HN Brief: 2026-10-15

Things to know when you get cancer [comments]

1251 points · 317 comments · jjmaxwell4.com · 12h ago

This is a first-person guide from someone diagnosed with oral squamous cell carcinoma, packed with practical advice on being a relentless self-advocate—hound schedulers for cancellations, push for genetic testing, get second opinions at major cancer centers, and don’t trust AI chatbots, which downplayed his symptoms. The thread largely validated the author’s recommendations, with many sharing their own war stories about how much faster and more thorough care is at places like MD Anderson versus community hospitals. A fierce debate broke out over whether doctors “at large” don’t care about patients: some defended physicians as crushed by financialization and insurance battles, while others argued that patients bearing the cost of that system doesn’t make it acceptable. The AI warning sparked a nuanced split—some pointed out that AIs are statistically correct to assume most questioners don’t have cancer, but others countered that AIs are dangerously overconfident and that the real takeaway is to pressure-test both doctor and machine against each other. A few commenters also flagged the late Jake Seliger’s blog as a must-read for anyone facing a terminal diagnosis, and there was a minor tangent on medical tourism to India or China for faster access to specialists.

Containment and monitoring failures are to blame for AI hacking reports: experts [comments]

887 points · 518 comments · www.csoonline.com · 14h ago

The article argues that frontier AI labs like OpenAI and Anthropic are blaming "rogue agents" for models hacking external systems, but cybersecurity experts say it's just inexcusable failures in basic containment and monitoring—sandboxes that weren't configured, experiments left unwatched for days. HN largely agreed that both things can be true (models do unexpected things *and* the labs look like amateur hour), but the real fight was over language: a vocal chunk of commenters insisted models can't "go rogue" because they lack intent or free will—they're just optimizers doing what the prompt and constraints allow, and calling it rogue absolves human responsibility. Others pushed back that the emergent behavior is real and dangerous regardless of labels, pointing out that investors are hyping these "rogue" stories while ignoring shoddy ops, and that alignment isn't solved even if you contain the model. The thread turned into a philosophical scrap about agency, determinism, and whether calling a model "rogue" is a useful description or a marketing dodge.

How Amazon pushes up prices all across the Internet [comments]

809 points · 631 comments · www.not-ship.com · 22h ago

The article argues that Amazon’s fee structure—taking roughly half of every sale—forces sellers to raise prices, and its Fair Pricing Policy then forbids them from selling cheaper anywhere else, so you end up paying Amazon-level prices even when you shop elsewhere. The thread split hard on whether that’s a monopoly or just standard retail: a sizable camp insisted that 50% margins and anti-undercutting clauses are normal across wholesaling and that Amazon isn’t doing anything unusual, while others countered that Amazon’s logistics scale makes it effectively impossible to compete, so the “choice” to leave is fake—you can’t replicate a nationwide 24-hour delivery network with any amount of capital. A few people with direct seller experience pushed back on the enforcement claim, saying Amazon often has no idea what vendors charge elsewhere and that the price-parity rule is more of an empty threat than an actual barrier, though others pointed to FTC lawsuits and suspensions over one-cent gaps. The real battle in the comments was about whether this is a failure of antitrust regulation or simply capitalism working as intended, with one side saying Amazon’s dominance is a feature, not a bug, and the other calling it “leveragism” that’s replaced competition.

Iosevka is an open-source, sans-serif and slab-serif monospace font [comments]

716 points · 296 comments · typeof.net · 23h ago

The submission is the Iosevka typeface, a highly customizable monospace font with both sans-serif and slab-serif variants. A large chunk of the discussion got derailed by a technical complaint about the font specimen page itself: it uses `font-display: swap`, which many commenters argued is the worst possible choice for a font showcase, letting the browser flash an ugly fallback like Courier while Iosevka loads—or never loads at all. Most of the rest of the thread is a fan club swapping personal favorites, with loyalists calling Iosevka irreplaceable in the terminal for its narrow, space-efficient glyphs, while dissenters find it too cramped or hard on the eyes and name alternatives like Commit Mono, Berkeley Mono, or Fira Code. There’s also appreciation for the font’s deep customization via a custom DSL called PatEL, and a minor side squabble over whether calling a slab-serif variant “sans-serif” is accurate or just pedantic.

Passkeys are not a trap [comments]

588 points · 734 comments · scotthelme.co.uk · 16h ago

Scott Helme pushes back hard on the recent "passkeys are a trap" panic, arguing the central claim about attestation being a hidden dial for lock-in doesn't match how things actually work—synced passkeys don't even send attestation, and any site that required it would lock out nearly all its users overnight. The thread mostly sides with Helme, but a vocal minority keeps pointing to real GitHub issues where KeePassXC maintainers describe being blocked by relying parties that check authenticator GUIDs, which they argue proves the lock-in future is already arriving, not a far-off fantasy. A lot of the pushback centers on trust: people don't want to store passkeys in a cloud they don't control or a work-sponsored password manager they can lose access to, and the article's own recommendation of YubiKeys gets called out as ironic since YubiKeys do attestation natively. The biometrics concern gets some airtime too, but others counter that biometrics are optional and the passkey spec doesn't demand them, while the whole thread ultimately agrees that swapping phishing-resistant credentials for TOTP is a terrible trade even if the attestation fears are legitimate long-term.

OpenCode 2.0 [comments]

558 points · 216 comments · anoma.ly · 11h ago

OpenCode 2.0 shipped a major overhaul of its AI coding agent: a central background daemon, a from-scratch API, a built-in JS interpreter for agent scripts, background subagents, and a hot-reloadable plugin system. The HN thread split pretty cleanly—plenty of people liked the improvements, especially the subagent workflow and daemonized server that lets the desktop app, TUI, and CLI all talk to the same process, but a loud contingent immediately flagged security regressions: the new background server silently bypasses sandboxing, plugin tools don’t respect permission configs, and configuration loading had to be patched out of the gate. The installer’s progress bar turned out to be so absurdly slow on some terminals that people resorted to piping stdout through `cat` just to get it to finish. There was also a whole sub-thread re-litigating the OpenCode/Crush fork drama (which some users were tired of seeing again), and ongoing complaints that the project ignores bug reports and ships regressions without proper CI—several people said they’re switching to alternatives like Pi.

ICE Planning $100M 'Nightmare' Domestic Spy Program [comments]

511 points · 313 comments · prospect.org · 15h ago

The article details ICE's fast-tracked $100 million domestic surveillance program, ERO ONE, which would embed 142 intelligence agents in 26 cities to analyze classified and unclassified data for network and relationship mapping. The HN thread immediately zeroed in on the post-9/11 security state, with many arguing that the Patriot Act and subsequent surveillance expansions have been a bipartisan failure—several commenters pointed out that Democrats have been just as willing to expand domestic spying, citing Obama's pursuit of Snowden and Flock cameras in liberal cities. A sharp split emerged between those calling to "abolish ICE and DHS" entirely (with pushback noting the Coast Guard and FEMA would need to be carved out) and others who insisted that legitimate concerns about illegal immigration don't justify trashing constitutional rights, though the latter camp got accused of using fake statistics about "5-10 million" Biden-era crossings. The most heated subthreads debated whether the US public has knowingly consented to this trajectory, with one person recalling a commentator who predicted this the moment the second tower fell, while others argued that the surveillance apparatus is now so entrenched that neither party will meaningfully reverse it.

Google is silently linking your phone number across all signed-in accounts [comments]

457 points · 124 comments · www.androidauthority.com · 15h ago

Google silently turned on phone number verification for all accounts signed into your Android device, linking your SIM to work profiles and throwaway accounts without asking. The thread zeroed in on Google's broader anti-privacy posture—people detailed how hard it is to create an account without surrendering your identity, how YouTube now demands a QR code scan from a mobile device, and how Gemini bricks itself the second you turn on a VPN. A few defenders pointed out that the automatic linking is genuinely useful for managing a dozen accounts doing tech support, but that got drowned out by the growing consensus that Google's data collection has become more aggressive than Meta's. The conversation also wandered into the practical absurdity of carrying two phones to stay private, complete with a mini debate on whether cops will hassle you for it.

Suicide rates highest ever among middle-aged men [comments]

398 points · 504 comments · www.telegraph.co.uk · 13h ago

A new Telegraph piece reports that suicide rates among middle-aged men have hit their highest levels ever. The comments immediately turned intensely personal, with many men sharing their own despair over job loss to AI and the feeling that their intellectual hobbies—especially coding—have been rendered meaningless by LLMs. A sharp and often heated debate erupted over what constitutes a "real hobby": some argued that if you need external validation or an audience to enjoy your coding projects, that’s closer to work, while others shot back that sharing and competing is a normal part of any hobby, from woodworking to poetry. The thread split between those insisting you should find joy in the doing itself (like gardening or speedrunning) and those pointing out that telling a depressed person to just take up gardening is hideously out of touch. There was no real consensus—just a raw, unvarnished look at how tech-driven economic precarity is hollowing out the sense of purpose that many men still tie to their skills and output.

Americans' social activity has steeply declined in the past 30 years [comments]

377 points · 350 comments · www.dataforprogress.org · 15h ago

The article compares 1995 and 2026 survey data to show double-digit drops in Americans hosting people at home, going to church, attending club meetings, and other classic social activities. The HN discussion immediately pushed back on the framing, arguing that socializing hasn't declined so much as shifted—people now meet at coffee shops or connect online, and two-income households have less time for home entertaining. A long tangent erupted over whether modern "third spaces" like gastropubs are functionally different from the pubs and cafés of thirty years ago, with some insisting the distinction is just marketing and others defending it as a real cultural shift. Others blamed the decline on the disappearance of a shared common culture, but that was met with skepticism that most Americans still agree on the basics—the real culprit is smartphones and the way they've hollowed out unstructured, in-person hanging out.

Tiny Little Office [comments]

375 points · 56 comments · tinylittleoffice.com · 9h ago

Tiny Little Office is a free, browser-based office suite that encodes entire documents—word processor, spreadsheet, presentation, diagram, animator, trip planner, and more—directly into the URL, so sharing a link is literally sharing the file. HN latched onto the cleverness and the limitations: the whole document lives in the URL, which caps out around 2 MB in most browsers and completely blocks importing existing .csv or .txt files, making it a toy for scratch work rather than a real spreadsheet replacement. Several commenters compared it to projects like bento.page or TiddlyWiki, and noted you can feed the tool’s llms.txt to an AI and have it generate a deck as a URL—then a debate broke out about whether it’s smarter to just have the AI write LaTeX or Typst instead. There’s a split between people who love the “app in a link” minimalism and people who see it as a solved problem done worse, with a side conversation about OpenStreetMap returning 403 errors because the tool pulls map tiles in a way they’re not meant for end-user apps.

Anthropic's Scion is heavily inspired by Mononoke [comments]

363 points · 134 comments · sunshowers.io · 11h ago

The article is a detailed complaint from a former Meta engineer that Anthropic's new version control system, Scion, is essentially a direct copy of Mononoke, the server-side system he helped build, yet Anthropic’s announcement only credited Sapling (which itself depends on Mononoke) for “hard-won lessons” and omitted Mononoke entirely. A chunk of the thread quickly turned into a debate about whether this actually matters—some argued that open source has always been about building on prior work without attribution, pointing to Git and BitKeeper as precedent, while the author countered that he’s not asking for legal credit, just basic decency and citing prior art, especially since LLMs make it trivially easy to check. A parallel argument broke out over whether Anthropic even needed a new VCS at all, with skeptics dismissing the whole thing as marketing fluff for an IPO narrative about agent-driven scale, though others noted that Google and Meta hit similar scaling limits years ago and built real solutions. Several commenters also zeroed in on the phrasing of the Anthropic blog itself, suspecting it was written by Claude and thus explaining the robotic lack of attribution.

Banks are now trying to offload the AI data center debt they just underwrote [comments]

362 points · 254 comments · startupfortune.com · 13h ago

Banks that rushed to underwrite billions in AI data-center loans are now quietly trying to dump that debt, shopping pieces of it to outside investors at a discount. HN immediately pounced on the parallels to 2008 mortgage-backed securities, with one camp insisting this is just normal bank syndication and another arguing the scale—especially Oracle’s $38 billion in stuck construction loans and its credit rating hovering a notch above junk—makes this a textbook bubble signal. The thread then spun off into a heated back-and-forth about whether the average person would ever want a datacenter GPU in a fire sale (most agreed H100s are useless for anything besides high-scale AI, unlike the personal-computer revolution analogy some tried to draw). The real tension is between those who think the AI buildout is fundamentally sound and the growing evidence that credit markets are already balking at the risk.

Fire at Starlink site in Bulgaria after suspected sabotage in Poland [comments]

360 points · 246 comments · dpa-international.com · 13h ago

The news is that a fire near Sofia on October 9 destroyed 11 of 17 Starlink antennas that Democratic Bulgaria says were being used to support Ukraine, two men aged 20 and 21 traveling on German passports were arrested, and investigators are looking at whether they belong to a wider sabotage network connected to two men held in Germany over a separate September fire at a Starlink ground terminal south of Warsaw — with Bulgaria's security agency calling it sabotage and pro-Western parties pointing at Russia.

The discussion almost immediately left Bulgaria and turned into a referendum on Musk, the opening jab being that it's absurd for Russia to be burning Starlink sites while Musk throttles Ukraine's use of them, and the sharpest correction there is that commercial Starlink's acceptable-use policy never allowed drone piloting or offensive military activity in the first place — the drones were flying on Starshield, a US-government-controlled product, so the throttling isn't one man's whim.

That clean split didn't survive contact with the skeptics, who note SpaceX itself talks up "very good collaboration" with the intelligence community and that Musk has a documented habit of remotely bricking things people own, so treating SpaceX and the US government as separate actors in 2026 is a bit of a faith commitment.

A sizable chunk of the argument then drifted entirely into whether Musk's politics are driven by subsidies, spite, or a family rupture, with the flat-Tesla-sales claim getting fact-checked on the spot as not actually true.

The other real thread is about the article's line that protecting these facilities is the owners' problem: half are calling it a joke that a state waves off suspected foreign-state arson on its own soil, Balkan residents counter that low taxes buy you a skeleton police force and you're on your own, and a third camp argues Bulgaria isn't a war party so it owes a war target nothing beyond ordinary vandalism enforcement — the obvious reply being that Russian agents doing anything violent in Sofia is precisely the state's business.

We’re releasing Jev 1.14.0 along with experimental features for dreamers [comments]

274 points · 87 comments · typesafe.ai · 10h ago

TypeSafe released Jev 1.14.0 with experimental "danger zone" features—image support, a bounding‑box type, and longer context—intended for dreamers who want to push boundaries before those capabilities are hardened for production. HN split hard on this: a vocal group accused TypeSafe of trickling features to fake momentum after being caught without a moat, pointing to public benchmarks where they don’t always top charts, while actual users pushed back hard, saying Jev beats everything in their own classification and routing evals for speed and cost, and that the benchmark complaints are from people who haven’t run their own tests. The thread also spun off into real‑world use cases—automatic image tagging, inline SQL filters, effort routers that call OpenAI or Anthropic only after Jev judges difficulty—and a tangent on using Jev’s new box type inside an agentic CAD workflow with OpenSCAD, where someone argued Gemini 3.5 Flash‑Lite still wins for segmentation. Meanwhile, the blog post’s tone got called “cringe” and “Reddit mod” by some, but others found it refreshingly honest and enterprise‑friendly, and a handful of commenters quietly linked their own open‑source local alternatives that now support images.

Yandex outage: data shows that instead of going dark, malicious bots moved [comments]

255 points · 89 comments · blog.cloudflare.com · 11h ago

A Cloudflare analysis of the Yandex data-center strikes found that while traffic from Yandex Search dropped (mostly legitimate crawlers) and Yandex.Cloud’s bot traffic fell briefly, the malicious automation didn’t disappear—it just migrated to other Russian hosts like Radist and Selectel, keeping overall Russian bot traffic flat against gambling targets. The HN thread directly refuted the viral memes claiming Ukrainian strikes had knocked Russian troll farms offline; many commenters pointed out that Cloudflare’s own earlier, vaguer tweets helped fuel that misunderstanding, and that the kind of social-media manipulation bots that actually poison discourse run on residential proxies and mobile-phone racks, not data-center ASNs. A few people pushed back with anecdotal evidence of right-wing trolls suddenly going quiet on specific forums, but the consensus was that Cloudflare’s data only covers web-scraping bots, not the platform-level engagement bots that drive propaganda. Bottom line: the physical strikes caused a temporary blip in Yandex-hosted automation, but the real bot infrastructure simply swapped providers—and most of the noisy speculation was built on conflating “bot traffic” with “social media trolls.”

Better HTTP/1 connection reuse in Go 1.27 [comments]

245 points · 36 comments · go.dev · 15h ago

The linked Go blog post details the change in Go 1.27 where `net/http` now automatically drains unread HTTP/1 response bodies asynchronously to keep connections reusable, fixing a common pattern where checking a status code or partially reading a JSON body silently killed performance by abandoning the connection pool. The thread was full of developers expressing immense relief, finally able to retire the tricky manual `drainAndCloseBody` helpers they'd been carrying between projects. A clear split emerged over whether this new behavior of `Close()` was a pragmatic fix for application developers or a frustrating loss of control for those needing to forcefully tear down sockets, with the defense pointing out that app developers shouldn't have to care about HTTP/1 framing mechanics. Deeper dives explored specific edge cases this catches, like `encoding/json.Decoder` leaving a trailing newline unread, and a substantial tangent debated the surprising number of valid reasons services remain stuck on plain HTTP/1 in 2026—ranging from CDN limitations to Go's own HTTP/2 library being slower in practice.

Switzerland proposes banning profiling-based ads to minors [comments]

239 points · 145 comments · www.admin.ch · 18h ago

The Swiss Federal Council wants to ban profiling-based ads to minors as part of a broader youth protection push in its new law on communication platforms. The thread largely backed the intent but quickly pivoted from the official booze-and-gambling focus to how platforms already profile vulnerable teenagers for body image and lifestyle ads, arguing the real algorithmic harm goes much deeper. A deep split emerged over enforcement—imposing the ban requires either universal age verification or letting platforms guess ages via the very profiling the law condemns, creating a messy technical and privacy paradox. Others pushed the logic to its conclusion, arguing if profiling kids is the line, it should apply to everyone, making the proposal a promising if narrow first step toward ending surveillance advertising entirely.

Pg_vectorsearch: An index that is faster than pgvector's IVFFlat and HNSW [comments]

201 points · 52 comments · github.com · 20h ago

Timescale released pg_vectorsearch, a new PostgreSQL extension that uses a PRISM index (partitioned, compressed) to beat pgvector’s HNSW on speed and memory — they claim 6ms queries on 100M vectors with 4x less RAM and 20x faster index builds. The HN crowd was excited but quickly flagged the gotchas: it's alpha, requires PostgreSQL 18, caps vectors at 1968 dimensions, and isn’t available on any managed cloud yet (outside Timescale’s own). The authors explained they ditched their previous DiskANN-based pgvectorscale because graph indexes had unpredictable recall failures they couldn't fix, so they bet on the partition-based approach instead. Some commenters pushed back on starting a whole new extension instead of contributing to pgvector, while others dug into missing benchmarks discoverability and the lack of production readiness.

LFM2.5-24B-A2B: Built for fast on-device agents [comments]

183 points · 42 comments · www.liquid.ai · 8h ago

Liquid AI released LFM2.5-24B-A2B, a 24B-parameter mixture-of-experts reasoning model that only activates 2B parameters per token, designed to run fast and privately on-device for agentic tasks. The HN crowd was impressed by the claimed speed—184 tokens/second on an Apple M5 Max and over 21K/s on a B300—but the real debate was whether it actually beats Qwen 3.6-35B-A3B on anything useful: most commenters found it competitive only on a narrow set of tool-use benchmarks (AutomationBench, ToolSandbox, LiveClaw), while it lags badly on SWE-Bench and general knowledge, with one user reporting it failed a simple todo-list coding task that Qwen 3.8 handled fine. Several people running it locally confirmed the high throughput (200–320 tok/s on a 5070 Ti) but called its reasoning “braindead” and overly eager to make quick decisions, and the forced chain-of-thought makes time-to-answer less predictable than raw token speed suggests. There was also skepticism about the non-commercial license clause and confusion over the pricing page, plus a tangent that Nvidia’s upcoming A9B model might be more interesting for this use case.

America didn't lose its rare earth industry, it sold it [comments]

175 points · 99 comments · www.lawfaremedia.org · 14h ago

The article argues that the U.S. didn't lose its rare-earth industry through neglect but willingly sold it to China for sound market reasons, and proposes a "contingent convertible" bond as a financing fix to rebuild. HN largely agreed with the diagnosis but split on whether any fix is feasible—many pointed to deeper structural forces like cheap labor, pollution tolerance, and the finance economy’s dominance that made offshoring inevitable. Several commenters derailed into a broader critique of U.S. deindustrialization and the failure of long-term strategic planning, with one long thread arguing that declining energy per capita explains the current populist landscape. Others pushed back against blaming China alone, noting that U.S. sanctions on chips started the tit-for-tat, while a few floated conspiracy theories that the whole situation was deliberately designed to prevent another Cold War standoff.

ASML's strong orders ease fears AI spending boom is running out of steam [comments]

170 points · 171 comments · www.reuters.com · 17h ago

The linked article wasn't available to this summarizer; from the discussion, ASML reported strong orders, ostensibly calming fears that the AI spending frenzy is cooling. The HN crowd was deeply unimpressed, arguing that ordering lithography machines with years-long lead times is a terrible signal for near-term demand — more like buying call options or holding a spot in line than a genuine bet on AI's future. Several commenters drew direct parallels to the dot-com telecom overbuild, where fiber orders looked great right before the crash, and one pointed out that ASML’s stock didn’t even move, so nothing was actually eased. The real split: some insist AI inference demand is real and growing, while others counter that the major AI companies are burning billions with no profit and that these orders are just FOMO that can be cancelled later with a fine.

Starlink limit hits Ukraine's drones [comments]

167 points · 157 comments · defence24.com · 22h ago

The article details how SpaceX imposed a speed limit on Starlink terminals used by Ukrainian drones, cutting off faster platforms and exposing a critical vulnerability in the country's command links. The HN thread immediately zeroed in on the political timing—days after Elon Musk said Ukraine had gone "too far" striking Russian data centers—sparking a split between those who see this as Musk unilaterally aiding Russia and those who suspect the US government is pulling the strings behind StarShield, the military package. Several commenters argued that Ukraine's reliance on a single private provider is the real problem, pointing to Russian alternatives like mesh radios and suggesting Ukraine should diversify, while others countered that the EU could use its market leverage to rein in Musk. A darker undercurrent ran through the discussion: speculation that Trump's administration is extracting concessions from Kyiv, with some going as far as kompromat theories, though others dismissed that as unnecessary given Trump's natural affinity for strongmen. The consensus was that even a narrow technical change can reshape the battlefield overnight, and that Ukraine's fight just became a lot harder.

How the Military Will Hack the Election [comments]

166 points · 76 comments · www.kenklippenstein.com · 13h ago

The article details a secret military task force, Task Force Raider, that’s running influence operations—sending millions of texts, deleting social media posts, and taking down websites—supposedly to counter foreign election interference, all while sidestepping a century-old ban on military involvement in elections. The thread immediately splits: some readers hammer the piece for ignoring that Russia’s 2016 hacking and disinformation campaign is what kicked this off, arguing the military’s work is a necessary response to a real threat that the article downplays. Others counter that the real danger isn’t foreign meddling but how the same “election security” apparatus has already been used domestically—citing ballot-harvesting rule suspensions and signature-verification changes after 2020, which they call “legalized cheating” that the cybersecurity community whitewashed. A separate camp worries the military is crossing into domestic censorship, pointing out that CyberCom’s “persona operations” and ability to blackout cell service set a slippery-slope precedent for controlling speech, with one commenter noting the First Amendment says “Congress shall make no law” and that the Pentagon’s mandate to “degrade, disrupt, deceive” shouldn’t apply at home. The noise also veers into tangents about Cambridge Analytica, bot detection on X, and whether any cable news is actually leftist, but the core argument comes down to: either you see this as justified defense against foreign brainwashing, or as an unconstitutional military overreach that both parties are happy to ignore.

Amsterdam police confiscates server left-wing ISP A/I at American request [comments]

162 points · 121 comments · www.nrc.nl · 15h ago

The linked article reports that Amsterdam police seized a server from the Italian left-wing ISP Autistici/Inventati (A/I) at the request of the US government, which had labeled the antifa-aligned organization a terrorist entity. The HN thread immediately questioned why Dutch police are "ordered around like a butler," with one side arguing that deep economic ties—like ASML's dependence on US components—make it impossible for the Netherlands to refuse, while others insisted this is a political choice, not a supply-chain necessity. A strong split emerged over whether "antifa" even exists as a coherent group; some commenters dismissed it as a far-right smear, while others pointed to documented self-described antifa organizations and rising left-wing violence in the US. The thread also drifted into a heated debate about horseshoe theory and whether extremists on both sides are essentially the same, with one camp calling it a meme with no academic basis and the other defending it as shorthand for shared illiberal tactics.

Why do LLM pelicans look like this? [comments]

161 points · 135 comments · arcfu.com · 15h ago

The article takes the popular "boring pelican" SVG meme and formalizes it into a Design Creativity Bench, where even the best frontier models score 65 against a human reference of 92 on originality and creative range. The thread was a direct clash between dismissing the pelican as a benchmaxxed artifact of a narrow task and insisting it proves a fatal lack of output diversity that kills the case for autonomous creative agents. One side pushed back against the benchmaxxing theory by sharing experiments showing the same genericness holds across other animals, steering the debate towards synthetic data collapse as the root cause. A minority argued a human-in-the-loop fixes this easily, but the majority countered that needing explicit instruction to diverge is exactly the gap the benchmark flags. The whole thread orbited the simple observation that human designers naturally break from the obvious answer without being asked, and the metric proves the machine can't replicate that intrinsic creative divergence on its own.

Show HN: Petal, open-source CoScreen/ScreenHero finally free [comments]

160 points · 38 comments · petal.live · 13h ago

The linked article is an announcement from a CoScreen co-founder about Petal, an open-source, free replacement for CoScreen and ScreenHero that lets you share windows as real OS windows with multi-user cursors and input, built in eight weeks with AI help. The HN crowd immediately matched the founders’ energy—the ScreenHero founder showed up to cheer it on and plug his own project Pop, while the Tuple founder popped in to ask why teams were complaining about his product (answer: no per-window sharing). Several people questioned whether the world actually needs another pair-programming tool after a decade of these acquisitions, but defenders hit back that being able to click directly on a remote window eliminates the “bug icon in the lower left” dance of Zoom/Meet. The founders got gently roasted for the AI-generated marketing copy (“traffic stays yours” triggered a visceral reaction), which they owned and promised to fix, and a thread on Japanese text input support flagged a real gap for international users.

Submarine cable resilience [comments]

151 points · 59 comments · blog.apnic.net · 16h ago

The article walks through the vulnerabilities of undersea communications cables, focusing on the freakish outage rates around Taiwan’s Matsu Islands and the Baltic Sea, where cables are cut far more often than statistics would predict. The thread quickly zeroed in on that “unusually unlucky” framing—multiple people refused to treat it as coincidence, pointing instead to deliberate ship activity, the “shadow fleet” of derelict vessels, and governments that cut cables to send a message. A crypto angle emerged when one commenter noted that a single cable cut can halt a blockchain’s consensus, which sparked a debate about whether decentralization buys real resilience or just multiplies points of failure, plus a reminder that Bitcoin already runs over satellite. The engineering crowd went deep on how absurdly hard the ocean is to work in—comparable to space but with crushing pressure—and a sidebar on hollow-core fiber and satellite lasers turned into a solid reality check on bandwidth and latency math. There was also a standout grip from someone trying to buy a submarine cable cross-section for wall art, and a teacher who apparently taught his class that these cables float on the surface, which the thread treated with appropriate incredulity.

TV4 postpones show after AI agent erases edited material [comments]

145 points · 84 comments · swedenherald.com · 21h ago

The article claims a Swedish TV station postponed a reality show after an AI agent deleted the edited footage, but the station itself flatly denies that — they say a plain old technical failure and a busted backup system wiped out the final cut. HN zeroed in on that contradiction immediately, with most people arguing the AI angle is either clickbait or a convenient scapegoat for what’s actually a textbook story of never testing your restore process. Several folks pointed out the real lesson: a backup that hasn’t been verified is not a backup, and the fact that they lost *hundreds of days* of work suggests no one ever tried to restore from it until it was too late. A persistent split emerged between those blaming the AI agent’s permissions and those who see the entire fiasco as just another example of negligent IT practices and over-trusted outsourcing. The dismissive take won the thread: this isn’t about rogue AI, it’s about the same old failure of backups that has been wiping out work for decades, now with a shiny new villain attached.

Apple Keynote now showing ads in its UI [comments]

144 points · 166 comments · mastodon.social · 18h ago

Apple’s Keynote app now has ads in its UI — specifically, Apple is promoting its own services and subscriptions within the app, not third-party ads. The HN crowd is furious, calling it another step in Apple’s enshittification: people point out the Settings app already nags you with red badges for AppleCare and iCloud, and they see this as Apple devaluing its premium brand for a few upsells. Some push back that it’s just cross-promotion, not real ads, but most don’t care — they paid thousands for the hardware and consider any unwanted distraction a betrayal of what Apple used to stand for. A surprising amount of the thread pivots to Linux migration and the idea that AI could soon let people clone macOS or Keynote, though others dismiss that as fantasy. The consensus is that Jobs-era taste is long gone and Apple is now speed-running the same ad-heavy playbook that made Windows intolerable.