HN Brief: 2026-06-27
Today's HN was dominated by two converging themes: the U.S. government is clamping down on frontier AI with a two-tier access system, and a growing backlash against tech overreach is erupting across multiple fronts. The biggest story was OpenAI agreeing to let the feds vet who gets GPT-5.6, paired with Anthropic's Mythos release to only a hundred “trusted” companies—a pair that sparked bitter debates about regulatory capture and whether open-weight models are now the only real alternative. Meanwhile, a sharp satire about autonomous agents negotiating a malware treaty landed somewhere between comedy and prophecy, and a series of unrelated stories—Springer Nature retracting dead physicist Max Planck's papers and still selling the PDFs, California's 3D printer surveillance bill, and Sony deleting 551 purchased movies from users' libraries—all fed the same undercurrent of institutional betrayal and ridiculousness.
The threads most worth clicking into: “U.S. government will decide who gets to use GPT-5.6” for the explosive split between those who see inevitable crypto-wars-style futility and those who think enforcement is now airtight; “Incident CVE-2026-LGTM” because the satirical incident report is so plausible half the commenters thought it was real; “Springer Nature has removed two studies by Max Planck” for the pure fury of a dead man's out-of-copyright work being retracted and sold for $39.95; and “The National Parks Were Reportedly Told to Stay Silent on Deaths” for the chilling policy that centralizes incident comms under a narrative-control rationale, with a thread that devolves into arguments over whether it's just standard comms or cartoon villainy.
U.S. government will decide who gets to use GPT-5.6 [comments]
991 points · 1043 comments · www.washingtonpost.com · 13h ago
The Washington Post reports that OpenAI has agreed to let the U.S. government vet and approve which companies can access its latest model, GPT-5.6 (Sol), marking a sharp pivot from the Trump administration's earlier hands-off AI stance. The thread largely treated this as a confirmation that frontier AI is becoming a controlled technology, with a strong push toward open-source models as the only viable escape hatch—many argued that open models are already competitive for most practical use cases, citing GLM 5.2 matching Opus 4.8 on evals. A substantial split emerged between those who see this as a repeat of the 1990s crypto wars (export controls that eventually became futile) and those who think the infrastructure for enforcement is far more robust now, with some predicting the government will eventually go after open-source distribution via CDNs or even criminalize running certain models. The more cynical take was that this is less about national security and more about incumbent companies using regulation to lock out competitors, with everybody pointing to the revolving door—Dean Ball, a former Trump AI adviser, is joining OpenAI next month. A few commenters dismissed the whole panic, noting that the vetting is only for a preview period and that the real action is already shifting to Chinese and European open models that don't need US approval.
Previewing GPT‑5.6 Sol: a next-generation model [comments]
973 points · 602 comments · openai.com · 14h ago
OpenAI previewed GPT‑5.6 Sol alongside the smaller Terra and Luna models, touting gains in coding, biology, and cybersecurity benchmarks plus a layered safety stack and a limited release coordinated with the U.S. government. HN spent most of its energy dunking on the "Sol / Terra / Luna" naming—called it LLM-level creativity in the worst way—and questioning why OpenAI skipped version 6.0: the thread suspects vibe-versioning is meant to duck expectations (and possibly government scrutiny) so they can call it a minor bump if it flops. Pricing drew heavy fire too, with people noting Sol costs the same as GPT‑5.5, that Terra is just half-price 5.5, and that charging 1.25x for cache writes is a cringe Anthropic move when DeepSeek is basically free on cache hits. A big split emerged on whether this release is a genuine leap or a hasty answer to Anthropic’s Mythos/Fable—plenty of people betting Sol will trail Fable on real-world vibes and coding, while others counter that the two models always end up close and Sol at least has the price advantage.
Incident CVE-2026-LGTM [comments]
538 points · 86 comments · nesbitt.io · 19h ago
This is a satirical incident report about an AI-powered software supply chain where multiple autonomous agents fail spectacularly at detecting malware, then end up negotiating a truce over which hosts to exfiltrate. The thread immediately split between people who recognized the joke from the title ("CVE-2026-LGTM") and those who fell for Poe's Law, with several commenters insisting the article needed a more visible [Satire] tag despite it being clearly tagged at the top. The discussion latched onto the most absurd details—the CVE name itself, the legal disclaimer about a fox being "over eighteen," and the climax where two instances of the same LLM on opposing sides sign a treaty—while also nodding at the uncomfortable truth that none of this is far from believable. A few people pushed back on the idea that the scenario could actually happen with humans in the loop, arguing that the whole point of the satire is that the AI replaced the humans and then failed in ways that only speed and scale made worse.
U.S. allows Anthropic to release Mythos AI to ‘trusted’ US organizations [comments]
400 points · 401 comments · www.semafor.com · 9h ago
The US government lifted its block on Anthropic’s most powerful model, Mythos 5, letting it out to about 100 vetted US companies and agencies after two weeks of tense negotiations over jailbreak risks. The comments immediately fixated on the corruption angle—many argued the administration is simply picking favorites, creating a pay-to-play class system where only companies that can afford lobbying or "donations" get frontier access. A significant split emerged: some pointed to existing precedent for regulating dangerous products (explosives, ITAR-controlled information) without a new act of Congress, while others countered that restricting a general-purpose knowledge tool is fundamentally different and should require explicit legislative delegation. The timing alongside OpenAI’s own restricted release of GPT-5.6 didn’t go unnoticed, with a contingent suspecting the government deliberately tied Anthropic up until a competitor had something comparable ready to release on the same terms.
We can still stop California's 3D printer surveillance scheme [comments]
374 points · 124 comments · www.eff.org · 10h ago
The EFF is sounding the alarm on California’s AB 2047, a bill that would require all 3D printers sold in the state to run spyware that blocks prints it deems firearm-related. The thread immediately spotted that the article’s illustration of California looks an awful lot like an AR-15 pistol grip, which sparked a whole tangent on the inevitable false positives—people are already planning to print cute llamas that get flagged, or embed prompt injections into STL files to crash the censorware. A lot of the pushback hammered on how laughably unenforceable the law is: you can’t stop someone from just building their own printer or driving to Nevada, and the existing ban on unlicensed firearm manufacturing already covers the actual problem. A smaller but pointed faction pushed back against that, arguing that DEFCAD’s First Amendment victory made it impossible to ban publication of plans, so regulating the printers themselves is the only lever left—but they got met with detailed takedowns about how you can’t actually print a functional gun on a consumer printer, and the whole thing is just a quirk of the law treating the lower receiver as the firearm. The conversation kept circling back to how this is straight out of the East Bloc playbook—typewriter registration, EURion constellations—and the consensus was that this bill will only criminalize hobbyists and open-source experimentation while doing nothing to stop anyone determined to make a gun.
Springer Nature has removed two studies by Max Planck [comments]
368 points · 176 comments · www.science.org · 17h ago
The linked article wasn't available to this summarizer; from the discussion, Springer Nature retracted two papers by Max Planck — the dead physicist — and posted a blank white page with a vague violation notice, yet still sells the empty PDF for $39.95. The thread lit up with fury over the brazenness: Planck's work is long out of copyright, so this is a pure scam, and the comments zeroed in on the broader dysfunction of for-profit academic publishing, where public money funnels into journal pockets while reviewers go unpaid. Several people dug into the history, pointing out that Robert Maxwell (Ghislaine's father) helped push the industry toward profit-seeking, and that an earlier 2007 licensing cancellation by the Max Planck Society shows how intertwined Springer and the Institute have become. There was also a split over whether the system is "broken" or simply working exactly as designed under capitalism — with a sarcastic subthread arguing that if you don't like it, you can choose not to play, countered by replies noting that grants and institutional KPIs often mandate publishing in those very journals.
Om [comments]
332 points · 15 comments · daringfireball.net · 8h ago
John Gruber wrote a long, personal tribute to Om Malik after his death, recounting their decades-long friendship, Malik's pivot from breaking-news blogger at GigaOm to thoughtful essayist and investor, and the quiet dignity with which he faced a heart condition that finally killed him. The HN thread is almost entirely a chorus of people who knew him or his work sharing their own grief and admiration, with several pointing to a separate memorial thread for more anecdotes. One commenter dug up a recent essay Malik wrote from the ICU without disclosing his condition, which landed in their inbox and now reads as hauntingly prescient. There's a brief oddball tangent about luggage stores in New York and San Francisco being possible money-laundering fronts, sparked by Gruber's story of Malik selling suitcases near Yankee Stadium—but it's quickly overshadowed by the overwhelming sense that the tech world lost someone genuinely beloved.
Jolla Phone (October 2026) [comments]
291 points · 162 comments · commerce.jolla.com · 17h ago
The linked article pitches the Jolla Phone as a privacy-first, community-designed Linux phone assembled in Finland, with a 649€ price tag and a 99€ deposit to lock an October 2026 delivery. The thread immediately drilled into what “assembled in Finland” actually means—most people assumed it’s just final assembly of Chinese-made mainboards, batteries, and screens, not real component-level manufacturing, which undercut the European-autonomy marketing. A sharp fight broke out over Sailfish OS itself: several people pointed out that the OS has large closed-source components (especially the UI), making it less open than AOSP, and dragged up Jolla’s past investment from Russian state-owned Rostelecom, with one faction calling it a permanent trust-burn and another insisting the company restructured and is now fully Finnish-owned post-2022. That privacy comparison with GrapheneOS got ugly—advocates for Sailfish argued it’s better against big-tech surveillance, while critics said its sandboxing is laughably weak compared to GrapheneOS, which runs on a Pixel and actually supports modern apps. A handful of nostalgic comments wished Ubuntu Touch would get native support, but the thread mostly settled into two camps: people genuinely excited about a daily-drivable Linux phone, and people who think it’s a half-baked, geopolitically tainted device that’s less private than a de-Googled Android.
Ultrasound imaging of the brain [comments]
271 points · 111 comments · alephneuro.com · 20h ago
The post describes a startup’s ultrasound technique that uses injected microbubbles to produce high-resolution 3D images of brain vasculature through the skull, aiming toward a non-invasive brain-computer interface. The HN crowd quickly zeroed in on the gap between the bubble-based results and the promised contrast-free future — several people did the math on bubble density versus red blood cell density and concluded the leap requires orders of magnitude more signal than the article’s hand-waving about ML and better hardware accounts for. Others pushed back on the mind-reading hype, arguing that functional imaging of blood flow is a blunt proxy for thoughts and that the real value here is likely cheaper, faster diagnostic imaging for stroke and Alzheimer’s rather than telepathy. The thread also turned to the Theranos parallel, especially after someone dug up a tweet from the team saying they taught themselves ultrasound from scratch, which fueled skepticism about whether this is another Silicon Valley medical imaging hustle. Meanwhile, the privacy dystopia angle got plenty of airtime, with people noting that even imperfect brain-reading tech would be terrifying if deployed coercively, though others countered that the current technique requires an IV infusion and careful contact, making it far from a stealth interrogation tool.
PlayStation Is Deleting 551 Movies from Customers' Accounts [comments]
258 points · 142 comments · kotaku.com · 11h ago
Sony is pulling 551 StudioCanal movies—including *Terminator 2* and *Total Recall*—out of PlayStation users' purchased libraries with no refunds, just a terse "you will no longer be able to access" notice. The thread latched onto the word "buy" being on the store button: a solid chunk of the discussion argued that calling it a purchase when you're actually getting a revocable license should be illegal, with one former industry insider confirming that a major studio forced them to drop "purchase" language a decade ago precisely to avoid this liability. Others pushed back that this is just how software licensing has always worked and that everyone implicitly agreed to it in the EULA, but the counterargument came down hard on contract-of-adhesion grounds and pointed to UK unfair-terms regulation as a potential escape hatch. A few people openly recommended small claims court or piracy as the only sane responses, and the thread fractured into a side debate about whether self-hosting via Jellyfin actually delivers better quality and convenience than any streaming store ever could.
Why does kinetic energy increase quadratically, not linearly, with speed? (2011) [comments]
220 points · 101 comments · physics.stackexchange.com · 9h ago
The linked article is a Physics Stack Exchange question and the top answer uses Galilean invariance — the idea that physics works the same on a moving train — to cleverly prove that kinetic energy *has* to be proportional to mv². The HN thread immediately latched onto a popular anecdote about two cars braking at the same deceleration, where a blue car going 70 stops just in time and a red car going 100 crashes at about 71, which sparked a multi-comment fight over whether the scenario confuses braking force with energy dissipation rate, with some pointing out that downforce from aero completely changes the assumption of constant deceleration. A separate, longer thread turned into a confessional from people who find physics frustrating and unintuitive compared to math or CS, arguing that physics is a messy collection of forced models that have to match reality, while math lets you pick elegant axioms — so the feeling of "I just don't get the tricks" isn't a personal failure, it's a feature of the discipline.
The gap between open weights LLMs and closed source LLMs [comments]
188 points · 158 comments · blog.doubleword.ai · 10h ago
The article tries to predict when open-weights LLMs will catch closed-source ones by fitting a trendline to benchmark lags — and gets a split answer: one benchmark suggests parity by Christmas 2026, while averaging across all 18 metrics shows a flat ~5-month gap that might even be widening. The comments immediately dragged the post for conflating "open weights" with "open source," though most people shrugged and said nobody actually cares about that distinction in practice. A big thread spun off on whether open models are really sustainable, since they currently depend on corporate philanthropy (or smart business plays) that could dry up, but others countered that once weights are out they can't be taken away — and then a surprisingly long argument erupted over whether the government could actually enforce bans on local AI, with the consensus being that enforcement is laughably impossible. The whole thing ended up feeling less about benchmarks and more about whether "open" means anything when the training data is still a black box and the spigot could turn off any day.
Data centers trigger voter backlash [comments]
184 points · 332 comments · www.newsweek.com · 14h ago
Data centers are becoming political liabilities: a wave of voter anger over massive projects—like a 9-gigawatt proposal near the Great Salt Lake—has ousted multiple incumbents, from a Utah state senate president to county commissioners, with similar results in Oregon, Virginia, and Missouri. The HN thread largely saw this as a rational backlash, not NIMBYism, driven by concrete costs—power and water consumption that get passed to residents, tax breaks with few local jobs, and air-quality damage in already-strained regions. A vocal chunk of commenters argued it’s specifically about AI: voters resent subsidizing a technology that replaces workers and generates slop, and the "privatize gains, socialize losses" framing resonated widely. A few dissenters called it misdirected FUD or suspected astroturfing, but the dominant take was that communities are correctly punishing politicians who greenlit projects offering no upside to locals while letting tech CEOs pocket the benefits.
Show HN: Smart model routing directly in Claude, Codex and Cursor [comments]
166 points · 96 comments · github.com · 15h ago
A team built a model router that sits between coding agents like Claude Code and Cursor and intelligently sends each prompt to the cheapest/fastest model that can handle it, trained via RL on real agent traces. The HN crowd immediately zeroed in on the cache-miss problem—switching models mid-session nukes the prompt cache and can cost more than it saves—but the team clarified their router is cache-aware, raising the switching threshold when a model already has warm cache, which apparently made the difference in practice. Skeptics argued that in practice you’d end up using at most two models per session (planner + executor) and that small models are unreliable, while the authors countered that subagents with fresh contexts give more degrees of freedom and that their fallback mechanism (“rescue” by a bigger model) handles failures. A vocal faction also questioned whether prompt style differs per model enough to break the routing assumption, and the team acknowledged that harness (e.g., Claude Code vs. Copilot CLI) often matters more than the model itself, making the routing decision more about context than raw capability.
The AI industry is pouring millions into US elections [comments]
115 points · 82 comments · www.bloodinthemachine.com · 16h ago
The article announces a new podcast tracking the hundreds of millions of dollars AI and crypto companies are spending to sway U.S. elections, with a specific focus on a project called Tech Influence Watch. Instead of engaging with the actual data or the podcast’s premise, the Hacker News thread immediately derailed into a philosophical argument over whether corporations should be considered people and allowed to vote, with one user pointing out that a town in Delaware already lets LLCs vote on local matters. The conversation then twisted further into a debate about whether “money is speech” under Citizens United, with one side calling it legalized bribery and the other arguing you can't limit campaign spending without violating the First Amendment. A smaller but vocal thread speculated that people will soon just ask their favorite AI who to vote for, and that candidates will use AI to write emotionally manipulative stump speeches. The net takeaway was less about the actual influence of tech money and more about a weary, cynical read of America’s campaign finance system as a permanent, unchangeable feature of doing business.
What Is a Nomogram and Why Would It Interest Me? [comments]
111 points · 19 comments · lefakkomies.github.io · 14h ago
The linked article is an introduction to nomograms—graphical slide-rule-like diagrams for solving equations with a straightedge—and the PyNomo library for generating them. The thread immediately took a Bayes-ian turn, with multiple people sharing a nomogram for Bayes' theorem (the Fagan nomogram from the New England Journal of Medicine) and diving into log-odds form and explainable naive Bayes classifiers; one person claimed you can internalize it well enough to use with your eyes closed, though others questioned that level of precision. Several electrical engineers chimed in to point out that the Smith chart is essentially a nomogram and still divisive from college. A skeptic misread the title as “Nonogram” (Picross puzzles), while someone else noted the US Navy still uses nomograms for chemistry control on nuclear reactors—a concrete, real-world counterpoint to the article’s historical framing. Overall, the crowd treated this less as a tutorial and more as an invitation to share favorite surviving examples of analog calculation in the wild.
AI in mathematics is forcing big questions [comments]
108 points · 78 comments · spectrum.ieee.org · 9h ago
The article explores how AI systems like DeepMind's Aletheia and OpenAI's latest models are starting to produce publishable mathematical results, forcing mathematicians to confront whether their slow, human-centric process of struggle and insight still matters. The thread landed hard on a persistent skepticism: multiple people pointed out that you basically have to be Terence Tao to know when an LLM is wrong, so AI doesn't actually make mathematicians obsolete—it just shifts the bottleneck to verification. The real pushback came from people familiar with proof assistants like Lean, arguing that the workflow isn't about trusting an AI's output but about using it interactively with a formal verifier that catches mistakes, which changes the calculus entirely. A split emerged between those who see AI as a tool for search and formalization versus those who worry that skipping the struggle atrophies the intuition needed to ask the right questions in the first place. The most pointed correction to the article's framing was that we're not close to an oracle; for every headline-grabbing result, there are untold hours of human prompting and checking, and the real value so far is in automating grunt work, not replacing judgment.
OpenTTD 16.0-Beta1 [comments]
103 points · 12 comments · www.openttd.org · 3h ago
OpenTTD 16 beta 1 is out, and the headline feature—trains can now drive backwards—got a nod for being realistically slow, with the flip suggesting two locos as a workaround. A bigger thread spun off around the game’s notorious setup complexity: one player wished for curated NewGRF packs instead of digging through forums, and another noted that v16’s new “collections” feature might actually let you share presets, though it wasn’t clear if that’s what they wanted. Meanwhile, someone else cheerfully admitted they spent an hour confusing the four signal types, got told to RTFM, and agreed they weren’t the target audience. On a totally different tangent, one developer casually drops that they use OpenTTD as a test harness for new CI tools and orchestrators—simple enough to deploy, fun to tinker with.
The National Parks Were Reportedly Told to Stay Silent on Deaths [comments]
101 points · 49 comments · www.outsideonline.com · 10h ago
The article reports that the Interior Department has issued an internal memo barring National Park Service staff from publicly confirming deaths or the severity of injuries, a policy experts argue will make trails deadlier by hiding risks like recent bear activity or waterfall falls. On Hacker News, a sharp split emerged: many commenters dismissed the outrage as overblown, pointing out that the policy simply centralizes incident communications—common in large organizations—and that existing NPS warnings about bears or heat stroke are already more than adequate, with one arguing that the risk of driving to the park dwarfs anything inside it. Others pushed back hard, insisting that real-time fatality data helps visitors calibrate risk (e.g., knowing a bear attack just happened in a specific area) and that the administration’s motive is transparently authoritarian—control the narrative to suppress bad news, part of a broader pattern of flooding the zone and undermining trust. A few technical voices noted that NPS already publishes aggregate mortality data online, but that the only way to get the full picture is through a FOIA request, and that the two unlisted deaths cited in the article don’t prove policy change caused the silence. The thread predictably veered into a wider indictment of the current administration’s “cartoon villainy,” with some commenters insisting this is a deliberate distraction from more consequential dismantling of federal institutions.
Pre-Modern Armies for Worldbuilders, Part III: Paying for It [comments]
92 points · 16 comments · acoup.blog · 13h ago
The article dives into the nitty-gritty of how pre-modern states actually paid for their armies—not just with cash, but through devolution of costs onto individuals, communities, or elites, and through redistribution economies that moved grain and labor instead of coins. The HN crowd mostly loved the series, but one sharp thread of pushback hit hard: several people argued that paying off invaders (a la Danegeld) did work in the short term, citing the Byzantines paying Attila for a decade and the English Danegeld lasting 150 years, and that calling it a failure depends on the timescale you’re looking at. Others countered that those payments were exceptions that proved the rule, since most states were too cash-strapped to even raise a proper army, let alone bribe their way out—the article’s core point about limited liquid capital got strong reinforcement. A tangent emerged around the paradox that raising an army with devolved costs actually made the state more vulnerable to internal coups, while paying off external threats could buy time, even if it never solved the underlying problem.
The AI backlash is only getting started [comments]
89 points · 262 comments · www.economist.com · 18h ago
The linked article wasn't available to this summarizer; from the discussion, it's about a growing societal backlash against AI. The thread quickly split into a fight over whether that backlash is really about the hype cycle or the technology itself—some argued people conflate "overhyped" with "useless," while others insisted the tech has real, harmful consequences like datacenter pollution, job displacement, and opaque decision-making. A major tangent broke out around automating government bureaucracy, with one camp claiming AI could kill 90% of useless jobs and nobody would notice, and another camp pushing back hard that you'd absolutely notice a gutted workforce and that you can't solve social problems with technical fixes. Several commenters pointed to practical failures—LLMs that lie and mislead in apps, the enormous subsidies with no profit, and the way AI is being forced onto engineers who are then blamed for not "prompting right"—as evidence this isn't just a stock-market panic but a genuine, grounded resentment toward the companies and the ruling class pushing it.
A C++ implementation of a fast hash map and hash set using hopscotch hashing [comments]
89 points · 15 comments · github.com · 10h ago
A C++ library implementing hopscotch hashing for fast, cache-friendly hash maps and sets, presented as an alternative to `std::unordered_map`. The HN discussion quickly moved past the library itself, noting that its benchmarks date to 2019 and that the project was essentially outshone by Abseil's Swiss tables almost immediately after release. Several people argued that Swiss tables (and now Boost's `unordered_flat_map` with rapidhash) are the de facto state of the art, though a few pushed back, pointing out that Swiss tables have a degenerate O(N) lookup case while hopscotch offers better strict worst-case guarantees. There was also a side debate on whether any hash table can be definitively fastest across all workloads, with one commenter flatly stating "it's impossible to be faster" for general-purpose use and another calling that claim silly.
The open source DOCX editor submitted to HN a few weeks ago has been deleted
86 points · 38 comments · news.ycombinator.com · 10h ago
The linked article wasn’t available to this summarizer; from the discussion, the open-source DOCX editor that hit HN a few weeks ago has vanished—GitHub repo and website both gone, no explanation from the author. The thread quickly zeroes in on the author’s heavy use of AI to build it, with several people arguing that an LLM likely regurgitated copyrighted Office code (or even disassembled old leaks), making the project a legal target for Microsoft. Others push back that DOCX is a standardized format and the real vector could be security exploits in docx-handling code, not copyright. A handful of users had already integrated the editor and are scrambling for alternatives, while someone points to a partial mirror and a few other open-source docx tools as replacements.
Mullvad founder gave millions to extremist far right party [comments]
81 points · 182 comments · www.flamman.se · 16h ago
The linked article reports that Mullvad’s founder donated about 5 million SEK to Örebropartiet, a small Swedish party that calls for mass deportation of immigrants—a policy widely described as far-right ethnic cleansing. The discussion immediately split: some commenters argued the party’s economic platform is actually far-left, making its label ambiguous, while others pushed back hard, quoting the party leader’s own words about deporting “parasites” and noting that forceful removal of people by ethnicity is the textbook definition of ethnic cleansing, not simply a “regressive migration policy.” Several longtime Mullvad customers said they’re switching services, pointing to recent similar scandals at Proton and questioning why people are surprised that a for-profit VPN founder might hold fringe political views. A deeper tangent emerged comparing the party’s rhetoric to historical Soviet “social parasitism” laws, with one faction claiming this proves the party is just radical populism and another retorting that trying to rebrand ethnonationalist expulsion as leftist worker-revolt is intellectual gymnastics that ignores the targeted group.
Gossamer: a Rust-flavoured language with real goroutines and pause-free memory [comments]
79 points · 70 comments · gossamer-lang.org · 13h ago
Gossamer proposes a Rust-inspired language that swaps the borrow checker for reference counting with arena regions, promising "pause-free" memory management. The HN crowd immediately pushed back hard on that claim, pointing out that cycle collection and eager RC still cause pauses—especially when dropping large cyclic structures—and that real pauseless GCs like ZGC or Go's already exist, so ignoring decades of GC research is a dubious design choice. A deeply technical back-and-forth erupted over whether sub-millisecond latency actually matters when you're not running a PREEMPT_RT kernel or pinning threads, with game developers arguing that 1ms is an eternity for 300+ Hz frame budgets. A few commenters also noted the name means "bitter kid" in French, and others dismissed the slick landing page as yet another LLM-generated vibe site propping up a language that doesn't feel genuinely novel.
Long Wave radio era set to end with switch-off [comments]
79 points · 33 comments · www.bbc.com · 12h ago
The BBC is finally turning off its Long Wave service from the Droitwich masts, ending an era that began in 1934 with a musical broadcast and included wartime resistance messages and the Shipping Forecast. The HN thread quickly got into the weeds on transmitter power: someone threw out a €2.6M annual electricity cost based on 500 kW, only to get corrected that effective radiated power vs. actual grid draw are very different things, with radio engineers explaining that Droitwich actually runs multiple services through combiners with significant losses. Others noted the real reason the BBC kept LW so long wasn't just Test Match Special—it was also used for Radio Teleswitching to control off-peak electricity meters, a service that's now being phased out. A surprising tangent emerged around the masts' role in nuclear continuity, with a link claiming the station is one of the signals UK nuclear subs check to assess whether the country still exists after an attack, which landed differently than the article's heritage-preservation angle. The consensus was split between nostalgia for a reliable signal that worked everywhere and the hard reality that maintaining kilowatt-scale vacuum tube transmitters for a shrinking audience is no longer justifiable, especially when FM and DAB already cover the same ground.
Captcha proves you're human. HATCHA proves you're not [comments]
77 points · 82 comments · github.com · 19h ago
HATCHA is an open-source reverse CAPTCHA from monday.com — instead of keeping bots out, it makes you solve nasty math and string-reversal puzzles that are trivial for an LLM but miserable for a human, so only agents make it through. The thread mostly treated it as a joke or thought experiment, with a lot of people trying to figure out who at monday.com actually signed off on this and whether the devs were just pranking management. A few practical ideas surfaced: hide the challenge off-screen as a bot honeypot, or gate content behind it and serve a different version to humans who can't solve it, though the obvious rebuttal came up fast — a bot can simply be told to fail on purpose. The time limits (30 seconds per challenge) sparked debate about whether a human with an LLM open in another tab could cheat their way in, and the "humans need not apply" tagline earned a digression into CGP Grey's old video. The general consensus was less "this is useful" and more "funny idea, but it collapses the second anyone decides to exploit it intentionally."
AMD Readies Full Open-Source HDMI 2.1 Support for Linux [comments]
71 points · 18 comments · www.techpowerup.com · 21h ago
The linked article wasn't available to this summarizer; from the discussion, AMD is finally prepping open-source HDMI 2.1 support for Linux, thanks largely to Valve pushing negotiations with the HDMI Forum. The thread credits Valve for advancing Linux readiness, even if it’s ultimately self-serving for Steam hardware — and most people are fine with that alignment of interests. A split runs through the comments over whether including HDCP DRM is hypocritical for open-source values: some argue it’s just a choice users can make, others warn it’s a Trojan horse that will eventually let DRM lock down displays, though technically HDCP is handled on the GPU firmware, not the kernel, so tampering is harder than with kernel-level anti-cheat. The anti-cheat tangent pops up repeatedly — people note that kernel-level anti-cheat is a support matrix nightmare on Linux, not a fundamental impossibility, and that publishers like Facepunch deliberately block working solutions like EasyAntiCheat.
WordStar: A Writer's Word Processor (1996) [comments]
70 points · 26 comments · www.sfwriter.com · 4h ago
Robert J. Sawyer makes the case that WordStar for DOS is still the best tool for long-form creative writing because its control-key interface keeps touch typists on the home row and treats the whole document as a fluid workspace, unlike modern word processors that break rhythm with mouse and function-key hunting. The thread quickly latched onto the practical question of how to run a 1978 DOS program today, with people trading tips on FreeDOS, DOS emulators, USB floppy drives, and even ISA slot CH375 chips to transfer files off an old writing deck. Someone pointed out that while the distraction-free focus is great for fiction, it's a pain for research-heavy projects with cross-referenced files and graphics, and the conversation split there: creative writers want simplicity, others need modern export and multitasking. A few commenters name-dropped George R.R. Martin's continued use of WordStar, which triggered a dry aside about his slow output, but most of the thread stayed grounded in the nitty-gritty of whether you can actually get a modern writing workflow out of a 40-year-old editor—conclusion: yes, if you're willing to strip away everything else.
The Exhaustion of Talking to a Tool [comments]
69 points · 74 comments · ohadravid.github.io · 16h ago
The article argues that talking to LLMs is exhausting because it imposes a "social tax" without the meaningful rewards of human interaction. The HN thread mostly rejected that framing—many commenters said human conversation is far more draining, and they appreciate being able to be blunt with a tool that never gets offended or tired. A significant split emerged over whether the frustration comes from social energy at all: some insisted LLMs exhaust you through their slowness, relentless BSing, and even gaslighting, not through social ritual. Others took a different practical tack, arguing the real fix is to stop treating LLMs like conversation partners—use them as pipelines or direct query tools, not chatty coworkers. A handful of deeper comments also debated whether "understanding" even exists in humans in any meaningful sense, so worrying about an LLM lacking it may be misplaced.
Generated 2026-06-27 08:30 UTC
Generated by Sauron from Hacker News discussions and linked articles.