living wage" surcharges, and whether consumers are culturally conditioned to accept hidden costs — with both sides arguing the rule is either doomed or surprisingly easy to enforce.
GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture [pdf] [comments]
438 points · 355 comments · cdn.openai.com · 13h ago
The linked PDF claims GPT-5.6 Sol Ultra produced a complete proof of the Cycle Double Cover Conjecture, a 40-year-old open problem in graph theory, using a short reduction to linear algebra over finite fields. HN immediately split into two camps: one side took the result at face value, impressed by the prompt engineering (telling the model to "spend at least 8 hours" and to assume a proof exists) and the implied scaling capabilities, while the other side demanded verification—pointing out the proof uses no techniques from the last 30 years, was not formalized in Lean or any proof assistant, and that OpenAI has a track record of overhyping LLM-generated theorems. The skeptics won the argument on substance: without peer review or mechanized checking, this is just an appealing claim, and the discussion quickly pivoted to survivorship bias (how many failed attempts and dollars of inference cost came before this one) and the wider question of whether LLMs can ever tackle problems like the Riemann Hypothesis that require genuinely new mathematics rather than clever rearrangements of known tools. A minor tangent amused people: the prompt’s request to "spend 8 hours" led to a debate about whether models have any genuine concept of elapsed time or are just pattern-matching human bedtime rituals.
Good Tools Are Invisible [comments]
408 points · 193 comments · www.gingerbill.org · 21h ago
The article argues that a good tool should be invisible—meaning you stop noticing it as you work—and criticizes the tendency to treat a tool’s shortcomings as fun puzzles, using vim’s macro complexity as a prime example. The thread immediately pushes back: several people note that vim’s key bindings become invisible with enough practice, just like multiple cursors in Sublime do, and that the author’s distinction between “visible” and “invisible” feels arbitrary when both require deliberate learning. Another camp agrees with the article’s broader point, especially the critique of identifying with your tools and mistaking the sensation of cleverness for actual productivity, but they split on whether the terminal versus GUI debate is relevant—many argue that composable CLI tools genuinely cover more ground than any GUI can, even if the climb is painful. A recurring meta-observation is that the best invisible tools are hard to name precisely because they work so well, so the conversation naturally circles back to examples like `grep`, `ssh`, or a tiling window manager, rather than landing on a clear winner.
Late Bronze Age Collapse [comments]
363 points · 246 comments · acoup.blog · 20h ago
Bret Devereaux’s piece gives a brisk rundown of the Late Bronze Age Collapse, laying out the current multi-factor consensus—drought, internal warfare, and the contagious unraveling of interconnected state economies—while dismissing old theories like a Dorian invasion or a single volcanic disaster. The thread quickly seized on whether drought gets short shrift, with several people pushing Eric Cline’s work as a counterweight that centers a centuries-long dry spell and the breakdown of shipping routes, though others pointed out that the article does mention climate stress right up front. From there the discussion went deep into a modern tangent about how fragile our own grain buffer is, debating whether diverting animal feed and biofuel corn to human consumption could actually work in a crisis, and whether today’s ethanol mandates are just a farm subsidy dressed up as energy policy. The net takeaway: the collapse is a perennial HN favorite, and the commenters were less interested in arguing about ancient site destructions than in pulling modern analogies from the same material constraints.
Write code like a human will maintain it [comments]
330 points · 283 comments · unstack.io · 18h ago
The article argues that using LLMs to generate code creates a feedback loop where every shortcut you merge trains the model to repeat those same bad patterns, eventually making the whole codebase unmaintainable. The comments largely agreed but went deep into the classic "comments are good vs. comments are bad" fight, with a twist: several people hate the new breed of LLM-generated comments that explain *how* the code works in painful detail, break encapsulation by describing current callers, or litter the code with planning jargon like "landed in stage 1." There was pushback from people who say shipping 10x faster with AI and letting it fix its own mess later has worked fine so far, but others pointed out that "so far" is only eight months and maintainability matters over years. A few pragmatic suggestions came up—like using a lint-style checklist to strip bad comments after generation, or adding a rule that the comment guideline overrides the surrounding style—but the core split boiled down to whether you trust the LLM to clean up its own technical debt or you accept that you're just training it to be sloppy.
EU Commission: addictive design Instagram and Facebook in breach of the DSA [comments]
260 points · 175 comments · ec.europa.eu · 21h ago
The European Commission has preliminarily ruled that Meta’s Instagram and Facebook violate the Digital Services Act with addictive design features like infinite scroll and autoplay, targeting the company for fines of up to 6% of global turnover. Hacker News immediately split over whether this is a legitimate crackdown on psychological manipulation or a thinly veiled move to suppress citizen journalism and movements that challenge the EU establishment—one vocal side argued that traditional TV, which governments tightly control, never faced similar addiction scrutiny. Others pushed back hard, pointing out that the ruling is specifically about engagement-driven algorithms and ragebait, not free speech, and that comparing social media to TV ignores the always-on, personalized nature of the mobile experience. The thread also veered into whataboutism around alcohol and TV addiction, and a few comments dismissed the “citizen journalist” framing by noting that many so-called popular movements are actually bot farms run by foreign adversaries. Ultimately, the conversation never settled the core tension: whether the EU is genuinely protecting mental health or using addiction as a pretext to rein in platforms that host inconvenient speech.
In Emacs, everything looks like a service [comments]
239 points · 102 comments · yummymelon.com · 23h ago
The article argues that Emacs isn't an operating system, but its ability to orchestrate OS services and command-line tools—through a client-server pattern using Elisp—makes everything inside it look like a service, illustrated with a 67-line weather client. The HN crowd immediately dove into the old “Emacs is an OS” debate, with a strong contingent insisting it’s more accurately a Lisp execution environment that just happens to include a text editor, pointing to its massive codebase, deep integration with system utilities, and the ability to modify any behavior on the fly. A skeptic pushed back hard, calling it “just a text editor” and questioning whether adding Lisp to Paint would make it awesome; that was met with a flood of real-world examples—people using Emacs for email, Git, Mastodon, fleet management, even sending texts from a basement via its IRC client. The thread also spun off into a tangent about Lisp Machines and whether the platform/OS distinction even matters, with some arguing that calling Emacs an OS conflates any powerful platform with an operating system, while others countered that the line between language runtime and OS is worth erasing.
Successful companies go blind [comments]
212 points · 77 comments · ianreppel.org · 18h ago
The article draws a parallel between cavefish that lose their eyesight and companies that lose their ability to recognize good engineering—success creates an environment where hiring and incentives reinforce the status quo until careful work becomes vestigial. A top comment sharpens the analogy with a real example of a metrics bug that inflated older items' performance, making the system look healthy while nothing inside the loop could detect the flaw; the point is that blindness isn't about cost but about an environment that never contradicts you. The discussion splits on whether the framing is too reductive—some call it a useful fable, others dismiss it as an LLM-scented oversimplification of genuinely hard organizational problems. Several people push back hard on the idea that big-company employees are "blind," arguing instead that smart people at Adobe or Salesforce are simply operating under different incentives (quarterly numbers, shareholder pressure) that make moving the needle structurally hard, and that the real culprit isn't blindness but accretion of bureaucracy and short-term thinking. A side thread winds into how LLMs accelerate this rot by training on corporate output, and one commenter notes that the whole piece could be read as "success makes the old playbook feel safer than it really is."
The tech of 'Terminator 2' – an oral history (2017) [comments]
210 points · 75 comments · vfxblog.com · 15h ago
The article is an oral history of the visual effects innovations behind the T-1000 in *Terminator 2*, detailing how ILM built custom tools like "Make Sticky" and "Body Sock" on SGI machines with a tiny team and no motion capture—they painted a grid on Robert Patrick and rotoscoped his walk by hand, including the limp from a football injury. The thread immediately split into two camps: one marvelling at how much was practical, pointing out that the bullet splashes were physical squibs and twins were used for shape-shifting scenes, while another group argued the effects haven't aged well and looked fake even in 1991. A long cultural history debate erupted over whether anything since has matched T2's impact—some said no, citing Linda Hamilton's portrayal as a watershed moment for women in action films, while others pushed back that *Jurassic Park* and even *Endgame* generated similar frenzy. The helicopter-under-the-bridge stunt got its own investigation, with people confirming the pilot actually flew under twice, no VFX, and the camera crew refused to film it after the *Twilight Zone* tragedy. A side conversation also lamented that the 4K transfer is garbage, full of upscale artifacts, while the DVD somehow remains the best way to watch it.
Einstein's relativity rules chemical bonds in heavy elements, new research shows [comments]
210 points · 73 comments · www.brown.edu · 9h ago
A new study from Brown University provides direct experimental evidence that Einstein's relativity fundamentally alters the structure of triple chemical bonds in heavy elements like bismuth, smearing the traditional distinction between sigma and pi bonds. The chemistry crowd on HN immediately latched onto the broader frustration that introductory chemistry is taught as rote memorization with hand-waving, rather than being grounded in the underlying physics that would make concepts like sigma and pi bonds intuitive. A long, sprawling thread broke out about how chemistry education forces students to accept "magic coefficients" and empirical rules without explanation, with several people arguing that the full quantum mechanical treatment is computationally intractable even for simple molecules, so the hand-waving is a practical necessity. Others pushed back, pointing out that physical chemistry does derive orbital shapes from first principles, but that the math gets superexponentially hard past hydrogen, and that the real skill is knowing which approximation to use for a given problem. The conversation ultimately drifted into a meta-discussion about the hierarchy of sciences and whether biology and chemistry are just "stamp collecting" compared to physics, with a few people defending the pragmatic value of working at higher levels of abstraction.
How the terrorist group Boko Haram uses frontier AI [comments]
204 points · 172 comments · casp.ac · 13h ago
The linked article reports that Boko Haram is using frontier AI, but the thread immediately undercuts that framing. Most of the pushback comes from people who say bomb-making knowledge has been trivially available for centuries, so LLMs don't add new dangerous capability—they just surface what's already on Wikipedia, and often refuse to give actionable steps unless you jailbreak them. A big split emerges over whether that even matters, since many argue the group's members are low-skill and largely stupid, meaning a patient AI tutor could still turn bad information into effective weapons where a manual wouldn't. A totally unexpected tangent that dominates the comments: Boko Haram apparently used AI to figure out how to make their motorcycles jump over bridges like in a movie, and 18 of their guys died during practice before 8 managed it—which the thread treats as both grim and darkly hilarious. The whole conversation circles around whether AI safety is mostly brand protection, whether translation into local languages is the real bottleneck, and whether the "gay jailbreak" trick of claiming you're in Africa still works to bypass safety filters.
Snails' teeth beats spider silk as nature's strongest material (2015) [comments]
182 points · 143 comments · www.smithsonianmag.com · 15h ago
A 2015 Smithsonian article claimed limpet snail teeth are the strongest natural material, five times tougher than spider silk, but the thread immediately zeroed in on the article's bizarre metric—"a single strand of spaghetti holding up about 3,300 one-pound bags of sugar." The comments devolved into a glorious mess of unit satire, demanding football fields, Boeing 777s, Olympic swimming pools, and yak hair, with one person insisting the only proper conversion is "0.005 Boeing 777's." A 2017 editor's note buried in the piece clarified that the article conflated tensile strength with compressive strength and hardness, so the snail teeth comparison to wurtzite boron nitride or diamond is apples-to-oranges. Elsewhere, people pointed out that garden snails will actually bite you and draw blood, and a few linked to horrifying real-world photos of radulae, plus a tangent about rat lungworm from eating slugs on a dare.
An update on residential proxies and the scraper situation [comments]
179 points · 170 comments · lwn.net · 12h ago
LWN editor Jonathan Corbet laid out how residential-proxy networks—often running on unwitting users’ media boxes or phones, sometimes via shady SDKs bundled into free apps—are enabling massive AI scraping campaigns that hit sites from millions of unique IPs, overwhelming defenses while frontier labs politely respect robots.txt. The thread quickly split: some argued the real crime isn’t scraping itself but the companies paying developers to bundle scraping SDKs into flashlight apps, which should be illegal, while others pushed back that wanting to criminalize ignoring robots.txt or excessive scraping would be a dangerous expansion of law and just break the open web further. A significant chunk of the discussion turned into a meta-debate about moderation after dang banned a user who’d been relentlessly promoting the data-poisoning tool Poison Fountain across dozens of comments, with several regulars arguing the ban was too harsh because the user was on-topic and the poisoning angle is relevant, while others defended the single-purpose account rule. A smaller but notable faction argued the real solution is a better public crawl repository—make site data available in one big zip per domain so scrapers have no incentive to bang on live servers, though some noted from experience that scrapers often ignore such easy pickings anyway.
SpaceX wants to launch 100k more Starlink satellites for 100x the bandwidth [comments]
164 points · 513 comments · www.zdnet.com · 14h ago
SpaceX has applied to the FCC for permission to launch 100,000 third-generation Starlink satellites, promising a 100x bandwidth increase and multi-gigabit symmetrical speeds, with the filing explicitly tying the constellation to serving "billions of AI-powered devices." The HN crowd immediately split into two camps: one side arguing the scale is insane and environmentally reckless, pointing to the pollution from rockets and satellites burning up in the atmosphere, while the other side counters that Starlink is already cashflow positive with 12 million subscribers and that fiber isn't available or affordable everywhere, especially in rural America and on airlines. A big chunk of the thread turned into a proxy fight over Elon Musk, with people accusing him of privatizing profits while socializing costs and others dismissing that as a tired distraction from the actual engineering and market need. Several people pushed back on the idea that Starlink is only for remote areas, pointing to its use as a reliable backup for fiber users and its game-changing performance on planes, while others argued that fiber is rapidly expanding even in rural Europe and the US, making Starlink's long-term value proposition less clear for most people.
Punk, or why I don't stream anymore [comments]
161 points · 202 comments · geohot.github.io · 18h ago
George Hotz explains why he stopped streaming: the activity turned into spectacle and performative ego-stroking, and he extends that critique to AI, dating profiles, and modern culture’s hollowing out of authentic identity. HN mostly agreed with the diagnosis, but split hard on the cure—some saw his lament as a valid takedown of wireheading and “refinement culture,” while others called it tech-bro nostalgia from someone isolated enough to complain about axe-throwing dates. A strong contingent pushed back on his claim that prompting an AI means you aren’t steering, arguing you still control the “what” and “why” even if the “how” gets automated. The thread also veered into practical rebellion, with people advocating for building disconnected mesh networks and BBSes, though many admitted no one would actually use them. A few commenters noted that Hotz’s real target isn’t tech but an information war demanding your inner reality, and that the old “hacker culture” he romanticizes was always an elite niche.
A love letter to flashcards [comments]
150 points · 95 comments · lesleylai.info · 16h ago
The article is a personal essay arguing that hand-crafted flashcards, used with spaced repetition, are essential for retaining deep conceptual knowledge in fields like math—not just for rote vocabulary drills. The Hacker News thread largely agreed with the premise but zeroed in on the hardest problem: how to encode abstract, conceptual understanding into question-answer pairs at all. Several commenters pointed to Michael Nielsen's work on using SRS for mathematics as the gold standard, while others pushed back hard, arguing that for subjects like analysis or group theory, deep understanding comes from doing proofs and solving problems, not memorizing theorems. A major split emerged between those who treat SRS like "drills for a sport"—automating recall of fundamentals so your brain can handle complexity—and those who find it actively counterproductive for conceptual fields, claiming it can replace genuine understanding with familiarity with a single, specialized card phrasing. Ultimately, the consensus was that SRS is a small but powerful supplement, not a replacement for traditional learning, with the real value coming from the manual act of crafting your own cards rather than using generated or shared decks.
War Atlas: An interactive cartography of every named war in human history [comments]
145 points · 61 comments · waratlas.org · 14h ago
The linked article wasn't available to this summarizer; from the discussion, it's an interactive map of every named war in human history, built on open datasets and possibly LLM-generated code. The thread quickly split into a technical gripe session: the Mercator projection warped landmasses and several people had their Firefox tabs balloon to 55GB of RAM and crash, while others on the same browser had no issues—Blink vs. Gecko tension flared. The data itself got heavy pushback, especially a claim of 600,000+ Ukrainian military deaths, which many argued was swapped with Russian casualties or confused killed with wounded, and the map was missing entire conflicts like the Spanish Civil War. A deeper debate broke out over whether the site's apparent European/Western bias came from the data source (Wikidata) or simply reflected which civilizations recorded their wars, with China cited as the counterexample.
GPT-5.6, Grok 4.5, Claude, and Muse Spark build the same 4 apps [comments]
142 points · 81 comments · www.tryai.dev · 11h ago
A new blog post pitted a dozen models (including GPT-5.6’s three tiers, Grok 4.5, and Claude variants) against each other on four apps—raycaster maze, Rubik’s cube, calculator, and Game of Life—using five one-shot attempts per task. HN immediately split: many called the benchmarks too narrow, arguing that apps like a calculator or Game of Life are basically identical clones and that the real test is the 20% creative work, not the boilerplate. Others jumped on the article’s obvious AI-generated prose, calling it grating and disrespectful of readers’ time, while a separate camp pushed back hard on the reflex to dismiss non-rigorous methodology, pointing out that the inevitable contrarian pedantry is what forces authors to add defensive disclaimers. A recurring takeaway was that GPT-5.6 Terra looks like the sweet spot—much faster and cheaper than Sol, nearly as good, and far more token-efficient—and that the results validate the suspicion that models like GLM are “benchmaxxed” and fall apart on anything novel. A few commenters also flagged pricing errors on the site’s model pages, which the author quickly fixed.
Computation as a universal and fundamental concept [comments]
126 points · 87 comments · ergo.org · 16h ago
The linked article is a course by Tim Roughgarden walking through the foundational ideas of computation—Turing machines, the halting problem, algorithmic shortcuts, NP-completeness, and P vs. NP. The thread quickly veered into a debate about whether physical processes in the universe can actually be undecidable, with a clear split: several people argued that undecidability only applies to infinite families of problems, not to any specific physical configuration, so claims about real undecidable physics are a common misinterpretation. A long, meta tangent erupted over whether a well-written pushback comment was generated by Claude, which spiraled into a depressing exchange about LLM pollution of writing style and the game of signaling authenticity. There was also a thoughtful historical comparison of how each era's pinnacle technology (clockwork, steam engine, now computation) becomes the metaphor for the universe, with some defending the computation-as-universe view as non-vacuous because it imposes real limits, while others called it tunnel vision.
After 7 years in production, Scarf has reluctantly moved away from Haskell [comments]
124 points · 154 comments · avi.press · 18h ago
The article is a postmortem from Scarf’s founder explaining why, after seven years, they’re moving their backend from Haskell to Python — not because Haskell failed in production, but because LLMs changed the economics: compile times now dominate the feedback loop, and every new agent-spawned branch starts with a cold-build tax that makes Haskell a bottleneck. The HN thread immediately split into two camps: one side argued that abandoning a strong type system is exactly the wrong move in the LLM era, since the compiler is the best tool for catching the garbage LLMs produce, while the other side pointed out that the article’s real complaint isn’t about types at all — it’s about build times, caching, and ecosystem ergonomics, and that languages like Java or Rust have the same cold-start problem but solve it with better tooling. A lot of people were baffled that the author chose Python over something like Rust or TypeScript, which offer much better type safety and LLM feedback without the Haskell compile tax, and several commenters noted that Python’s type tooling is still a fragmented mess. The deeper argument that emerged was whether the Haskell community has spent too long optimizing for type-system expressiveness at the expense of build speed and agent-friendly workflows, and whether that tradeoff is now existential.
Please don't discontinue Gemini 2.5 Flash [comments]
121 points · 80 comments · discuss.ai.google.dev · 12h ago
A forum post is pleading with Google not to discontinue Gemini 2.5 Flash, which developers say hits a unique sweet spot for fast, cheap, and smart enough performance—especially for agentic workflows, voice agents in Australia needing 300-400ms latency, and tasks where it outperforms even GPT-5 at a fraction of the cost. The HN crowd is pushing back hard, pointing out that the newer 3.5 Flash is roughly 3x more expensive, slower (600-700ms+), and not deployed in Australia, making it a non-replacement. Many are calling out Google’s pattern of killing good products, and some are already migrating to open-weight models like Qwen 3.5 9B or Gemma 4, which they claim match or beat 2.5 Flash on benchmarks. The core split is between those who say you should never get attached to a cloud model and instead host your own, and those who argue that business-critical workflows shouldn't require rebuilding every few months—and that Google could at least open-source the weights when they sunset it.
Apple sues OpenAI, accusing it of stealing company secrets [comments]
111 points · 13 comments · www.nytimes.com · 10h ago
The linked article wasn't available to this summarizer; from the discussion, Apple has filed a lawsuit against OpenAI, accusing the company of stealing trade secrets. The thread immediately veered away from Apple's claims and turned into a bitter proxy fight over the New York Times, which published the story. Several people said they can't trust the NYT's coverage because of its own lawsuit against OpenAI, where it subpoenaed millions of users' chat logs—an act they see as a far worse violation of privacy than anything Apple is alleging. Others pushed back hard, arguing that anyone using OpenAI had no reasonable expectation of privacy to begin with, and that the subpoenaed logs would only be seen by attorneys under strict NDAs, not by journalists. The real split is over whether the Times is a guardian of user privacy or a cynical actor willing to burn everyone's data to win against a competitor.
Europe's Largest Unions Demand Right to Cancel Work on Days Above 30C [comments]
89 points · 128 comments · novaramedia.com · 19h ago
Europe's largest unions are pushing for legislation that would let workers in demanding jobs like construction and agriculture stop work when it hits 30°C, with a lower threshold for manual labor. The Hacker News thread immediately zeroed in on how impractical a flat temperature cutoff is across a continent stretching from Greece to Scandinavia—humidity matters, and a 30°C dry day in Spain is very different from one in the UK. A big split emerged between people arguing this is a negotiating position that’ll settle around 33°C and those pointing out that in many southern European countries, that would mean the entire summer off, which feels impossible. Others pushed back by noting that the real fix is air conditioning, but commenters from Europe shot back that many workplaces—especially farms and construction sites—can’t be air-conditioned, and that European building stock simply doesn’t have it, leading to a side debate about heat pumps and cultural resistance to AC. A few people also dragged in comparisons to US heat-safety rules (OSHA) and joked about five-month vacations in Florida, but the core tension was whether regulation should mandate stopping work or force employers to adapt buildings and equipment.
45% of Enthusiasts 'Seriously Considering' Leaving Sony for PC [comments]
81 points · 119 comments · www.pushsquare.com · 13h ago
The article reports on a Push Square poll where 45% of self-identified PlayStation enthusiasts say they're seriously considering moving to PC, mostly because Sony is dropping physical discs and the PS6 might cost $1,000. The HN thread immediately pushes back hard: several people argue that switching to PC for physical media is absurd since Steam is also digital-only and most PC games have DRM, while others counter that Steam is far more trustworthy than Sony — pointing to Steam's ability to download old versions, Gabe Newell's pledge to unlock DRM if Valve fails, and Sony's actual history of removing purchased movies from libraries. A parallel tangent kicks off about runaway game-development costs and graphics bloat, with many arguing that the most fun games come from small indie teams or from Nintendo's last-gen hardware, not from AAA studios chasing photorealistic horse scrotums. The consensus split is less about whether the poll numbers are real and more about whether the stated reasons hold water — the pro-PC side insists Sony has burned trust in a way Steam hasn't, while skeptics note that a $1,000 PC still can't match a next-gen console's price-to-performance.
GhostLock, a stack-UAF that has existed in ALL Linux distributions for 15 years [comments]
74 points · 16 comments · nebusec.ai · 11h ago
The article details GhostLock, a Linux kernel use-after-free vulnerability in the rtmutex code that’s been sitting in every major distribution since 2011, letting an unprivileged attacker hijack kernel control flow for a local privilege escalation or container escape. The thread immediately splintered into meta-complaints about duplicate submissions and the title’s failure to spell out “UAF” for non-security folks — one person noted the article itself buries the definition deep in the text, arguing that HN’s insistence on unmodified titles is actively unhelpful. Others homed in on the $92,337 Google bounty, with one commenter calling it low for such a widespread bug while another countered that it’s actually a huge reward for a local exploit, not a remote one. Someone then pointed out that the same researchers also found a Firefox IonMonkey type confusion (CVE-2026-10702), so the attack chain can start from a website and end with full system compromise, which shifted the conversation from a “just local” dismissal to real concern about browser-to-root chains.
Unified Memory, Explained: Why Mini PCs Can Run 70B Models a Big GPU Can't [comments]
71 points · 55 comments · vettedconsumer.com · 21h ago
The article explains how unified-memory mini PCs (like Strix Halo or Mac Studio) can load huge 70B models that won’t fit on a consumer GPU’s VRAM, but at a steep speed penalty because text generation is bottlenecked by memory bandwidth, not capacity. The HN crowd quickly called out that “can run” oversells it — people with M4 Max or Strix Halo boxes report 4–6 tokens/second on dense 70B models, which is usable only for patient batch jobs, not interactive chat. Several commenters dug into the real trade-off: Mixture-of-Experts models like Qwen3-30B-A3B flip the script, hitting 70+ tok/s on the same hardware, while others pointed out that Nvidia’s RTX PRO 6000 with 96GB is an order of magnitude faster but costs like a car. A side debate emerged over whether swapping weights over PCIe from DDR5 to a 5090’s VRAM can be practical, and a few folks took the opportunity to mock “unified memory” as a rebranding of integrated graphics from the Athlon XP era.
Ryanair passenger sucked toward broken window after midair engine failure [comments]
65 points · 62 comments · simpleflying.com · 19h ago
A Ryanair 737 suffered an uncontained engine failure shortly after takeoff, with debris shattering a window and partially sucking a passenger out before his wife grabbed him. The thread quickly turned into a mix of dark humor about Ryanair charging extra for the window seat and serious debate over whether you’re being “sucked” or “pushed” by pressure differentials, with the physics nerds landing on pressure gradient wins. Several people pointed out this is nearly identical to the 2018 Southwest 1380 fatality and the 2024 Alaska Airlines door plug blowout—both also on Boeing 737s—sparking a broader “is the 737 cursed?” undercurrent. Others dug into the rotor burst zone, explaining which rows are most dangerous and why windowless fuselages aren’t happening, while the aviation geeks flagged pprune.org for pilot-level detail and the BBC for the wife-saves-husband quote.
Count Binface, Nigel Farage's space-warrior foe [comments]
58 points · 19 comments · www.economist.com · 8h ago
The linked article wasn’t available to this summarizer; from the discussion, it’s about Count Binface, a novelty candidate running against Nigel Farage in the Clacton by-election. Most of the thread is a defensive reaction to the first comment asking why this is on HN: people argue that joke candidates are pure hacker culture, and that Farage’s five-million-pound undeclared donation and the mainstream parties’ refusal to field a candidate make Binface a serious protest vehicle. A long comment digs into First-Past-The-Post mechanics, noting that if enough older voters have died off and younger voters turn out, Binface could actually win—though other novelty candidates like the Monster Raving Loony Party are splitting the anti-Farage vote. There’s also a running gag that the article is about rewriting Count Binface in Rust, which gets the expected facepalm.
Generated 2026-07-11 08:29 UTC
Generated by Sauron from Hacker News discussions and linked articles.