HN Brief: 2026-06-30
Today’s Hacker News was pulled between platform power struggles and technical pragmatism. The biggest story—a bogus DMCA takedown against a postmortem of a startup collapse—exposed how Google rubber-stamps fake claims from fake people on uninhabited islands, with a Streisand effect already unfolding. That thread sat alongside a Supreme Court ruling that geofence warrants need Fourth Amendment protection, though the conversation quickly soured into whether the executive branch even follows such orders. Meanwhile, the Qwen 3.6 27B model sparked a hardware class-war: the framing demanded a $10,000 laptop, but actual users were running it on 24GB cards and finding it shockingly good for agentic coding, while a separate debate raged over whether Mullvad’s CEO personally funding a far-right party makes any VPN a political choice.
Threads most worth clicking into: “Pollen tried to remove my article and Google is assisting with it” — a fake DMCA from Bouvet Island gets a post pulled, igniting a fight over the perjury-free takedown system that’s never been prosecuted. “The CEO of Mullvad is the main financer of the Swedish Örebro party” — a VPN founder bankrolled 72% of a far-right party, splitting HN between cancel-subscription fury and a messy defense of anti-immigration nationalism as left-populist. “US Supreme Court rules geofence warrants require constitutional protections” — the ruling is celebrated, but the thread quickly pivots to private data brokers filling the gap and a dark debate over whether paper constitutions matter when the President controls the guns. “Rocketlab acquires Iridium” — an $8B satellite deal that pits L-band’s irreplaceable portable antennas against imminent direct-to-device services, with the room split on whether Neutron’s unproven rocket makes the whole thing a binary bet. “What happens when you run a CUDA kernel?” — a rare end-to-end walkthrough from source to warp scheduler, with a sharp tangent on whether NVIDIA driver bugs are real or just application bugs, and whether AI can ever dethrone manual kernel optimization.
Pollen tried to remove my article and Google is assisting with it [comments]
881 points · 126 comments · blog.pragmaticengineer.com · 22h ago
Gergely Orosz's 2022 postmortem on events startup Pollen's spectacular collapse—double-charging customers, unpaid wages, lost pensions—was yanked from Google results by a bogus DMCA filed under a fake name from an uninhabited island, copying zero sentences from a 1998 NY Post article. The thread's main takeaway is that this is a well-known playbook: reputation management firms abuse a system where Google rubber-stamps takedowns because the law incentivizes compliance over verification, and the perjury penalty for false claims has literally never been enforced. A big split emerged between those arguing the DMCA process itself needs a court-order requirement and those saying the real fix is for Google to stop using its cheapest safe-harbor path and actually verify claimants—since even basic identity checks would kill the "Ellie Piee from Bouvet Island" trick. There's also a quieter angle: a few people traced the double-barreled Negus-Fancey name to a music-industry family, and multiple replies pointed out that the article being suppressed is now necessarily streisanding all over the front page.
Qwen 3.6 27B is the sweet spot for local development [comments]
875 points · 597 comments · quesma.com · 14h ago
Piotr Migdał argues that Qwen 3.6 27B is the first local model that genuinely works as a general intelligence, posting benchmark scores that rival GPT-5 and Claude Sonnet 4.5 while running on a Macbook Max M5 with 128GB or an RTX 5090. The thread immediately pushed back on the hardware gatekeeping — most people pointed out that the model runs fine on a 24GB 3090 or 64GB M1 Max at lower quants, and that the $10,000 laptop framing is misleading. The real split is between people who are actually running it on modest hardware and saying it’s shockingly competent for agentic coding, and others who swear the upcoming Qwen 3.7 will make this version obsolete overnight. A significant tangent emerged around Gemma 4 12B being a frustrating dud for coding tasks compared to Qwen, and some users outright said Llama 3.3 simulates doing work and then gaslights you about missing artifacts. The discussion also settled the MLX vs. llama.cpp debate by noting they’re now nearly identical on Apple Silicon, with M1 Max still being the value king for memory bandwidth despite slow prefill.
The CEO of Mullvad is the main financer of the Swedish Örebro party [comments]
596 points · 1324 comments · www.flamman.se · 21h ago
The linked article reports that Mullvad VPN’s CEO personally bankrolled 72% of the far-right Örebro party’s 2025 funding—a party pushing “remigration” (mass deportation of non-white residents) and whose leader has called immigrants “damn parasites.” HN largely reacted by immediately canceling subscriptions, with many drawing parallels to Proton’s recent far-right vlogger scandal and calling this a pattern of VPN founders funding extremism. A vocal minority pushed back, arguing the party is actually left-populist on economics and that the “extremist” label is overblown, leading to a messy fight over whether anti-immigration nationalism can coexist with socialist welfare policies—and whether calling it “ethnic cleansing” is accurate or hyperbolic. The thread also veered into broader ideological wrestling: some claimed all VPN services must be either far-right or far-left because they serve marginalized users, while others dismissed the whole privacy-via-VPN premise as “trust me bro” theater regardless of the politics.
US Supreme Court rules geofence warrants require constitutional protections [comments]
513 points · 243 comments · www.theguardian.com · 16h ago
The Supreme Court ruled 6-3 that geofence warrants require a warrant with Fourth Amendment protections, rejecting the government's argument that voluntarily handing over location data to Google isn't a search. The thread broadly cheered the outcome, but quickly pivoted to debating how the ruling might actually play out — one comment linked to a company that sells aggregated location data to law enforcement, and several people pointed out that Google doesn't *sell* your data, so the real money is in private data brokers filling the gap. A fair amount of discussion centered on Justice Barrett joining Alito and Thomas in dissent; some argued she's simply more aligned with property-rights-focused interpretations of the Fourth Amendment, while others floated the theory that Chief Justice Roberts has lost control of the court and can't swing conservative votes to the center anymore, leading to fractured rulings. That opened a wider, somewhat fractious argument about whether the Supreme Court's legitimacy depends on the executive branch actually following its orders — with people pointing to both the current and past administrations ignoring rulings, and one commenter bluntly noting that paper constitutions don't matter when the President controls the guns.
.self: A new top-level domain designed to support self-hosting [comments]
480 points · 271 comments · hccf.onmy.cloud · 12h ago
The Human-Centered Computing Foundation is pitching a new `.self` top-level domain to simplify self-hosting, offering free subdomains, shared mail servers, and integrated TLS certs—all governed by community input. Hacker News immediately questioned the need for yet another gTLD, arguing that `.name` or `.me` already serve the same purpose and that the real value would come from the actual services, not the TLD string itself. Several people pushed back hard on the idea of polluting the root namespace, with one calling it a "commons-pollution problem" and noting that `.duckdns.org` or even plain `.net` would work without the bureaucratic overhead of an ICANN application. The thread also split over whether the initiative's reliance on a shared mail server undermines the whole "self-hosting" premise, and the site getting hugged to death on its own underpowered server only added to the irony.
Rocketlab acquires Iridium [comments]
401 points · 262 comments · investors.rocketlabcorp.com · 17h ago
The linked article wasn't available to this summarizer; from the discussion, Rocket Lab is acquiring Iridium in a deal that values Iridium at $8B against Rocket Lab's $57B market cap, giving the launch company a profitable satellite operator with valuable L-band spectrum and a guaranteed 23 launches per year. HN quickly split into camps: one side argued the spectrum and recurring revenue stream justify the price, while the other pointed to Iridium's heavy debt and the risk that its low-bandwidth, power-efficient niche gets squeezed by direct-to-device services from AST SpaceMobile and Starlink that work with regular phones. A deep technical thread erupted over whether Iridium's terminal advantages are real or just a consequence of slow speeds and physics—some commenters insisted that beamforming and spectrum reuse make modern Ku/Ka-band constellations superior, while others countered that L-band's omnidirectional antennas are irreplaceable for portable, battery-powered use. The biggest pushback centered on Rocket Lab’s valuation itself, with skeptics calling Neutron an unproven carbon-composite rocket that's years behind schedule and warning the stock price is pricing in binary success or fail—if Neutron doesn't work, the grand vision (and the debt used to buy Iridium) collapses.
Samsung, SK Hynix, Micron Sued in US over Memory Price Fixing [comments]
391 points · 179 comments · en.sedaily.com · 20h ago
The linked article wasn't available to this summarizer; from the discussion, Samsung, SK Hynix, and Micron are being sued in the US for allegedly fixing DRAM and NAND prices again. The thread immediately recalls the industry’s long history of price-fixing convictions and notes that a similar 2022 lawsuit was dismissed for lack of evidence of explicit agreement. Many argue that with demand surging and barriers to entry astronomical, the three companies don't need a backroom deal—tacit collusion or simple market dynamics like OpenAI's 40% supply grab explain the price hikes just as well, making prosecution nearly impossible. A sharp split emerges: some say the real harm is the outcome (persistently high margins with no new entrants), so the law should punish the appearance of collusion, while others counter that building new fabs takes years and anyone who overbuilt in past booms went bankrupt, so the current pricing isn't suspicious at all.
European ISPs Want Rightsholders Held Accountable for Overblocking Damage [comments]
376 points · 111 comments · torrentfreak.com · 15h ago
European ISPs are lobbying the EU Commission to hold rightsholders financially accountable when aggressive site-blocking orders take down innocent sites and services. The thread zeroed in on Spain and Italy, where soccer leagues like LaLiga and Serie A have pushed blocking so far that shared IP addresses took out banking apps and developer tools for millions of users — and a Portuguese hosting provider lost email with Italy for 16 days. Several commenters argued this is just bread and circuses: soccer gets special treatment, money, and street closures while hospitals and parks go underfunded, and they see the same pattern in US high school football stadiums. The discussion also split into a sharp debate on whether censorship is ever good, with one side insisting it's always bad and the other pointing to child porn — but the practical retort was that if you can't reliably identify what to block, overblocking is inevitable even with good intent.
Instagram is incorporating users' photos in ads for Meta Glasses [comments]
341 points · 145 comments · x.com · 18h ago
The linked article wasn’t available to this summarizer; from the discussion, Instagram is now pulling users’ photos into ads for Meta Glasses. The thread quickly became a long sigh of “this again” — people tracked the exact same language back to Facebook’s 2013 terms of service, where you already grant permission to use your name, profile picture, and content in sponsored posts with zero compensation. A lot of pushback focused on the futility of reading or fighting those terms: you either delete your account or accept that every social platform claims the same rights, with several people pointing out that even Hacker News’ own ToS contains nearly identical language. A handful of commenters argued that the real issue isn’t the legal fine print but Meta’s proven track record of abusing user trust (Cambridge Analytica, emotional manipulation experiments, and scam ads), making this just one more predictable move from a company that treats users as inventory.
Tidal AI Policy [comments]
299 points · 330 comments · tidal.com · 18h ago
Tidal quietly published an AI policy that accepts AI-generated music but requires it to be labeled and explicitly bans monetization—no royalties for AI tracks. The HN crowd largely cheered the labeling and demonetization, with many saying “now give us a toggle to hide all AI content” and wishing YouTube would follow suit, though a few pointed out that enforcement will be a nightmare and that most people uploading AI slop will simply lie about it. The conversation quickly moved beyond Tidal to the broader flood of AI slop on subscription services, with people drawing parallels to Kindle Unlimited and arguing that the real problem isn’t AI itself but the collapse of discoverability—one longtime strategy being to just consume media more than ten years old. There was also a sharp split between listeners who want AI for background noise and those who see it as a copyright evasion scam, plus a practical gotcha: the policy page returns a 404 if you’re logged into Tidal.
A native graphical shell for SSH [comments]
291 points · 159 comments · probablymarcus.com · 16h ago
The article pitches a new paradigm: using SSH to serve web-based graphical apps from remote servers via Unix sockets, rather than the usual terminal or port-forwarding dance. The thread immediately split into a semantic war over whether “shell” can mean a GUI (with people citing DOS Shell, GNOME Shell, and Windows Explorer as precedent), while others dismissed the whole idea as unnecessary—terminals work fine, or you can just use X11 forwarding (which the author says is too slow). Several people pointed out that tools like MobaXterm and Windows Admin Center already fill this niche, and a bunch of commenters got confused about the difference between Outer Loop and Outer Shell, especially since the client appears to be macOS-only. A vocal minority argued these graphical shells are genuinely useful for less technical team members or for situations where the browser is the only available platform, though the counter came back fast: if you can’t SSH, you shouldn’t be touching the server.
What happens when you run a CUDA kernel? [comments]
252 points · 30 comments · fergusfinn.com · 18h ago
The article walks through the entire stack of a simple CUDA vector-add kernel, from source code down through PTX and SASS assembly, the driver’s doorbell and QMD mechanism, and finally to warp scheduling and static control codes on the GPU. Several people pointed out that using the CUDA driver API instead of the runtime API gives far better visibility into launches and even allows hot-reloading kernels, with a couple of library wrappers offered as cleaner alternatives. A heated split emerged over NVIDIA driver reliability: some engineers said dealing with driver and library bugs eats a huge percentage of time at scale, while others countered that most reported bugs are actually application bugs and that NVIDIA’s driver team is transparent when the bug is theirs. A separate tangent debated whether AI or open-source libraries will eventually dethrone companies that specialize in manual kernel optimization, with skepticism that current models can handle weird memory layouts and skinny shapes, though KernelBench scores suggest models are good on happy-path kernels but weak on sustained or shape-robust workloads. The post itself was widely praised for connecting the CPU-side launch syntax to the actual hardware submission path, something most explanations skip.
Studio Canal Movies purchased on PlayStation Store removed without refund [comments]
219 points · 125 comments · www.playstation.com · 18h ago
Sony posted a legal notice that it will delete over 500 Studio Canal movies from PlayStation customers’ video libraries on September 1, 2026, with no refund—movies people paid real money to “buy.” The thread immediately shredded the fiction that a digital “purchase” means anything, with most people arguing that this is exactly why you either buy physical discs or pirate, since Sony’s licensing deal with Studio Canal didn’t include perpetual rights for already-sold copies. A lot of pushback landed on Sony for signing a contract that allows this, but the real split was over whether the average consumer should have known better—some say convenience is king and it’s unreasonable to expect a $6 movie to be hosted forever, while others insist the industry deliberately uses the word “buy” to trick people into renting. The conversation also veered into a practical debate about kids’ movies (the main use case for digital purchases), self-hosting with Jellyfin, and the grim truth that even Apple or Amazon could pull the same stunt tomorrow.
South Korea to spend $1T on more memory chip production and humanoid robots [comments]
212 points · 135 comments · arstechnica.com · 9h ago
South Korea's government and companies are pouring a trillion dollars into doubling DRAM production, building AI data centers, and mass-producing Boston Dynamics' Atlas humanoid robots by 2028, aiming to lock down the AI supply chain. The HN discussion immediately zeroed in on a translation oddity in the president's speech—"great leap forward" sparked comparisons to Mao's campaign until Korean speakers clarified the actual phrase "대도약" carries no such historical baggage and is just a generic "big jump." From there the thread veered hard into why Germany didn't grab this industry, with detailed postmortems on Qimonda's bankruptcy and the argument that state intervention was politically toxic after 2008, while others countered that chip fabs require grinding highly educated workers around the clock in ways Western labor markets won't tolerate—citing TSMC's Arizona delays as proof. The labor angle got explicit pushback from someone who said the "Asian work ethic" framing obscures aggressive state investment and chaebol politics, though the original poster doubled down that inhumane shift work for degreed engineers is the real structural barrier, not culture.
Why did this journal retract two 1940s papers by Max Planck? [comments]
208 points · 27 comments · arstechnica.com · 23h ago
The article reports that a journal retracted two 1940s papers by Max Planck for alleged copyright violations, which historians argue stems from modern publishing algorithms misapplying contemporary standards to historical work where duplicate publication was normal. The HN thread quickly zeroed in on the absurdity of copyright claims against a scientist who died in 1947—German copyright expires 70 years after death, so these papers should be public domain—and debated whether “self-plagiarism” is a meaningful concept at all, with several people arguing it’s a made-up rule designed to protect journal turf rather than a genuine ethical breach. Others flagged the portmanteau “Zensurheberrecht” (censorship + copyright) as fitting perfectly here, and a side conversation compared the title’s clickbait quality to the earlier Science.org article, noting Ars Technica’s phrasing was slightly less baity but still annoying. The consensus was that this is a clear case of algorithmic overreach and publisher incompetence, not scientific misconduct.
WATaBoy: JIT-Ing Game Boy Instructions to WASM Beats a Native Interpreter [comments]
202 points · 33 comments · humphri.es · 16h ago
The article describes a Game Boy emulator that uses a JIT compiler to turn instructions into WebAssembly, exploiting the browser's own JIT to get around iOS's ban on native JIT — and it actually beats a native interpreter in benchmarks. HN immediately zeroed in on the iOS angle, with some pointing out that there are already sideloading workarounds for apps like Dolphin, though others counter that those require tethering or debug entitlements, making a WebView-based approach far more practical for distribution. The performance results sparked a split: several commenters note that of course a JIT beats an interpreter (interpreter overhead is ~1000% vs WASM's ~20%), but others insist a proper natively-coded emulator from twenty years ago still crushes this on modern hardware. A side tangent dug into why Firefox is 25% slower than Safari and Chrome on the benchmarks, and someone linked Andrew Kelley's old Jamulator article on static recompilation, noting the author there had predicted a JIT approach would be the way through the same kind of tooling headaches.
Ornith-1.0: self-improving open-source models for agentic coding [comments]
200 points · 39 comments · github.com · 14h ago
The linked article announces Ornith-1.0, a family of open-source coding models fine-tuned on Qwen and Gemma 4 with a reinforcement learning framework that jointly optimizes solution rollouts and the harness driving them, claiming state-of-the-art benchmarks for their sizes. The HN crowd immediately pushed back on the “self-improving” label—it’s about the training process, not runtime behavior—and several people called it a benchmark-maxxed fine-tune with misleading marketing. Some users who actually ran it reported it’s faster than Qwen 3.6 for C++ tasks and occasionally creative, but others found it hallucinated heavily without tools and failed follow-up sessions; one pointed out that a previous independent test on the same benchmarks had ranked it poorly. A split emerged between those who see this as yet another hype-driven fine-tune from crypto-world refugees and those who find it genuinely useful for local agentic workflows, though the missing 31B weights and a cited benchmark that bizarrely puts Kimi K2 models below Ornith 35B kept the skepticism loud.
Mag 7 starting to underperform [pdf] [comments]
188 points · 149 comments · www.apollo.com · 17h ago
An Apollo Global Management analyst report argues the Mag 7 are starting to underperform, pointing to falling free cash flow, soaring capex that now exceeds it, and earnings growth converging with the rest of the S&P 500. The HN discussion immediately split: one camp trusts the report because Apollo manages a trillion dollars and its claim that the Mag 7 are burning cash is factually backed, while the other camp dismisses the deck as a sales pitch full of sloppy research—showing public API prices and OpenRouter token usage as if it represents all LLM usage is plainly dumb. Several people jumped in to say this looks healthy anyway, since a market less dependent on seven names reduces concentration risk, while others leaned on historical mean reversion data showing that recent winners almost always underperform over the next decade. A sharp back-and-forth emerged over whether Meta’s Metaverse waste proves Mag 7 CEOs allocate capital by vibes rather than the day job of capital allocation, and one person pointed out that the “good runs end” observation is nearly a tautology if you define “underperform” relative to an average containing them.
Sandia National Labs SA3000 8085 CPU [comments]
171 points · 44 comments · www.cpushack.com · 21h ago
The article digs into the Sandia SA3000, a radiation-hardened CMOS version of the Intel 8085 built in the early 80s for nuclear weapons and space probes, still flying in Trident II warheads today. The thread quickly split into two main camps: one side piled on details about modern rad-hard CPUs from Moog and BAE using POWER architecture, plus SPARC, ARM, and RISC-V options, making the case that the state of the art has moved on but the market is tiny and conservative. The other side got caught up in whether the article was "slop" because it rendered radiation doses as "1x106 rads" instead of 10^6 — the formatting looked like a bad copy-paste, and people dug up the original DTIC report to compare, with most dismissing it as a trivial typesetting error rather than real slop. A strong undercurrent of dark humor ran through it all: people are both amazed and horrified that a W88 warhead runs on eight 8085-class chips, but the material physics crowd shut that down hard, pointing out that trajectory and fuzing calculations are computationally trivial when the airframe itself is falling apart at the molecular level.
Working With AI: A concrete example [comments]
152 points · 50 comments · htmx.org · 17h ago
HTMX author Carson Gross wrote up a real debugging session with Claude fixing a hyperscript parser bug, showing how the AI was great at root-cause analysis and test generation but kept proposing hacky fixes that would have introduced technical debt. The HN crowd largely agreed with that assessment, with several people noting the AI’s inability to feel out code smells or think through design trade-offs — one person described it as “code inpainting” that fills boilerplate but can’t synthesize novel, clean solutions. A few commenters pushed back on Gross’s worry about intellectual dulling, arguing the brain is plastic and you regain skills quickly after relying on AI, while others countered that the real damage is in *initial learning* since you never form those memory traces in the first place. There was also a side tangent about the new htmx homepage being built with Astro and Tailwind, which felt ironic given the project’s plain-HTML ethos, though Gross said he just let a younger team member try something different.
CachyOS June 2026 Release [comments]
132 points · 65 comments · cachyos.org · 18h ago
The CachyOS June 2026 release adds a Hyprland Noctalia desktop option, DNS-over-QUIC, and various performance fixes for Python and GCC. The thread mostly ignored the release notes and turned into a near-evangelical discussion about gaming on Linux, with multiple people recounting how they wiped Windows for CachyOS and haven't looked back, praising its snappiness and gaming performance. Several commenters pushed back hard, naming specific pain points that still keep them tied to Windows: kernel-level anti-cheat in esports titles like Valorant and Fortnite, broken overlays and third-party utilities in sim racing and Elite Dangerous, and a dead end for professional audio production due to missing VST plugin support and DRM workflows. Others countered that anti-cheat is slowly being enabled for Proton (Ubisoft, EAC), that Bitwig and Renoise work fine for hobbyist music production, and that for single-player and non-competitive multiplayer games, Linux is now the better experience. One tangent revealed you can run CachyOS-optimized kernels on Fedora via a COPR repo.
The Radiation Exposure Lie [comments]
119 points · 95 comments · worksinprogress.co · 15h ago
The article argues that the linear no-threshold (LNT) model for radiation—which says even tiny doses cause harm—is poorly supported, using the Taiwanese cobalt-60 building cohort and nuclear worker studies to claim chronic low-dose exposure may not increase cancer risk, and that this overregulation has crippled nuclear power. The thread split hard: one side seized on the article to argue that “absence of evidence is evidence of absence” and that the NRC is already considering relaxing the rules, while the other side shot back that the article is propaganda, pointing to a 2026 Nature study showing higher cancer mortality near nuclear plants—though defenders of the article countered that actual measured exposure near those plants is negligible and the correlation is likely confounded by demographics. A top critic detailed how modern sequencing can’t detect radiation-specific translocations in non-cancerous tissue, so the absence of such findings doesn’t prove safety, and accused the article of p-hacking the Taiwanese data. Another tangent emerged: even if LNT is wrong, making concrete walls thinner won’t fix nuclear’s real cost problem—abysmal construction productivity in the West. The beer analogy (365 beers at once vs. one a night) got batted around: obvious that acute is worse, but the article’s critics insist that doesn’t mean chronic is harmless, just harder to detect.
LongCat-2.0, a large-scale MoE model with 1.6T total and 48B Active [comments]
118 points · 34 comments · longcat.chat · 7h ago
The linked article wasn’t available to this summarizer; from the discussion, LongCat-2.0 is a 1.6T-parameter Mixture-of-Experts model from Meituan, a Chinese food delivery company, that builds on DeepSeek V4’s architecture with some claimed architectural additions. Most of the thread focused on whether it’s really a novel contribution or just a post-trained DeepSeek variant—one commenter initially dismissed it then backtracked, but the model weights are still missing, so skepticism runs deep. The real headline for many is that they apparently trained and deployed this on tens of thousands of Huawei Ascend 910C chips, skipping NVIDIA entirely, which people saw as a huge infrastructure flex for China’s AI independence. A domain expert tested it on a tricky nuclear fuel question and found LongCat-2.0 gave the wrong answer while Gemini Flash and Qwen 3.7 got it right, leading to a side debate about whether one-off tests on non-deterministic LLMs are meaningful. Meanwhile, the model’s MoE structure (48B active) sparked the usual “can I run this on my laptop?” discussion—answer: no, you need enterprise hardware, though some argued heavily quantized large MoE models can squeeze into high-end unified memory systems.
Building Principia for Windows XP [comments]
114 points · 32 comments · voxelmanip.se · 18h ago
The article is a deep technical post from Principia’s maintainer on the grueling work of backporting a modern open-source physics sandbox to Windows XP, swapping out toolchains, patching GCC, downgrading curl, and replacing GTK3 with Dear ImGui. HN mostly applauds the effort, with one side celebrating the thriving XP enthusiast community and the other immediately litigating whether it’s safe to connect an XP machine to the internet—triggering a detailed subthread on routers, passive infection risks, and the difference between practical hobbyist setups and the old horror stories. The conversation quickly spirals into what people actually *do* with XP today: running legacy POS systems in former socialist countries, hacking DX12 onto XP-x64 for faster game performance, or just driving LLM agents from a Winamp-themed desktop—nobody really stays on the technical build details after the first few comments.
What can you confidently guarantee about your software? [comments]
111 points · 58 comments · queue.acm.org · 17h ago
The linked article wasn’t available to this summarizer; from the discussion, it argues for formal verification as a way to guarantee correctness in software, giving an e-commerce refund logic example. That claim got immediate, sharp pushback from multiple sides: the top reply dismisses formal verification as a siren song for CRUD apps, insisting that verifying the core logic doesn't touch the database or UI (the actual 99% of the code) and that even property-based testing is rarely useful there. Others counter that verifying *some* dimensions—like enforcing invariants on well-defined business rules, or proving email validation against RFCs—still eliminates whole classes of bugs, and that the same "model mismatch" skepticism could be leveled at functional programming in CRUD apps, which clearly works fine. A deep tangent spun off into whether email validation is even well-defined (RFC vs. real-world mail servers), with one side arguing formal verification would prove a wrong spec and the other saying a simple "@" check plus PBT already beats the typical broken implementation. One memorable anecdote: a former student formally verified that Lamport’s bakery algorithm fails when ticket numbers overflow—got a B because the professor didn't expect that result, but the real issue is that the proof just confirmed a known limitation of finite counters.
Rebuilding the Computer Room [comments]
109 points · 57 comments · alexwlchan.net · 20h ago
The article is a personal essay about deliberately retreating from ubiquitous computing by confining the author's digital life to a single desktop in a dedicated room, reversing the decades-long trend toward portability. HN ran with it mostly approvingly, with many sharing their own strategies—setting up separate workstations for different tasks, keeping phones on charging docks, even wearing pocketless clothes at home—but the thread quickly split over whether smartphones are "real computers" or not. A vocal contingent insisted there's no functional difference, pointing to Termux or Blender-on-Android as proof, while others countered that ergonomic and artificial restrictions still make phones a different category entirely. The housing reality check landed hard too: several people remarked that dedicating a whole room to a computer is a luxury of older or larger houses, not a realistic option for most modern European apartments.
Memory Safe Context Switching [comments]
108 points · 23 comments · fil-c.org · 7h ago
This post is from the Fil-C project, documenting how their memory-safe C dialect now supports the notoriously dangerous setjmp/longjmp and ucontext APIs without allowing stack corruption. The HN thread mostly dives into the technical weeds: several people point out that even with memory safety, these APIs still risk resource leaks, signal mask corruption, and other non-memory bugs—the project's author concedes those aren't covered but argues they're less catastrophic. A sharp commenter also corrects the article's claim that Boost uses ucontext for fibers, noting it's only a slow fallback; the author agrees and says a sigmask-free backend is a future step. The most interesting tangent is a deep back-and-forth about why C's stack-allocated pointers prevent the copy-and-relocate trick that languages like Go use for delimited continuations, with a couple of people swapping links to papers on stack remapping and segment-based relocation.
Scientists find molecular-level evidence for two structures in liquid water [comments]
102 points · 76 comments · www.nature.com · 9h ago
The article reports on a Nature Physics study that used unsupervised deep learning on molecular dynamics simulations to find molecular-level evidence for two distinct, interconvertible local structures in liquid water—supporting the long-debated "two-state" model that explains water's anomalous properties. The HN thread largely ignored the actual physics and immediately swerved into a heated, sprawling debate about homeopathy, with one side insisting the paper would inevitably be co-opted as pseudoscientific "evidence" for water memory, while others argued that homeopathic products often contain non-trivial amounts of active ingredients and that the comparison is unfair. A few commenters brought up polywater as a historical cautionary tale, and there was the expected round of jokes about firmware updates for dehumidifiers and ice-9. The science itself—how the reaction pathway changes from a full-loop with three saddle points near the phase boundary to a semi-loop with a single saddle point away from it—barely got a mention amid the culture-war tangents.
Popping the GPU Bubble [comments]
101 points · 18 comments · moondream.ai · 2h ago
The article details how Moondream's Photon inference engine eliminates "GPU bubbles"—idle GPU time caused by CPU round-trips during autoregressive decoding—by overlapping GPU work on the next token with CPU housekeeping from the previous one, using ping-pong slot buffers, deferred sampling, and zombie sequence management. Most of the HN discussion immediately seized on the title being misleading, since "GPU bubble" usually evokes the AI hardware pricing bubble, not a pipeline stall; several people argued it should be called a stall, underutilization, or simply a CPU bottleneck, though a few graphics programmers noted "bubble" is an established term in GPU pipeline profiling. One commenter used the thread to plug HAMi for VRAM allocation, steering the conversation toward the broader infrastructure challenge of optimally assigning GPU memory across models and context sizes, which they felt was tangentially related to the post's CUDA-focused optimizations. The author showed up to thank the kind words and promised more technical posts, while another reader linked a recent paper on "Speculative Pipeline Decoding" as a different approach to the same problem.
NixOS 26.05 [comments]
100 points · 87 comments · nixos.org · 18h ago
The linked article announces NixOS 26.05 "Yarara" — new packages, systemd-based initrd by default, deprecation of x86_64-darwin, and GNOME 50 — but the thread barely touches the release notes. Instead, the comments are a near-unanimous lovefest: people describe switching from Windows, Arch, or Ubuntu and now treating NixOS as their forever OS because rolling back a broken config is effortless and declarative setups kill driver headaches and config drift. A few users veer into wanting static typing for the Nix language (typenix, nix-compile) and one person tries to bait drama about community conflicts but gets completely ignored. The consensus is that NixOS makes you braver with experiments and that every other OS feels "dirty" by comparison; no real pushback or technical corrections surface.
Generated 2026-06-30 08:31 UTC
Generated by Sauron from Hacker News discussions and linked articles.