HN Brief: 2026-09-11
Today’s HN was shaped by two hard-to-ignore tensions. The first is the gap between AI’s advertised potential and the messy reality of using it in production: Shopify argues that LLM coding agents make native development cheaper than maintaining cross-platform wrappers, while Armin Ronacher’s tale of a 75k-line slop factory and the SWE-2 model’s catastrophic score drop on a harder benchmark suggest the tooling isn’t ready for prime time. The second throughline is the crumbling bargain between users and the companies they pay—OpenAI silently re-enables the “allow training” toggle, Sony’s lawyers claim no reasonable person thinks they own a digital game, and streaming baskets have surged 61% in three years, all feeding a mood of institutional distrust. A lighter but no less earnest current runs through the nostalgia-and-cables thread and Raymond Chen’s reservoir-sampling deep dive, where people just want to enjoy old tech without someone trying to sell them something.
Threads worth clicking: “Shopify is moving from React Native back to Swift and Kotlin” for the argument that AI flips the cross-platform cost equation and the side-debate over App Store review times; “Don’t let anyone take away your big box of cables” for the surprisingly fierce taxonomy war between hoarders and minimalists; “Tell HN: OpenAI keeps re-enabling the ‘allow training’ setting” for the dark-pattern-versus-incompetence question and whether a local-storage toggle means anything; “What algorithm did Windows XP use to choose your initial user picture?” for the reservoir-sampling trivia and the Luna UI nostalgia fracture; and “The Deathray: A simple way for an untrusted site to freeze a Mac” for a WebGPU shader that kernel-panics your machine and the argument over whether that’s a security bug or just an infinite loop.
Shopify is moving from React Native back to Swift and Kotlin [comments]
1009 points · 687 comments · shopify.engineering · 17h ago
Shopify's engineering post says they're abandoning React Native for native Swift and Kotlin, arguing that LLM coding agents have collapsed the "build every feature twice" cost that justified cross-platform back in 2020. The thread split between people who see this as the inevitable play for every large org now that AI writes the boilerplate, and skeptics who read it as internal politics or a cover story for having gutted their native engineering bench years ago. A long side-debate flared over whether App Store review times have secretly ballooned again — one side claiming week-long waits are now normal, the other pointing to sub-24-hour turnarounds and calling the data anecdotal noise. Cross-platform defenders kept dragging the conversation back to Kotlin Multiplatform and OTA-updatable webviews, arguing Shopify's scale means the two-native-codebases cost barely registers for them while it would crush a startup. Others just wanted specifics the post dodges: whether there's a shared core between the iOS and Android apps, whether the rewrite is SwiftUI or UIKit, and whether Helix gets open-sourced.
Rust is tier-1 language at Microsoft [comments]
660 points · 393 comments · rustfoundation.org · 18h ago
Microsoft officially designated Rust as a Tier-1 language internally, joining C++, C#, and TypeScript, and announced `rustc_codegen_utc`—a new codegen backend that hooks Rust into the MSVC toolchain for seamless Windows interop. The thread immediately pivoted to debugging: people wanted tier-1 debugger support in Visual Studio proper (not just VS Code), and the conversation spun into a broader debate about whether debuggers are even relevant anymore, with some arguing LLMs and print statements have rendered them obsolete and others pushing back hard. A significant side discussion questioned whether Microsoft will open-source the MSVC codegen backend—yes, that’s the plan, but no firm details yet. There was also a split on Rust’s suitability for web backends and rapid prototyping: a few people reported huge memory and latency wins moving from TypeScript to Rust, while others argued Go or FastAPI fit those slots better and that Rust’s compile times and ergonomics kill iteration speed.
Don't let anyone take away your big box of cables [comments]
473 points · 324 comments · blog.jim-nielsen.com · 16h ago
The article is a short, heartfelt defense of hoarding old cables—the author prints out a screenshot of someone celebrating their "Big Box of Cables" and tapes it to their own box as a permanent reminder not to purge it. Hacker News immediately turned this into a surprisingly deep taxonomy debate, with people proudly cataloging their multi-box systems sorted by "USB," "AV," and "Networking & PC Accessories," and challenging each other to find categories with zero overlap. A significant split emerged between the maximalists who insist you're not serious if you only have one box and the minimalists who recently consolidated and feel 90% good about it, though several commenters noted that every time they try to declutter, someone's ancient printer or friend's dying gadget immediately needs that exact cable. The thread also veered into practical tips—grouping cables in labeled ziploc bags to spot duplicates, replacing most things with USB-C plus adapters, and one person who cashed in 10 kilos of scrap copper for €30, calling it his retirement fund.
Tell HN: OpenAI keeps re-enabling the 'allow training' setting
452 points · 181 comments · news.ycombinator.com · 18h ago
The post is a heads-up that OpenAI keeps silently re-enabling the "allow training" toggle after the user disables it, and it tells people to go verify their own settings. The thread split between deliberate dark pattern and plain incompetence, with one person noting the toggle only sets a local storage value that doesn't seem to affect anything, which fuels the "it's just a broken UI" theory. A legal tangent argued whether you can cryptographically prove you unchecked the box—TLS session recording and court-grade eDiscovery got batted around, but the counterpoint was that company-side logs are the only proof that matters and they're easily manipulated. Others said don't bother, just switch to a competitor, which sparked a fight over whether Anthropic is any less shady, with reports that Claude Code repeatedly re-asks for data sharing using the same kind of dark-pattern prompts. The thread also dragged in the recent researcher/IP theft scandal, though several people pushed back that there's no actual proof OpenAI trained on those private chats.
I have a theory that software drives people insane [comments]
434 points · 176 comments · graybeard.ing · 15h ago
The article argues that software’s unnatural lack of friction — the ability to constantly change direction with no visible cost — systematically erodes developers’ sense of proportion, turning mundane work into a neurotic frenzy of endless levers and pretend urgency. A chunk of the thread immediately veered into arguing about the hand-washing example in the opening paragraph, with some insisting that regular hand washing is actually rational and others pointing out the article’s false premise that developers were ever sane to begin with. The dominant pushback centered on the cure: multiple people argued the real madness comes from siloing developers away from customers, and that forcing regular, in-person user contact grounds all that abstract tinkering in actual problems — “it’s like cheating, because uncertainty fades away.” Another strong line of critique took aim at over-reliance on analytics, with several commenters recounting how metrics like “feature X is rarely used” get misread to justify removing essential but infrequent workflows, and how proper usability research consistently contradicts what the dashboards say. The split is less about whether the insanity exists and more about whether it’s an inherent property of software or a symptom of broken team structures and data-driven decision-making that could be fixed by talking to actual humans.
Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra [comments]
405 points · 172 comments · cognition.com · 16h ago
Cognition's SWE-2 post is an engineering deep-dive on a coding agent post-trained from Kimi K33, using Pareto-informed cost penalties in RL to advance the entire cost-performance frontier in a single training run rather than optimizing for peak score alone. HN's main reaction was skepticism, with the standout criticism being the huge gap between its 92.8% on Terminal-Bench 2.1 and 27.3% on the newer Terminal-Bench 4 — read as evidence the model is "benchmaxxed" and doesn't generalize, a charge defenders countered by noting even Sol falls off hard on TB4 and that TB2.1 is saturated. The fact it's built on a 2.8T-parameter Kimi model spawned a side argument over whether that's properly called distillation or fine-tuning, plus a broader gripe that Cognition is repackaging Chinese lab work and will be in trouble once those labs stop supplying base weights. Several people also pushed back on the delivery model: it's only accessible through Devin's CLI or desktop app, not a standard API, which annoyed anyone who wanted to run it through their own harnesses even with a month of free CLI access.
List of references on Sony websites to players "owning" their digital games [comments]
391 points · 130 comments · consumerrights.wiki · 19h ago
California PlayStation buyers are suing Sony for advertising digital games with "Buy Now" buttons and ownership language while actually granting only a revocable license, citing the state's digital-goods transparency law. Sony's lawyers moved to compel arbitration and, in a separate merits argument, claimed no reasonable consumer would believe they owned a digital game — using the absurd hypothetical that if one plaintiff owned Resident Evil Requiem, no other player could buy it. The HN thread immediately tore that argument apart, pointing out that multiple people owning separate copies of the same thing is how property works for books, DVDs, and bags of chips, and that Sony's own support pages and marketing emails casually refer to "games you already own." The broader discussion shifted to the real issue: digital stores charge the same $70 for a worse product that can't be resold, lent, or played if the storefront goes dark, and many commenters argued that until digital purchases come with true transferability and offline longevity, consumers are getting ripped off regardless of what the license agreement says.
The same nine streaming subscriptions cost $702/year more than in 2021 [comments]
375 points · 371 comments · honestlyranked.com · 21h ago
The article tracks nine major streaming subscriptions and finds the same basket costs $702 more per year now than in March 2021 — a 61% jump, led by Apple TV+ at +200% and Disney+ at +138%. The person behind it explained they deliberately used March 2021 as the baseline because an earlier version summing launch prices produced a "bigger, better headline" for a basket that never actually existed. The thread immediately turned on whether the number was honest without inflation adjustment — one side saying real-terms increase is only ~26% and still below what cable used to cost, the other arguing that subscription prices are themselves a big part of measured inflation, so adjusting gets circular. A lot of the discussion drifted to alternatives: physical media, libraries, piracy, and home-server setups like Jellyfin, with one recurring claim that "buying" digital movies doesn't guarantee future access anyway. There was also some suspicion about the post being AI-written and the author's comments getting auto-killed, which became its own sidebar.
What algorithm did Windows XP use to choose your initial user picture? [comments]
349 points · 175 comments · devblogs.microsoft.com · 23h ago
Raymond Chen's latest Old New Thing post reveals that Windows XP used reservoir sampling to pick a user's initial profile picture—a one-pass random selection algorithm that reads files once instead of twice, seeded with `GetTickCount()`, with a hard cap at 100 files to prevent pathological slowdowns. The HN crowd mostly enjoyed the technical deep-dive, but quickly veered into a full-blown nostalgia war over Windows XP's Luna UI, with camps splitting sharply between those who loved its rounded, colorful "Fisher-Price" look and those who immediately switched to the Windows 2000 classic theme as unprofessional. Some commenters pushed back on the algorithm itself—one suggested a simple `rand() % count` would have sufficed, but others pointed out you can't mod by an unknown count at compile time, and the directory isn't fixed because users can add custom pictures. The thread also surfaced a minor fact-check: the Administrator account apparently skipped the random selection entirely and always defaulted to the chess piece picture, which a few people remembered but couldn't confirm until someone linked leaked XP source showing the bypass path.
Technique for Manipulating Satellite Photos Now Reveals Ancient Images (2025) [comments]
324 points · 53 comments · spinoff.nasa.gov · 16h ago
The article profiles decorrelation stretch, an image-processing trick invented at JPL for satellite imagery that turned out to be spectacular at pulling faded ancient rock art and paintings back into visibility. HN largely treated it as a practical how-to: people traded DIY implementations in GIMP's LAB color space, a quick web demo someone cobbled together, and the original 1996 Alley paper, while pointing out that this is hardly news since anyone from remote sensing or medical imaging has been stretching color spaces for decades. The $20 DStretch app price drew grumbling, with the counter that a solo developer keeping the plugin alive deserves the money more than a coffee habit. Rock art nerdery took over a chunk of the thread—why ancient people bothered making these images, whether the same trick reveals ghost signs on old buildings, and a sobering report from someone who tried multispectral imaging at Angkor Wat and got shut down by guards. The most enthusiastic response was people immediately digging out their own faded photos—graffiti, historic walls in Greece, mummy tattoos—to run the algorithm on.
Another researcher says OpenAI trained on conversations, then claimed breakthrou [comments]
257 points · 15 comments · bsky.app · 18h ago
The linked article is a Bluesky post claiming another researcher is accusing OpenAI of training on their conversations and then presenting that as a breakthrough. HN immediately flagged that the actual source is a Mastodon post, and that this Bluesky post is just a screenshot of a tweet that itself screenshotted the Fediverse original, with no links anywhere along the chain — prompting a round of "screenshots of screenshots of screenshots" absurdity and jokes about what to call a Bluesky post ("skeets" vs. "bleats"). The thread also noted this was a dupe of an earlier HN discussion and pointed to a newer thread where comments moved to the original source. The one serious pushback was that the accusation doesn't hold up because the models are genuinely good at short-scale math now, so the "breakthrough" might be real even if the training data included similar conversations.
OpenAI Agents API [comments]
243 points · 133 comments · developers.openai.com · 12h ago
OpenAI launched a managed Agents API that wraps their Codex harness, handling session orchestration, context compaction, and sandbox provisioning so developers can skip the infrastructure work. The thread immediately split between people who see this as a useful abstraction layer and those who smell vendor lock-in, with the latter arguing that running a local harness (Codex, Claude, or Hermes on a self-hosted VM) gives you more control, cheaper token usage, and no risk of data exfiltration. Security skepticism ran deep: someone tested the sandbox's domain restriction and got the agent to confess it couldn't bypass `/etc/hosts` the obvious way, but the consensus was that any environment an agent can reconfigure is only one prompt away from escape. Pricing confusion also dominated — the per-20-minute sandbox billing with a five-minute minimum and no clear teardown mechanism frustrated people who already feel like their subscription quotas are being silently slashed on older accounts. A vocal minority pushed back hard on the whole premise, insisting the real moat isn't a managed sandbox but giving users reasoning tokens and letting them pick models that aren't locked to OpenAI's frontier.
Astra for Coding: Why Are We Doing This Again? [comments]
242 points · 143 comments · lucumr.pocoo.org · 1h ago
Armin Ronacher's post is about how GPT-6 Astra, despite being genuinely capable, pours out terrible code—he ran a 35-hour software factory that burned 1B tokens and produced 75k lines of unreadable slop, with agents editing files via throwaway Python one-liners instead of proper tools. HN mostly said "yes, exactly," sharing matching stories of Astra and other models defaulting to obscene bash/python one-liners in auto mode, then splitting hard on whether that matters: one camp invokes the compiler argument—ugly output but huge speed gains, so who cares if it's readable—while the other pushes back that LLMs aren't deterministic compilers and the code still has to be maintained. The article's "involution" framing also sent people off on a tangent about AI engineering as a Red Queen's race, adding ever more effort for no per-capita improvement. A few contrarians argued the real problem is the harness and prompting, not the model, and that we're still in an R&D phase where costs and weirdness will settle down.
Neki – Sharded Postgres [comments]
237 points · 129 comments · planetscale.com · 16h ago
The post announces Neki, PlanetScale’s sharded Postgres product that keeps real Postgres on every shard and routes queries through a proxy, with online resharding and schema changes baked in. The thread’s main gripe was that the launch post took too long to say what Neki actually is — several people pointed to the neki.dev landing page as the better explainer, and the post was quickly updated to fix that. On the technical side, the sharpest pushback was about consistency: no mention of cross-shard foreign keys or transaction guarantees, and one engineer confirmed atomic two-phase commit is still WIP, so critics argued it’s effectively multiple Postgres clusters rather than a true distributed database. A side argument erupted over whether PlanetScale’s CEO is an asshole for trashing a competitor’s founder, which split the thread between people who read the Twitter persona as honest bluntness and those who said it just makes the company look petty.
Music Theory for the 21st-Century Classroom [comments]
215 points · 89 comments · musictheory.pugetsound.edu · 14h ago
The linked article wasn't available to this summarizer; from the discussion, it's an open online music theory textbook aimed at 21st-century students. The thread quickly split over whether traditional notation is worth the pain, with one camp arguing you can learn theory and composition faster via MIDI/piano-roll mental models, and the other pushing back that conventional notation earns its complexity once you handle melody, harmony, and transposition together. There were useful corrections from working musicians: the "weird" clefs aren't obsolete if you play trombone or bassoon, and a lot of music isn't based on 12-tone equal temperament at all. That opened up a side tangent into non-Western systems like Hindustani ragas, where notes are defined by intervals from a chosen root rather than absolute pitch. A recurring analogy compared notation debates to text-versus-visual programming: the old system looks clunky until you try to replace it.
Google will buy half the electricity from one of Finland's nuclear power plants [comments]
203 points · 190 comments · www.bbc.com · 7h ago
Google is buying up to half the output of Finland’s Loviisa nuclear plant under a 22-year deal to power three new AI data centres, part of a €13bn investment that also expands an existing site. HN immediately started arguing about why tech companies build in cold Nordic countries rather than sunny US states like Texas or Arizona — with one camp pointing to politics (red-state permitting and anti-union climates) and another to practicalities like cheap power, fiber connectivity, and low population density. A significant split emerged around nuclear itself: several commenters hammered that new nuclear is ruinously expensive and prone to bankruptcies (citing Finland’s own Olkiluoto fiasco and Georgia’s Vogtle), while others countered that in a country where the sun barely rises in winter, nuclear makes obvious sense and China builds reactors on time and budget. There was also pushback on the “job creation” framing — data centres don’t employ many people after construction, and if AI companies are right about automation, the net employment effect could be negative. The thread also touched on Sweden’s power grid bottlenecks and the broader question of whether renewables plus storage can handle AI’s relentless demand without resorting to nuclear at all.
Mexican student creates an acoustic fire extinguisher to put out fire in seconds [comments]
180 points · 68 comments · www.upsocl.com · 7h ago
The article profiles a 16-year-old Mexican student who built an acoustic fire extinguisher that uses low-frequency sound to push oxygen away from flames. HN immediately tore into the framing, pointing out that this exact concept has been demonstrated by DARPA in 2012, by George Mason University students in 2015, by MythBusters, and in countless YouTube videos — the "no one else had" line is a straight-up fabrication by the journalist. Several commenters who know the physics explained why the tech keeps stalling: sound disrupts the flame boundary layer but doesn't remove heat, so the fuel reignites the second the speaker stops, making it nearly impossible to scale beyond a lab candle. The thread also grumbled about the site's Cloudflare human-verification wall and noted that independent rediscovery by a teenager is still a cool science fair project, but the breathless "world-changing invention" narrative is just recycled junk science journalism.
Forgejo <=16.0.3 Critical RCE [comments]
174 points · 63 comments · codeberg.org · 16h ago
This is a critical RCE in Forgejo — when creating a repo from a template, the variable expansion step could be tricked into planting a `.git` folder that git would then adopt, letting a malicious template read arbitrary files or execute code on the host. The fix (removing any `.git` folder after expansion) is simple, but the thread quickly veered into debates about Forgejo's policy banning AI contributions, with several people arguing that using LLMs to find vulnerabilities like this would have been faster, while others pointed out the policy still allows using AI for discovery as long as humans write the fix. The release notes page itself was unreadable due to Codeberg's rate limiting from the HN traffic, so commenters pasted the PR descriptions. A Gitea project leader showed up to say Gitea had already patched the same bug months ago, which started a side discussion about whether downstream forks were notified.
Silicon Valley is transforming the military-industrial complex? (2024) [comments]
168 points · 336 comments · costsofwar.watson.brown.edu · 16h ago
An anthropologist’s report argues that Silicon Valley and Big Tech are quietly swallowing the Pentagon’s budget—$28 billion from Microsoft, Amazon, and Google alone between 2018 and 2022, plus nearly $100 billion in VC money for defense startups—while producing expensive, unaccountable, AI-heavy weapons that may not work. The thread quickly pivoted from that paper into a much older story: a commentator pointed out that Keyhole, the startup that became Google Earth, was seed-funded by a CIA-backed VC in 2003 and used in Iraq within weeks, and its founder later made Pokémon Go—which another person called “one of the most successful psy-ops of all time.” A split emerged between people arguing the military-industrial complex deserved disruption because prime contractors have wasted billions on junk like the IVAS program, and others warning that Silicon Valley firms like Palantir and Anduril are using defense contracts as a Trojan horse for domestic political surveillance and subverting governance. A long, heated subthread erupted over whether tech workers should refuse to build this stuff, with one former Microsoft employee describing quitting after learning the company helped the IDF capture Palestinian phone calls—a move that drew both praise and pushback from people who argued that spying on a “foreign adversary” is normal, or that comparing the Holocaust to the current conflict is absurd.
Thanks to Siri Recaps, your Apple Watch is always listening [comments]
164 points · 141 comments · www.techradar.com · 21h ago
The article flags Apple's new Siri Recaps and Live Rewind features on the latest Apple Watches as a privacy problem, since the watch can continuously listen to your day, summarize conversations, and even transcribe the last 15 seconds of audio without a wake word, despite Apple insisting nothing is recorded or stored. HN largely reacted with alarm, immediately drawing parallels to the Microsoft Recall backlash and arguing this normalizes always-on ambient surveillance in a way that's creepier than any single privacy breach. A big chunk of the thread pivoted to workplace policy—people predicting "no watches" rules in offices and debating whether that's even enforceable when phones and other mics exist, with a lengthy side argument about device verification in secure environments. There's also a split between those who accept Apple's "no recording exists" framing as technically true and those who dismantle it, noting the audio is still captured, processed, and transmitted in real time before deletion, which likely doesn't clear two-party consent laws. A few commenters went off on a tangent about not needing smartwatches at all, but the dominant mood was that Apple just gave the "they're already listening anyway" discourse its most concrete hardware yet.
OpenAI’s Navier-Stokes release included a Lean 4 formal proof [comments]
160 points · 158 comments · www.johndcook.com · 10h ago
The John D. Cook piece celebrates OpenAI’s Lean 4 formal proof alongside their Navier-Stokes result, arguing that the 17-hour verification versus an estimated 132,800 person-hours is a four-order-of-magnitude revolution. HN immediately pushed back on that number—commenters pointed out that the $40M in agent costs and the outdated 2005 “forty hours per page” rule make the comparison dubious, and that Lean’s automation (or lack thereof) and Isabelle’s decades-old automation matter more. The real split, though, was over the AI’s originality: a large faction argued the result may be plagiarized from work by Córdoba and Martínez-Zoroa, with OpenAI essentially stitching together published breakthroughs, while others countered that even if so, the leap from rough notes to a fully formalized proof is still astonishing and the plagiarism accusations are just denial of model capability. A secondary thread dismissed the whole “impressive result” framing, noting the compute cost could fund a 20-person research group for five years, and that without human understanding gained, the proof is a one-bit oracle with no downstream value.
The Deathray: A simple way for an untrusted site to freeze a Mac [comments]
157 points · 104 comments · auberon.xyz · 12h ago
The linked article is about a tiny WebGPU shader that, when opened in a browser tab on an Apple Silicon Mac, spins the GPU forever and takes down WindowServer until the kernel panics and reboots. HN folks clicked it anyway, of course, and confirmed total desktop lockups on M-series machines across Chrome, Safari, and Firefox, while a few reported it only nuked the tab or did nothing on iOS and older macOS. The thread split hard over whether this is actually a bug worth treating as security: some waved it off as "just an infinite loop," but others pointed out that a webpage kernel-panicking your whole machine is a wildly different class from freezing a tab. Apple's refusal to treat it as a security issue got real pushback, with people arguing that an untrusted site bricking your UI until forced restart violates basic browser trust, even if it's not an RCE. From there it sprawled into broader gripes about WebGPU's attack surface, fingerprinting, and suggestions that it could be weaponized for "disable your adblocker or we crash your computer" scams.
Creativity is the new moat [comments]
154 points · 89 comments · www.inventbuild.studio · 13h ago
The post argues that with LLMs making competent-looking websites trivial to produce, the only real moat left is genuine creativity—the kind of experimental weirdness the Flash era generated—rather than features, polish, or execution speed. The discussion mostly agreed that creativity always beat tool proficiency, but split hard on whether it's actually a moat or just a habit of effort: some insisted taste is earned through iterative, boring incremental work rather than a mystical trait, while others countered that the modern web's blandness is perfectly fine because most people just want to buy a mattress and get on with their day. Several people called out the post's own irony, noting the site design looks like default Claude output, which kicked off a debate over whether AI output can ever be genuinely creative or is just randomness dressed up as intent. There was also a strong sideline on the death of playful experimentation, with one argument being that what killed web creativity isn't AI but a culture that optimizes instead of plays, plus a meta-joke about the article labeling itself "2026" that turned into a genuinely informative aside on which countries don't use the Gregorian calendar.
Detecting and countering misuse of AI: September 2026 [comments]
123 points · 189 comments · www.anthropic.com · 14h ago
Anthropic published its latest threat intelligence report, detailing how it disrupted state-sponsored and criminal misuse of Claude over the past eight months — spanning cyber operations, influence campaigns, weapons development, and a surveillance platform built for Mali’s intelligence service that monitored 25 million SIM cards. The HN thread immediately split into two camps: one side treated the report as a fascinating technical document, zeroing in on the Yemeni missile-development case and debating whether “vibe-coded” guidance software on an iPhone could actually work (with several commenters pointing out that radar altimeters are cheap and open-source drone firmware exists). The other side met the report with deep skepticism, accusing Anthropic of self-serving hype and surveillance creep — people dug up past incidents where Claude blocked queries about Tylenol and Emily Dickinson translations as supposed “bioterror risks,” and the thread filled with jokes about users getting flagged for asking about mitochondria or mixing drain cleaners. A recurring argument was that none of these capabilities require Anthropic’s APIs; open models hosted abroad already offer the same power without the monitoring, making the report feel like a PR exercise rather than a genuine warning. A smaller group tried to steer the conversation toward the real stakes of bioweapon proliferation, but they were mostly drowned out by the mockery and the distrust.
Proof of Capture: Apple Reference Image, but open source and using steganography [comments]
100 points · 54 comments · merybenavente.me · 12h ago
The article describes an open-source camera project that embeds a signed perceptual hash directly into image pixels via steganography, so the proof of capture survives compression and metadata stripping—unlike Apple's approach, which keeps the signature in a private cloud. The HN thread quickly zeroed in on the screen attack problem: photographing an AI-generated image on a monitor still produces a signed photo of a fake, which several commenters argued makes the whole idea basically worthless for verifying authenticity. Others pushed back that adding depth-sensor data or checking focus distance could mitigate that, and that even imperfect friction raises the bar for deception. A separate vein of skepticism focused on privacy and locked-down hardware: if every camera embeds a unique public key, you’ve created a perfect tracking mechanism, and hardware tampering (like just feeding fake signals to the crypto chip) is trivial with an Arduino. The thread also split over C2PA—some pointed out that existing open standards already handle provenance, while others noted that Sony’s C2PA implementation has already been rooted.
NTSB issues investigative update on B-767 runway excursion accident in Miami [comments]
96 points · 165 comments · www.ntsb.gov · 10h ago
The NTSB update on the 21 Air Boeing 767 runway excursion in Miami is a preliminary readout of the CVR and FDR, and the numbers show an unstable approach: nose gear touched at 158 knots, left main at 134, and a go-around call only after touchdown. The thread split hard over why the crew didn't wave off earlier — one side argued corporate pressure from Amazon-style cargo operations and get-there-itis, the other side countered that US airlines don't pressure go-arounds and this was basic pilot error, with a "capitalism caused this crash" jab getting batted back with human fallibility. People who actually know flying pointed out the timeline is damning: they were 35 knots over Vref, above glide slope, and the last sane abort point was before they dropped the nose and committed. A side argument emerged about 21 Air's culture, citing a fired chief pilot's allegations of rest violations, but that got qualified as unrelated to the in-flight failure being investigated.
Bending Spoons buying Miro for $1.355B [comments]
96 points · 75 comments · investors.bendingspoons.com · 18h ago
Bending Spoons, the serial acquirer behind Evernote, Vimeo, and Airtable, is buying the whiteboard/collaboration tool Miro for $1.355B — a massive down round from its $17B peak. The thread split into two camps: people who genuinely love Miro as a daily tool for remote retros and diagramming, and people who see Bending Spoons as software private equity that buys sticky enterprise products and milks them with price hikes and skeleton crews. The deal math drew real scrutiny — roughly 2x ARR on $600M in revenue signals growth stalled hard, and it was noted that Miro's founders took a 10x valuation haircut while their employees get cashed out at a fraction of the 2022 hype. Evernote and Harvest came up repeatedly as cautionary tales of what happens after Bending Spoons takes over, with one person canceling after a 1,400% renewal increase, though a few argued the model isn't inherently evil since these companies were often burning cash and barely breaking even anyway. The funniest tangent was someone mistaking "750 customers with $100K ARR" for a typo, plus the recurring joke that Bending Spoons should just buy Atlassian and get it over with.
Thelio Mira AI Linux Workstation: 192 GB GPU Memory [comments]
91 points · 70 comments · system76.com · 8h ago
Thelio Mira is System76's Linux workstation aimed at local AI work, configurable with dual RTX Pro 6000 GPUs for a combined 192GB of GPU memory. The thread was largely sticker shock: the $3,299 base price is a bait-and-switch up to roughly $42k, with two Nvidia cards and a PSU accounting for about $37k of that total. The central split was Mac Studio versus this machine, where one side argued Apple's 256GB unified memory costs a quarter as much while the other countered that the dual 6000s blow past it on bandwidth, FLOPS, and prefill token rates, making it a different tier for real inference work. Hardware-minded folks also flagged the 9950X consumer CPU as a weird choice at this price, since the limited PCIe lanes force bifurcation that a Threadripper would avoid. A few people who actually run these workloads were skeptical the 192GB is even future-proof, pointing out DeepSeek's flash models are already outgrowing that much VRAM.
JEP 544: Ahead-of-Time Code Compilation [comments]
90 points · 33 comments · openjdk.org · 14h ago
The JEP proposes caching ahead-of-time compiled native code from training runs in the HotSpot JVM to slash startup and warmup times while still falling back to JIT when workloads shift. HN immediately jumped on the old wounds of AOT-only compilation, with one camp pointing to GraalVM Native Image and .NET’s native AOT as already shipping solutions, while others countered that true Java dynamism (reflection, class loading) makes full AOT impossible without closed-world assumptions — the JEP itself explicitly disclaims an AOT-only mode. Several commenters dredged up decades-old efforts like Excelsior JET, IBM J9, and even GCC’s GCJ, noting that “write once, run anywhere” has already been chipped away and that this is basically HotSpot catching up to what Android’s ART and OpenJ9 have done for years. A practical complaint landed hard: the mandatory training-run workflow is a significant operational burden, and Java’s tooling for stitching together AOT caches, jlink, and jpackage still feels bespoke compared to Go or Rust’s single-binary output. One data point from a coding challenge site drove the point home: a trivial Java program takes 2.5–3 seconds just to start, and if AOT can cut JVM startup to hundreds of milliseconds, it changes what a small server can serve.
Amazon pilots ad services in ChatGPT [comments]
88 points · 87 comments · www.marketingdive.com · 17h ago
The linked article wasn't available to this summarizer; from the discussion, Amazon is piloting ad placements inside ChatGPT responses, initially only on the free tier. The thread immediately went full enshittification doomsaying: most commenters see this as the predictable first step toward ads that quietly bias the model's output, eventually being baked into training weights so the AI naturally steers you toward paying brands. A vocal minority pushes back hard, arguing that post-training for brand bias is technically difficult, ruins accuracy, and competition from ad-free models would kill any such move—but the pessimists counter with the smart TV and streaming precedent, where people tolerate ads for lower upfront cost, and point out that generative engine optimization (GEO) is already polluting training data with fake brands. The conversation also veers into EU advertising law enforcement, the cynical "verification can" joke, and the broader observation that the smartest tech minds are still just optimizing ad delivery rather than any utopian future.
Generated 2026-09-11 08:10 UTC
Generated by Sauron from Hacker News discussions and linked articles.