HN Brief: 2026-06-20
Today’s threads were split between two moods: AI skepticism and platform-power pushback. Norway’s near-ban on AI in elementary schools kicked off a familiar fight between “prudent pause” and “this is just 1990s internet panic all over again.” Meanwhile, Google Workspace started warning Firefox users they’d better switch to Chrome—setting off a debate over whether it’s a genuine enterprise security feature or a textbook monopoly lever. A third throughline was the quiet but growing anxiety that AI tools are making us dumber, backed by a Nature study on skill atrophy and a wave of companies capping AI spend after budgets blew up. And just to keep things weird, someone claimed to have cracked Linear A using Claude Code.
“Norway imposes near ban on AI in elementary school” — because the split between cautious regulation and fears of repeating past tech bans is the sharpest it’s been all year. “Google workspace threatening to block Firefox access” — a perfect Rorschach test for whether you see enterprise security or antitrust abuse. “AI Engineer Claims to Have Cracked Linear A” — the skepticism is heavy, but if it holds up, it’s a genuine bombshell; if not, it’s a masterclass in how LLMs enable confident nonsense. “Is AI ruining our skills?” — senior engineers share first-hand accounts of watching their own judgment atrophy, which hits harder than most hand-wringing.
Hyundai buys Boston Dynamics [comments]
788 points · 348 comments · startupfortune.com · 15h ago
Hyundai is completing its acquisition of Boston Dynamics by buying SoftBank’s remaining 9.65% stake for $325 million, giving the automaker full control of the Atlas humanoid robot as it heads toward factory deployment at Hyundai’s Georgia EV plant by 2028. HN’s top comment quickly sidetracked into Hyundai’s notoriously non-transferable 10-year/100K-mile powertrain warranty, with a long back-and-forth about used-car values and which owners actually get covered — a tangent that had nothing to do with robotics but dominated the thread. Others called out the article’s AI-generated image and LLM-tinged prose as sloppy, sparking a meta-fight about speed versus quality in content production. A deeper split emerged over humanoid robots vs. purpose-built factory automation: some argued the humanoid form is over-engineered for most tasks, while others countered that flexibility to replace any human worker is exactly the point. Several commenters also rehashed Google’s decision to sell BD in 2017 as a major strategic blunder, given what DeepMind could have done with the platform — though others noted BD’s reliance on DARPA funding and dangerous hydraulic bots made it a poor fit for Google’s ethics at the time.
Norway imposes near ban on AI in elementary school [comments]
621 points · 428 comments · www.reuters.com · 15h ago
The linked article wasn't available to this summarizer; from the discussion, Norway's government has essentially banned AI use for students up to age 13, then allows cautious, supervised use for 14–16 year-olds, and mandates learning appropriate AI use for 17–19 year-olds. The HN thread split pretty cleanly: one camp argued the ban is prudent, citing evidence that AI exposure hurts learning and cognition, and that we haven't yet figured out how to deploy it without damage—so a blanket ban on younger kids while studying the problem makes sense. The other side pushed back hard, comparing it to 1990s schools banning internet research, and insisting that a well-designed, guard-railed AI tutor could accelerate learning (some even claimed personal experience with faster learning via LLMs, though that was met with skepticism about self-assessment versus external benchmarks). The most pointed criticism was directed at the reporting itself for lacking nuance—commenters noted that Norway's policy is actually a graduated, thoughtful approach, but the Reuters article flattened it into a sensationalized "ban," and there was frustration over the absence of any communicated plan for researching constructive AI use in education. A recurring sub-thread also noted the class irony of tech and AI elites strictly limiting their own kids' device use.
Google workspace threatening to block Firefox access [comments]
478 points · 150 comments · tales.fromprod.com · 15h ago
The article reports that Google Workspace has started warning Firefox users they must switch to Chrome, though access still works for now, and Google support later claimed it’s only a recommendation for admin.google.com that won’t be documented publicly. HN immediately zeroed in on Google’s Context-Aware Access (CAA) product as the likely culprit—admins can configure policies that require Chrome-specific attestation, and Firefox simply can’t provide that, so the warning is just the org’s own security rule surfacing, not a Google-wide block. But plenty of people aren’t buying that excuse: they argue Google deliberately only offers Chrome as a verifiable option in CAA, making it effectively a monopoly lever, especially since there’s no way to set a “require Firefox” policy even if your IT team wanted to. The conversation split between defenders pointing to legitimate enterprise security needs (managed extensions, device verification) and critics calling it textbook anticompetitive behavior, with several commenters noting that if Microsoft had done the same with Internet Explorer, the antitrust claims would be deafening.
There are no instances in ATProto [comments]
425 points · 221 comments · overreacted.io · 16h ago
Dan Abramov's latest post argues that the whole "where are the Bluesky instances" question is a category error—atproto isn't Mastodon, and the architecture deliberately separates hosting from apps, more like RSS plus Google Reader than a bunch of walled-garden fiefdoms federating with each other. The thread mostly split between people who loved the clarity of the analogy and those who pushed back hard, saying atproto does have "instances" (relays, PDSs, the single Bluesky AppView) just grouped differently, and that the costly, centralized firehose relay isn't meaningfully decentralized. The author jumped in to clarify that relays are just an optional $30/month optimization and you can run your own or skip them entirely, and that the whole point is you can swap hosting without changing apps, or build new apps without changing hosting—something Mastodon's monolithic instances don't let you do. Others argued that in practice, the friction of running your own relay or AppView is way higher than spinning up a personal Mastodon instance, and that the moderation benefits of atproto's separation are mostly theoretical when the default relay and app are still run by one company.
AI Engineer Claims to Have Cracked Linear A [comments]
423 points · 164 comments · aiclambake.com · 15h ago
The linked article describes an AI engineer and amateur linguist, Tom Di Mino, who claims to have deciphered Linear A by identifying it as an extinct Semitic language, using statistical analysis powered by Claude Code to produce over 300 proposed translations. The HN thread is sharply split: the blog's author (who knows Di Mino) vouches for the work's credibility, noting it's under review at Rutgers and Cambridge, but most commenters are demanding the actual manuscript or preprint before taking it seriously. A lengthy debate breaks out over whether an LLM can be a legitimate tool for reproducibility—some argue stochastic systems are fine and compare it to cooking or thermodynamics, while others insist that without the exact prompts and context, the results can't be verified. There are also comparisons to past crackpot claims in linguistics, like the Dutch professor pushing a Hurrian-Urartian theory, and a reminder that even strong statistical signals don't guarantee a correct decipherment—everyone is waiting for the experts to weigh in.
Court Records Should Be Free [comments]
351 points · 73 comments · www.eff.org · 14h ago
The EFF is pushing the Open Courts Act of 2026 to kill PACER's paywall and replace the clunky federal court system with a modern, free platform. HN largely agreed that public records should be free, but a vocal minority pushed back hard—arguing that zero-friction access would let data brokers, AI scrapers, and mugshot-extortion sites vacuum up everything, including unredacted PII and frivolous accusations, which is worse than a small fee. The thread split: one side pointed out that binding precedent is already free and that RECAP already liberates documents for the motivated, while the other side called any cost a deliberate barrier that only hurts ordinary people and benefits wealthy law firms. A few commenters also noted that state courts can be far pricier than PACER, and that the real issue isn't fee amount but the principle of paying twice for public work.
Bobby Prince, composer for Doom, Wolfenstein 3D, and Duke Nukem 3D, has died [comments]
342 points · 38 comments · www.legacy.com · 12h ago
The linked article wasn't available to this summarizer; from the discussion, composer Bobby Prince—responsible for the iconic soundtracks of Doom, Wolfenstein 3D, and Duke Nukem 3D—has died, and the thread is a mix of tribute and deep technical appreciation. People immediately zero in on how his music, much of it adapted from speed and thrash metal tracks (the E1M1 theme is Painkiller by Judas Priest), was a huge part of the games' immersive atmosphere, often more than the gunshots or positional audio. A real split emerges over sound hardware: some argue the OPL2 (AdLib/Sound Blaster) chip-music versions are more authentic and exciting than the CD-quality General MIDI recordings, because Prince composed on a Roland SC-55 but many players heard it through a waveform generator—others push back, saying the original Roland MIDI is the definitive version and nobody thought OPL2 sounded better in 1993. There’s also a thread on his sound effects work for Doom and a note that the Library of Congress added the Doom soundtrack to its registry just last month, cementing his legacy beyond the retro gaming community.
How many of the 170k English words do you know? [comments]
341 points · 432 comments · vocabowl-870366514258.us-west1.run.app · 18h ago
The linked article wasn't available to this summarizer; from the discussion, it's a 100-question multiple-choice vocabulary test that estimates how many of roughly 170k English words you know. The main complaint was that the quiz requires three clicks per question with no keyboard shortcuts, so many people bailed or wanted a skip button to avoid inflating scores with lucky guesses. A big thread spun off into the test's distractors being weak—often the longest answer or the obvious opposite was correct, leading people to suspect AI-generated slop. Others dived deep into English etymology, arguing that most of the obscure words are actually French or Latin loanwords, with a side debate about whether words like "bratwurst" become English through heavy use. Someone also pointed out a typo in "Hippopotomonstrosesquippedaliophobia" and argued about how the interrobang should be rendered, which felt very on-brand for a QI-style quiz referencing Sandi Toksvig.
A new bill takes aim at government pressure to silence lawful online speech [comments]
274 points · 122 comments · www.eff.org · 14h ago
The EFF is cheering a new bipartisan bill from Cruz and Wyden—the JAWBONE Act—that would let individuals sue federal officials who pressure tech platforms into removing lawful, First-Amendment-protected speech, citing the government's coercion of Apple to pull the ICE Block app. HN immediately dug into the familiar partisan split: one side argued the real problem is both parties doing this, pointing to the Trump administration's ICE Block crackdown and the Biden administration's alleged campaign to suppress discussion of ivermectin as a COVID treatment. The thread turned into a brutal, detailed argument over whether ivermectin was actually suppressed or just scientifically debunked, with one commenter walking through the TOGETHER and ACTIV-6 trials to argue the science was settled, while others countered that the FDA admitted it never looked at the data and that a small RCT showed a 75% mortality reduction. A separate line of discussion noted that the bill is a direct response to the Supreme Court's *Murthy v. Missouri* ruling, where the plaintiffs lost on standing rather than merits, and that this law would create a clear cause of action to force the courts to finally rule on whether government jawboning is unconstitutional. The real split wasn't about the bill itself—everyone agreed government coercion of platforms is bad—but about what counts as coercion versus legitimate public-health communication, and whether the bill will protect dangerous misinformation under the guise of free speech.
The room the economy can't see [comments]
233 points · 255 comments · wilsoniumite.com · 21h ago
The article makes the case that "third places" like a Swedish gaming club for lonely teens can't survive on market logic because their value is a positive externality, and argues for a basic income floor to let people afford to run them without needing a grant committee to hand-pick winners each year. A significant chunk of the thread pushed back hard, insisting the market *does* see that value because the room's existence reflects real wants that people could fund voluntarily if they cared enough, while others called that naïve and pointed to financialization—pension funds and REITs pricing out parks and youth centers as prime real estate—as the structural force actually destroying these spaces. The conversation spilled into a sprawling debate about whether any retirement savings system can work at scale without distorting the economy, with one commenter arguing that storing enough assets for 30 years of retirement is physically impossible and that pay-as-you-go pensions are the only honest alternative. A sharp split emerged between those who see market fundamentalism as a cognitive error akin to "liver fundamentalism" and those who insist that markets, properly understood, are just the aggregate of human choices and that blaming "the market" lets people off the hook for not organizing their own communities.
Is AI ruining our skills? Early results are in – and they're not good [comments]
221 points · 283 comments · www.nature.com · 14h ago
The Nature article reports on studies showing that physicians who use AI for colonoscopy screening and software engineers who rely on coding assistants both see their own skills degrade when the AI is unavailable. The thread immediately split between a skeptical camp that compared this to hand-wringing over calculators ruining arithmetic—arguing the net productivity gain is worth it—and a worried camp pointing out that unlike calculators, LLMs threaten "basically all knowledge and communication skills," not just a narrow task. A strong contingent of senior engineers pushed back hard with first-hand accounts: they've watched heavy LLM users in their own orgs lose judgment and produce terrible code, and several described deliberately quitting AI tools because they felt their own abilities atrophying in real time. The most pointed critique argued the real danger isn't individual deskilling but a societal innovation stall—if everyone offloads thinking to a non-innovating machine, the pipeline of human insight that produces breakthroughs dries up, and vibe-coding newbies building CRUD apps won't fill that gap.
Americans express unease over SpaceX's influence on retirement savings [comments]
214 points · 121 comments · www.theguardian.com · 9h ago
The Guardian piece rounds up reader anxiety about the SpaceX IPO forcing itself into 401(k)s through index funds—people feel trapped in a casino tied to Elon Musk’s valuation. The HN thread zeroed in on the rule change Musk pushed to get SpaceX into indexes after just 15 days instead of the usual year, with most of the pushback arguing it doesn't matter much because the stock would have ended up in every retirement account a few months later anyway. A strong split emerged: some commenters dug into the mechanics, pointing out that S&P and CRSP refused to bend, so only Nasdaq 100 funds are forced to buy at the inflated price, while others insisted the real scam is the tiny public float (around 5%) creating an artificial squeeze that drives the price even higher before index funds can accumulate. The most pointed criticism came from people who said this is classic Musk market manipulation—setting an insane IPO price, then pressuring index providers to lock in buyers—and that anyone who thinks they can’t avoid it is ignoring brokerage link options or simple value ETFs that barely touch tech. A few voices warned that the whole AI bubble underpinning SpaceX’s $1.77 trillion valuation is already baked into the S&P 500 via Nvidia and Microsoft, so even if you dodge SpaceX, you're still deep in the same casino.
Think of the children: How to force real ID for all internet traffic (2023) [comments]
196 points · 116 comments · nochan.net · 11h ago
The linked article wasn't available to this summarizer; from the discussion, it argued for requiring government ID to access the internet, using child protection as the rationale. The thread immediately pushed back hard, with people pointing out that KYC/AML-style laws just shift liability down the chain, leading to self-censorship and data leaks — Discord’s recent ID leak got called out as a perfect example. A strong faction argued the real solution is putting responsibility on parents to control their kids' devices, using technical measures like RTA headers instead of collecting everyone's PII, while another camp dismissed the whole premise as a billionaire-driven false flag to justify mass surveillance. The conversation veered into a surprisingly deep tangent about whether you can code your way around the state's monopoly on legitimate violence, with blockchain smart contracts getting dunked on for failing whenever the physical world pushes back. A minor but telling split emerged between those who think giving kids unfettered internet is like handing them a loaded gun and those who found that hyperbole ridiculous, pointing out that 17-year-olds need to order pizza and do homework online.
Amazon drops Sam Altman movie after announcing OpenAI partnership [comments]
190 points · 66 comments · www.the-independent.com · 11h ago
Amazon dropped Luca Guadagnino's nearly completed Sam Altman biopic months after inking a massive OpenAI cloud partnership, and the thread didn't buy the "we'll help it find a new home" line for a second. The dominant angle was pure conflict of interest — several people pointed straight to the 1940s Paramount antitrust case that forced studios to sell their theaters, arguing that streaming platforms owning studios is the same problem reborn. A skeptical contingent noted Amazon's similar willingness to distribute a Melania Trump documentary that looked like a payoff, which earned a look, but others pushed back on assuming causality from timing alone, noting the film is being shopped to other distributors. A side debate flared over whether Altman even makes a compelling movie subject — one camp argued that *The Social Network* proved a great director can make anyone fascinating, while the other countered that Altman is a wet blanket whose ouster drama lacks the narrative stakes of Facebook's founding. The underlying consensus was blunt: this is what oligarchy looks like, and without antitrust enforcement it'll keep happening.
Hey, n00b, we didn't hire you to complete tasks [comments]
162 points · 81 comments · newsletter.kentbeck.com · 7h ago
Kent Beck’s newsletter post tells new engineers that nobody cares about how many tasks they close—instead, seniors are trying to sort them into A (high-potential future leaders), B (solid performers), or C (won't last), and the real job is to signal learning and growth. The HN thread mostly pushed back hard: a lot of people said the article’s thesis is a luxury they’ve never seen in practice, arguing that most companies hire juniors precisely because they have junior-level tasks that need to get done, not as some long-term development bet. Others found the whole A/B/C framing toxic and egotistical, especially the implication that seniors are actively trying to minimize effort on the “C’s”—multiple commenters pointed out that if a junior breaks something, it’s the team’s failure for not having guardrails, and that the article reads like a recipe for a blame culture rather than mentorship. A few people defended the spirit of the advice, saying it works in places that actually invest in growing engineers, but the broad consensus was that the reality of modern startups and short job tenures makes the article feel dated and out of touch.
Big Banana Car [comments]
145 points · 75 comments · bigbananacar.com · 13h ago
The article is about Steve Braithwaite's giant motorized banana car, which he built and has driven across North America, and his stalled plans for a world tour due to funding, insurance, and a side project. The HN thread latched hardest onto the police stops mentioned in the story—Braithwaite says he's been pulled over hundreds of times for selfies and banana puns—and split sharply: one side saw harmless, whimsical fun and a positive police-community interaction, while the other argued every such stop is an unlawful seizure under the Fourth Amendment, with the banter about "peeled out" jokes not excusing the power abuse. A tangential thread debated whether open-cockpit driving is practical (bike commuters chimed in with cold-weather survival tips), and the Arrested Development "there's always money in the banana stand" line got its inevitable, long-awaited payoff from a username that had been sitting on the joke since 2012. The overall vibe was admiration for the builder's whimsy mixed with a strong undercurrent of frustration that cops keep illegally detaining him just because his car is fun.
Norway greenlights first full-scale ship tunnel [comments]
144 points · 81 comments · eandt.theiet.org · 21h ago
Norway approved the world's first full-scale ship tunnel, a 2km blast through rock to let ferries bypass a notoriously stormy peninsula—one that has killed 33 sailors since WWII. The HN thread immediately fixated on the article's terrible, wonky renderings, sparking a debate over whether they were AI slop or just bad 3D photoshop that's been circulating since 2017. Political and economic skepticism ran deep: locals pointed out that Norway's Labour party opposed the tunnel for years but caved to keep its coalition together, while eastern papers call it a boondoggle and western ones cheer it—and one skeptic questioned whether the route even connects anywhere significant. Defenders countered that the problem isn't the destination but the weather: the sea is so vicious that even Vikings dragged boats overland, and the tunnel shaves 56 km off a route used by coastal freight and the Hurtigruten cruise-ferries. A sidebar flared up about tautological place names—"Stadhavet Sea" meaning "the sea sea"—and about how Norway's mountain infrastructure basically requires tunnels everywhere, with one local noting that a quarter of a 360 km drive is spent inside them.
How to feed a dictator [comments]
134 points · 46 comments · www.theguardian.com · 6h ago
The piece is about a new documentary, *How to Feed a Dictator*, where personal chefs for brutal leaders like Idi Amin, Saddam Hussein, and Kim Jong-il recount their experiences—everything from pepperoni pizza orders to being forced to cook a human heart. HN immediately latched onto the director’s invocation of Hannah Arendt’s “banality of evil,” and most of the thread is a slugfest over whether that quote is a misrepresentation: critics insist Arendt was talking about bureaucrats like Eichmann who knowingly pushed paper to murder millions, not about cooks who were often held hostage with their passports seized, while defenders argue the chefs’ fixation on “a great gig” and doing good work despite the carnage is exactly the banality Arendt described. A secondary argument broke out over whether Arendt even got Eichmann right—some pointed to later scholarship showing he was a fanatical Nazi, not a thoughtless cog—and that spilled into a broader discussion about moral frameworks, coercion versus willing complicity, and whether a chef who benefits from a dictator’s luxury while people starve is morally different from a soldier following orders. A few commenters grumbled about politics creeping onto a tech forum, but the dominant energy was a pedantic but substantive fight over what “banality of evil” actually means.
John Jumper to join Anthropic [comments]
118 points · 91 comments · twitter.com · 14h ago
The linked article wasn't available to this summarizer; from the discussion, John Jumper—who led AlphaFold and won a Nobel—is leaving Google DeepMind for Anthropic. The thread immediately connects this to Noam Shazeer's recent departure, sparking theories about deep trouble at Google, from bureaucratic rot and safetyism to a talent exodus driven by pre-IPO stock at Anthropic. Several commenters push back on the idea that AlphaFold was brute force, with one clarifying it's "based on deep learning and it's not brute force" and others mocking the misunderstanding. The conversation widens into a split over whether Google is genuinely falling behind on model quality or simply doesn't care to compete on the same axis—one side argues Gemini Flash is good enough for search ads and cheap to run, while the other insists Claude and GPT leave it in the dust on real reasoning and coding, and that Anthropic is assembling an all-time great IC team. A recurring joke lands on the name itself: "John Jumper jumps to Anthropic."
Iran requires insurance on ships using Strait of Hormuz, fees likely to follow [comments]
110 points · 114 comments · www.lloydslist.com · 8h ago
The article reports that Iran has imposed mandatory insurance on ships transiting the Strait of Hormuz, with fees likely to follow after a 60-day grace period, effectively sidestepping a US-Iran memorandum guaranteeing toll-free passage. Hacker News immediately seized on the duplicity of the deal, calling it “tolls after all” and noting the US paid to end a conflict only to let Iran start charging anyway. A heavy split emerged between those who see Iran as having won the confrontation and those who cataloged the US military’s failures—losing an aircraft carrier to a laundry fire, bombing a school, failing to secure the strait—arguing the war was lost the moment the US wasn’t willing to occupy mountainous Iran. Others pushed back on any “mission accomplished” framing, pointing out that the strategic problem is Iran’s nuclear program and that tolls actually make US oil exports more competitive, so Washington may not mind. The thread largely abandoned the insurance details to re-litigate the whole war’s competence, with some calling the entire enterprise a payoff to make the midterms go away.
Companies rein in AI usage as costs strain budgets [comments]
109 points · 93 comments · archive.ph · 12h ago
The article reports that companies like Amazon, Walmart, Uber, and Cisco are capping AI usage and pushing employees to cheaper models because token-based billing has blown through budgets — Uber blew its entire 2026 AI budget by April, and one company’s spend jumped 7x overnight after switching to usage-based pricing. Hacker News largely seized on this as vindication that the AI hype is overblown: the top comment argues AI simply isn’t worth its cost for most use cases, and that since AI labs themselves are losing money, real prices will rise further, making the current spending look like a subsidy-driven bubble. Others pushed back with the math that $1,500/month per engineer is trivial compared to their salary, so even a tiny productivity gain justifies the outlay — but that drew the retort that executives are burning cash out of FOMO and groupthink, not because they’ve measured any real return. Several commenters shared inside stories of internal token-spending competitions that were abruptly killed, and an engineer admitted they now deliberately write vague documentation to avoid making themselves replaceable, reflecting a deeper distrust that AI adoption is really about firing people later. The thread overall split between “this is just a normal cost-optimization phase” and “this is the beginning of the AI hangover,” with nobody sure which side is right.
Zenzizenzizenzic [comments]
100 points · 30 comments · en.wikipedia.org · 10h ago
The Wikipedia article digs up "zenzizenzizenzic," a 16th-century word for the eighth power, coined by the same guy who gave us the equals sign. HN mostly ran with two tangents: the word's Scrabble impossibility (it's 16 letters long and needs four Zs, so you can't even start) and the fact that it just got a fresh wave of attention from a recent episode of "The Rest is Science" with Vsauce and Hannah Fry. A few people pointed out that Robert Recorde's whole naming scheme was a mess—one comment argued that bad naming conventions like this are a "smell" that the problem isn't fully understood. Others veered into comparing it to musical notation oddities like the hemidemisemiquaver, or joking about pharmaceutical branding and Powerade's original formula.
Surprising economics of load-balanced systems [comments]
100 points · 23 comments · brooker.co.za · 11h ago
The article breaks down why adding more servers to a load-balanced system can actually reduce latency (the M/M/c queue effect), countering the intuition that more servers just add capacity. The thread quickly zeroes in on the real-world gaps: the model assumes independent, stateless requests with Poisson arrivals, but in practice, correlated bursts like World Cup traffic or thundering herds break those assumptions entirely. Some pushback is that most cloud load balancers don't even implement the perfect central queuing the analysis depends on—they're stateless and random, so the latency benefit doesn't materialize the same way. Another strong angle is that the whole "pooling is cheap" result only holds for independent work; as soon as servers share mutable state or coordinate, you trade queuing benefits for coordination costs that scale with the number of pairs, not workers. Several people also note that real traffic patterns are non-stationary and non-ergodic, making the theoretical curves a useful starting point but far from a reliable economic story for production systems.
GPT-5.5 hallucinates 3x more than MIT-licensed GLM-5.2 [comments]
96 points · 18 comments · arrowtsx.dev · 15h ago
The article argues that bigger models aren't always better, using benchmark data to show that smaller, open-weight models like GLM-5.2 hallucinate far less than massive proprietary ones like GPT-5.5, suggesting intelligence gains from scaling have plateaued. The HN thread immediately got into a debate about how to interpret those hallucination rates—some pointed out the metric is conditional on the model not knowing an answer, so it doesn't measure everyday error probability, and others pushed back hard, saying a lie is a lie regardless of whether the model "knows" the truth. A few people raised the practical angle that for coding agents, these hallucinations make LLM-written code a ticking time bomb, with small anomalies that compound into unmaintainable enterprise codebases. There was also a split over whether the problem is model size or training methodology, with one commenter arguing these models lack anything like the human amygdala's fear signal, so they never learn to say "I don't know" because their training corpus and RLHF both penalize abstention and reward confident-sounding answers.
Ask HN: Will programmers write more efficient code during the memory shortage?
89 points · 152 comments · news.ycombinator.com · 9h ago
The submission asks whether the current memory shortage will push programmers to write more efficient code and revive interest in advanced algorithms. The thread overwhelmingly answered "no" — many argued that the real driver is abundance, not scarcity, and that users and companies will just pay more for inefficient software rather than optimize. A few commenters noted that server-side memory optimization does get prioritized because it directly cuts costs, but client-side bloat (like Electron apps) has little incentive to improve unless users revolt. There was a split on whether LLMs will write more efficient code themselves or simply perpetuate laziness, and one person at a big tech company reported that their team is already planning RAM reduction goals — though they admitted most gains come from "stop doing stupid stuff" rather than fancy algorithms.
I Stored a Website in a Favicon [comments]
84 points · 27 comments · www.timwehrle.de · 2h ago
The article describes encoding a tiny HTML page into a favicon by turning each byte into a pixel's RGB value, then using JavaScript to decode it back. HN quickly jumped on more practical alternatives—several people pointed out that SVG favicons can just hold the markup directly, and others noted PNG comment chunks or the ICO format's multiple-resolutions allow way more storage, which the author acknowledged as valid. A security tangent emerged about using favicon caching for cross-domain fingerprinting and tracking across private browsing sessions, with some wondering whether that’s been fixed. There was also a side debate about whether the article’s staccato, bullet-point style was LLM-generated, and a few found the prose hard to read—though others argued it’s just short sentences. One commenter linked their own project about storing a stock portfolio in a URL and favicon, calling it a timing coincidence.
Leave a Trace [comments]
74 points · 76 comments · www.jakeworth.com · 19h ago
The post argues that leaving a trace online—dropping a comment, thanking a solution author, or even telling a maintainer why you're leaving their tool—makes the internet more humane. But the thread immediately seized on the irony that the author’s own blog has no comment box and hides a contact form behind a consulting page, making it all but impossible to follow his advice. A few people dug into practical solutions (Disqus, giscus, just publishing an email), while others dove into a tangential but spirited debate about “leave no trace” hiking ethics versus digital traces, and whether the author’s plea is undercut by his own ghost town of a site. There’s a genuine split: some found the post motivating and promised to engage more, while others felt the hypocrisy undermines the whole premise—if you want feedback, make it easy to give.
AURpocalypse now: a look at the recent AUR attacks [comments]
71 points · 43 comments · lwn.net · 15h ago
The LWN article details a large-scale attack on the Arch User Repository where adversaries created accounts, adopted orphaned packages, and pushed malicious PKGBUILDs with npm-based malware and eBPF credential-stealers, hitting over 1,500 packages and forcing the project to disable new-user registration. The HN thread immediately split into two camps: Arch stalwarts defending the AUR as an explicitly unvetted "here's a PKGBUILD" system where the onus is on the user to review diffs, and critics arguing the flat namespace and orphan-adoption mechanic make the trust model uniquely broken compared to Fedora's Copr or openSUSE's OBS, which are namespaced and don't allow maintainer transfers. A practical tangent emerged around recent yay v13+ Lua hooks that can skip recently updated packages during upgrades, though someone correctly pointed out that hook doesn't protect against fresh `yay -S` installs, so it's not a real defense. A few commenters tied the AUR campaign to the broader miasma worm targeting multiple package ecosystems, while another subthread debated whether rolling release distros are inherently less secure than Debian's freeze-and-backport model, with pushback arguing that LTS maintenance is increasingly unsustainable as upstreams stop isolating security fixes.
The UK's new under-16 social media ban will cause more harm than it prevents [comments]
67 points · 76 comments · www.eff.org · 10h ago
The EFF argues that the UK’s plan to ban under-16s from social media by 2027 will backfire—privacy-invasive age verification is unreliable, and kids will lose access to educational content and connections with distant family. Hacker News mostly disagreed with the EFF, with many commenters saying the ban is a net good and that the EFF is ignoring the real harms of algorithmic feeds, grooming, and brainrot. A prominent counterpoint was that the real fix is to ban algorithmically driven social media for everyone, not just kids, sidestepping the privacy nightmares of age gates entirely. Others argued that the government is simply better at enforcing health boundaries than most parents, though critics shot back that this just hands control to a surveillance state and that YouTube isn’t social media—but plenty replied that YouTube’s recommendation engine is just as toxic as TikTok’s.
EU to soon classify AWS and Azure as gatekeepers under DSA [comments]
64 points · 25 comments · www.heise.de · 22h ago
The EU is reportedly about to slap AWS and Azure with gatekeeper status under the Digital Markets Act, citing the massive blast radius of recent cloud outages. The HN thread quickly zeroed in on the irony that many European services are hosted in us-east-1, noting that AWS's own control planes are locked to that single region—so an outage there can take down services actually running in eu-west-1. Several commenters pushed back on the political framing, arguing the DMA is being used to target US tech while giving a pass to European monopolies like ASML; others shot back that the market thresholds for gatekeeper status are clearly defined and that any monopoly, regardless of origin, should face the same rules. A separate technical tangent debated whether inflated egress pricing is the real problem, with one side insisting the Data Act already forces cost-based data portability and the other countering that current pricing still locks customers into single clouds by making multi-cloud setups prohibitively expensive.
Generated 2026-06-20 08:23 UTC
Generated by Sauron from Hacker News discussions and linked articles.