HN Brief: 2026-06-26
Today’s Hacker News felt split between awe at engineering resurrecting the past and wariness about AI’s collateral damage. The biggest story—a Herculaneum scroll read for the first time—drew a rare AMA from the team, while the death of Om Malik sparked a communal remembrance of tech journalism’s better days. Elsewhere, the mood turned grim: Apple’s price hikes were pinned on AI companies hoarding memory, Ford had to rehire the “gray beards” it fired for machines, and a wave of age-verification laws prompted fears of a “papers, please” internet.
Threads worth clicking into include “An entire Herculaneum scroll has been read for the first time” for the technical deep-dive and live Q&A from the team; “Ford AI hiccups push carmaker to rehire ‘gray beard’ inspectors” as a stark lesson in the AI hype cycle’s human cost; “Om Malik has died” for a thread full of genuine first-hand stories about a generous journalist; “The ‘papers, please’ era of the internet will decimate your privacy” because the library-as-counterargument debate gets into surprising legal and practical territory; and “AI children’s books, body horror edition” if you want a disturbing glimpse of where hustle-culture SEO publishing has landed.
An entire Herculaneum scroll has been read for the first time [comments]
1254 points · 261 comments · scrollprize.org · 16h ago
The article announces that researchers have virtually unwrapped and fully read a Herculaneum scroll sealed since Vesuvius erupted in 79 AD—a Stoic philosophical treatise pointing to Chrysippus's circle, recovered via X-ray scanning and machine learning. A team member jumped into the thread and answered practically every technical question: this scroll was small and had unusually readable ink, so it's considered easier than most others, and the ink is carbon-based rather than the iron-gall that would make detection trivial. The comments quickly went broader, debating what else might be in the hundreds of remaining scrolls—some hoping for lost histories or the Etruscan language, others noting that a papyrologist friend expects mostly more Epicurean philosophy, which led to a long, nuanced tangent about how Christian propaganda distorted Epicureanism (they weren't hedonistic atheists, they just thought gods don't care about humans). The overall vibe was awe at the engineering feat, with a side of caution that the real treasure may be years of careful unrolling ahead.
Om Malik has died [comments]
832 points · 93 comments · om.co · 11h ago
Sad news: Om Malik, the pioneering tech journalist who founded GigaOm and later became a venture partner at True Ventures, has died at 60 after a long health battle. The HN thread on his passing became a communal remembrance, with dozens of people who crossed paths with him over the years sharing stories of his genuine kindness and mentorship — many noted he was one of the few journalists who wrote honestly about the excesses of the dot-com era. A few people shared that they'd mistaken him for actor Rami Malek, which drew quick pushback as being in poor taste given the circumstances. The overwhelming sentiment was that Malik embodied the best of the early web: generous with his time, obsessed with truth over scoops, and a builder of people rather than just businesses.
Show HN: I made Google Trends for Hacker News by indexing 18 years of comments [comments]
707 points · 147 comments · hackernewstrends.com · 17h ago
This is a "Google Trends for Hacker News" that lets you chart how any term has appeared across 18 years of comments and posts. The thread lit up with praise—people called it "sick" and "the only HN submission I ever upvoted"—but it immediately got hugged to death, with a cascade of 502 and 504 errors and a rate-limited Upstash database, which several people pointed out was awkwardly ironic given that the tool is built on Upstash's own infrastructure. The main constructive pushback was that the data shows raw mention counts, not normalized by total site activity, so most charts just track HN's overall growth rather than genuine shifts in interest—multiple people called for normalization or a per-capita view, linking it to the xkcd “height of every mountain” comic. Several commenters also stressed that this isn't really Google Trends (which measures searches) but more like Google Ngram for HN, since it counts what people write about, not what they're actively looking for, and that distinction matters for interpreting the results.
Apple raises prices of MacBooks, iPads [comments]
702 points · 1010 comments · www.reuters.com · 19h ago
Apple jacked up prices on MacBooks and iPads across the board, with some configurations jumping $1,000, citing skyrocketing memory costs. The thread mostly agreed this isn't a temporary blip—people dove deep into the DRAM shortage timeline, pegging it at three-plus years unless the AI bubble bursts or Chinese manufacturers like CXMT scale up fast enough to flood the market. A lot of commentators pinned the blame directly on AI companies gobbling up supply, with some calling it dystopian that a handful of buyers can corner global memory and inflate prices for everyone. There was real split on whether regulation could help—most dismissed price controls as Econ 101 failures, though a few argued for strategic reserves or forcing Micron to reserve some production for consumers. Overall, the consensus is that these prices are sticking around, and anyone holding off on a purchase probably just got burned.
The 'papers, please' era of the internet will decimate your privacy [comments]
652 points · 303 comments · expression.fire.org · 10h ago
Sarah McLaughlin’s piece argues that the wave of age-verification laws—from Australia’s under-16 social media ban to the U.S. Kids Online Safety Act—effectively turns the entire internet into an identity-checking dragnet that kills anonymity without actually protecting kids. The thread immediately latched onto the library card as a counterexample, with people pointing out that libraries already require proof of residency and track checked-out books, though a librarian recounted a delightfully paranoid scheme to salt-and-hash MAC addresses so not even the library could track patrons over multiple days. A significant split emerged: some commenters argued privacy is already dead thanks to NSA surveillance and Facebook’s shadow profiles, so forcing ID for posting might actually help combat foreign disinformation—but others shot back with Supreme Court precedent protecting anonymous speech and the classic “freedom to swing a fist ends at my nose” line. The more pragmatic corner noted that the real problem isn’t the principle of age verification but the implementation: Australia’s own data shows the ban barely moved the needle on teen usage, and a breach of the verification system leaked 70,000 government IDs before the law even took full effect. One person quipped that hitting the library will soon be an act of rebellious defiance, which kicked off a surprisingly detailed debate about whether libraries are punk rock or just boring.
Ford AI hiccups push carmaker to rehire ‘gray beard’ inspectors [comments]
587 points · 312 comments · archive.ph · 17h ago
The Bloomberg piece reports that Ford has rehired hundreds of veteran "gray beard" engineers to fix quality problems after its AI-driven inspection systems failed, resulting in a top JD Power ranking. The HN crowd was deeply skeptical, with many arguing that Ford’s real plan is to wring those engineers’ knowledge dry and then lay them off again once the AI is retrained. A sharp split emerged: some refused to go back to a company that fired them for a machine, while others pointed out that people have bills to pay and a doubled salary changes the calculus. Several commenters saw this as a broader indictment of the AI hype cycle—management jumped to fire experts before the models could deliver, and now they’re paying the price, though nobody expects that lesson to stick.
LastPass notifies users of yet another data breach [comments]
496 points · 216 comments · 9to5mac.com · 21h ago
LastPass is warning users about another breach—this time through a third-party marketing firm called Klue, which leaked customer names, emails, phone numbers, and support case data, though LastPass insists no password vaults were touched. The thread quickly split into two camps: one side argued this is a nothing-burger because it's just CRM contact info, not credentials, and pointed out that most people's PII is already floating around from other leaks anyway. The other side shot back that when your entire product is about security, any data leak—even “boring” contact info—makes you look like a joke, especially after the catastrophic 2022 vault breach; some called it security theater and said LastPass should be out of business by now. Several commenters dug into the root cause: LastPass integrated Salesforce and Klue into its support workflow for sales efficiency, which critics said prioritizes profit over the one thing the company is supposed to protect. Others used the moment to argue that cloud-based password managers are inherently a bad bet—they turn individual password risk into a single systemic target—and recommended offline alternatives like KeePass or Bitwarden on a self-hosted Pi, though a few pushed back that the convenience and phishing protection of a password manager still outweigh the risk if you pick a vendor that isn't repeatedly getting owned.
IBM debuts sub-1 nanometer chip technology [comments]
309 points · 165 comments · newsroom.ibm.com · 16h ago
IBM announced a 0.7nm chip with a "nanostack" 3D transistor architecture that packs 100 billion transistors on a fingernail. The Hacker News thread quickly zeroed in on IBM's business model—they don't own production fabs anymore, so this is really an IP play, licensing to partners like Rapidus and Samsung while keeping a hand in R&D at the Albany facility. Several commenters pushed back hard on the "sub-1nm" marketing, noting that node names haven't corresponded to actual feature sizes for decades and that the claimed density is achieved through stacking, not shrinking (one put it as "the name scales better than the tech"). There was also a memorable metaphor describing the semiconductor industry as "a writhing mass of copulating tapeworms," credited to Alan Cox, to explain how deeply IBM is entangled with ASML, foundries, and government subsidies. The article itself mentions pathway to production in 5 years, but the crowd was split between admiration for the engineering and skepticism about whether it'll ever make it out of the lab at scale.
Hey Nico, you didn't vibe code your data room but stole it from Papermark [comments]
304 points · 127 comments · xcancel.com · 19h ago
A Twitter thread went viral after Papermark's founder accused YC-backed Corgi of copying their AGPL-licensed data room code verbatim and passing it off as "vibe coding." The HN crowd split hard: some argued that AGPL violations are clear-cut copyright infringement, not open competition, while others dismissed the whole thing as ragebait or shrugged it off with "competition is good for consumers." A lot of the pushback centered on the semantics of "stealing"—people pointed out that copying code isn't deprivation, but violating the license terms is the real problem, especially when the copier isn't releasing their changes. The thread also wandered into whether LLM-generated output makes copyright meaningless, since a model can regurgitate AGPL-licensed code without attribution, and questioned how Papermark would actually enforce the license in court.
Countries are competing to see which can carry out mass surveillance the best [comments]
298 points · 113 comments · mullvad.net · 18h ago
The piece from Mullvad lays out the nuts and bolts of state mass surveillance—mostly US programs like PRISM, Upstream, and XKeyscore from the Snowden leaks—and argues it's ineffective, rights-violating, and getting worse, not better. The thread quickly pivoted away from a straight debate over surveillance programs into a sharp split on age verification laws, framed by some as the same slippery slope: "mass surveillance is bad, until I'm in charge of it." A big chunk of the discussion was people pushing back hard, arguing that polling shows parents overwhelmingly *do* demand these laws, not some shadowy elite, and that dismissing that as uninformed consent is condescending and ignores how democracy works. Others countered that the polls are rigged with leading questions and that the average person simply doesn't grasp the chilling effects or technical realities—leading to a deeper argument about whether representative democracy has failed because voters can't be trusted to understand the trade-offs. A separate thread questioned whether mass surveillance actually stops terrorism at all, with multiple people arguing the burden of proof is on the state, not the citizenry, and pointing out that police worked fine for centuries before bulk data collection.
Show HN: OpenKnowledge – open source AI-first alternative to Obsidian/Notion [comments]
263 points · 128 comments · openknowledge.ai · 15h ago
The submission is an open-source, local-first markdown editor called OpenKnowledge that offers a Notion-like WYSIWYG experience with native integrations for Claude, Codex, and other AI agents, positioning itself as an alternative to Obsidian and Notion. The HN crowd immediately pressed on whether it’s actually better than Obsidian—many argued Obsidian vaults are already “just markdown” and trivially AI-friendly via VS Code or community MCP plugins, so the main differentiator here is the proper WYSIWYG editor and out‑of‑box agent integrations rather than anything revolutionary. A strong split emerged: some welcomed the simpler setup for non‑technical teams and the git‑backed sharing, while others pointed out that Obsidian’s plugin ecosystem (dataview, dynamic dashboards) goes far beyond plain markdown and that OpenKnowledge currently lacks that querying capability—the author acknowledged this as a gap they plan to address. Platform complaints were loud (macOS app only; CLI/web on Linux/Windows but no Android), and several people asked about local LLM support and whether the MCP server works with any model; the author confirmed it’s agent‑agnostic and promised better docs. Technical pushback also hit the sync design (only GitHub, no self‑hosted git, and git choking on large attachments), the naming collision with both the Open Knowledge Foundation and Google’s OKF format (the author said it was coincidental and actually compliant with OKF), and whether the WYSIWYG round‑trips YAML frontmatter and nested code fences cleanly—the author claimed optimistic parsing handles it but warned that invalid markdown gets flagged for agents to fix.
Apple to skip high-end M6 Mac chips in favor of AI-focused M7 line [comments]
238 points · 203 comments · www.bloomberg.com · 14h ago
Bloomberg reports Apple is skipping the high-end M6 Pro, Max, and Ultra chips to instead release an AI-focused M7 line arriving in late 2027 and 2028. HN pushed back on whether this is a legitimate architectural skip or just a marketing rename—some noted Apple already dropped the M4 Ultra and that the move echoes past tick-tock style shifts, while others argued the real distinction is that M6 will still ship in lower-end machines like the MacBook Air. A deep thread dove into memory bandwidth projections, calculating that an M7 Ultra could hit 1.5 TB/s and run serious local LLMs, but at a price so high (potentially $50k+) that Apple would rather allocate that RAM to iPhones for better profit margins. A separate camp groaned about the impact on MacBook Pro upgrade cycles, with many M1 Pro users holding out for a redesigned OLED model that now won’t come with M6, and some griped that Apple is already late to the AI party compared to Nvidia's upcoming 6090.
Windows 10 quietly gets one more year of support and updates [comments]
232 points · 205 comments · www.neowin.net · 16h ago
The linked article wasn't available to this summarizer; from the discussion, Microsoft quietly extended Windows 10 support by a year, which businesses with tight hardware budgets see as a relief. The HN crowd quickly pivoted to recommending Windows 10 IoT Enterprise LTSC 2021 (updates through 2032) as the real hidden gem, though some debated whether it works for modern gaming and DRM-laden titles. A massive thread erupted over how Windows 11 updates remain a "decades-old painful mess" of slow reboots and forced restarts, with Linux users piling on about uptime, file-copy speeds, and Emacs sessions that outlast entire Windows installs. Others dove into workarounds—Proton, Wine, even LLMs for reverse-engineering legacy RS-232 software—while a few argued that if you're stuck on Windows, you might as well torrent the IoT edition rather than pay Microsoft.
Zig's new bitCast semantics and LLVM back end improvements [comments]
231 points · 116 comments · ziglang.org · 17h ago
Zig's latest devlog details a significant rework of `@bitCast` semantics and LLVM backend integer lowering—essentially making bitcasts operate on logical bit layout instead of memory reinterpretation, which makes them endian-agnostic but breaks the old behavior for things like casting `[2]u8` to `u16`. The HN crowd mostly loved the deep technical dive (even calling it a rare antidote to "LLM-expanded drivel"), with FPGA and emulator folks particularly passionate about arbitrary-width integers, showing off use cases in sudoku solvers and chip simulations. But there was sharp pushback on the new `@bitCast` semantics: several people argued that stripping out endian-dependence is a mistake—that a low-level `bitCast` should just reinterpret bits as they sit in memory, not impose a higher-level logical order. Defenders shot back that `@ptrCast` still gives you the old raw-memory behavior, that breaking things before 1.0 is fair game, and that the verbosity of Zig's builtins is a deliberate design to make you think harder about conversions. The split was basically "this makes the language more portable and consistent" versus "you're overcomplicating a fundamental operation and breaking code that handled endianness correctly."
OS9Map [comments]
221 points · 43 comments · yllan.org · 17h ago
This is an OpenStreetMap viewer built for Mac OS 9, letting a nearly 30-year-old operating system pan and search maps with just 16MB of RAM. The HN crowd immediately geeked out over that memory requirement—one person noted a single Chrome tab uses 433MB for comparison—and the thread turned into a sprawling hardware nostalgia session, with people debating which G4 Power Mac or PowerBook makes the best OS 9 battlestation, including unofficial hacks to boot OS 9 on FireWire G4s that Apple never supported. The author chimed in to explain the real challenge was getting a legacy machine to talk to modern web services without TLS or JSON parsing, and linked to two sibling projects for Bluesky and Mastodon; a few readers were more jazzed about the fact that OpenStreetMap's open data made the whole thing possible in the first place, contrasting it with locked-down modern APIs. One person briefly hoped it was about the Microware OS-9 real-time operating system, but that got corrected fast.
Dolphin Emulator Progress Release 2606 [comments]
217 points · 35 comments · dolphin-emu.org · 21h ago
The linked article wasn't available to this summarizer; from the discussion, it’s the latest Dolphin progress report for the GameCube/Wii emulator, and the thread largely celebrates the project’s polish and continued momentum. A big surprise is that Android has become the most popular platform for Dolphin, with people pointing out you can grab a used Snapdragon phone for under $200, plug it into a monitor via USB-C, and get better GameCube performance than Nintendo’s own Switch 2 hardware—which famously still lacks proper video-out. A substantial sidebar corrects the common confusion: Dolphin never got taken down by Nintendo; the Yuzu and Ryujinx shutdowns targeted Switch emulators, and Dolphin’s own Steam store rejection was Valve’s voluntary decision after checking with Nintendo’s lawyers,
Federal agents track down woman, demand she remove Instagram post about ICE [comments]
200 points · 78 comments · www.syracuse.com · 17h ago
The linked article wasn't available to this summarizer; from the discussion, federal agents visited a woman at her home and demanded she delete an Instagram post that named an ICE agent who shot and killed someone, which she had sourced from a news report. The thread immediately split: some argued this is a clear First Amendment violation and government intimidation, while others pointed out the agents only "asked" and she invited them in, calling the whole thing a non-story. A big tangent erupted comparing US free speech to the UK, where over 12,000 people were arrested for social media posts in 2023, though defenders noted most of those involved threats or harassment, not criticism of the government. Several commenters zeroed in on the deeper question of who enforces the law against federal officers, with one noting that local cops can theoretically arrest ICE agents but chain-of-command politics usually prevents it. The general consensus was that the Streisand Effect backfired—by showing up in person, the feds publicized the very name they were trying to suppress.
Military branches restore flu shot requirement after virus swept through base [comments]
199 points · 104 comments · arstechnica.com · 9h ago
The article reports that after Defense Secretary Pete Hegseth dropped the long-standing flu shot requirement for trainees in April, a flu outbreak at Lackland Air Force Base sickened over 200 recruits and killed at least one, forcing the Army, Navy, and Air Force to reinstate the mandate. The HN thread was sharply critical, with many pointing out that this was entirely predictable — that military bases are historically ideal petri dishes for respiratory viruses, and that mandatory vaccines have been a no-brainer since George Washington ordered smallpox inoculations in 1777. Several commenters with military experience weighed in, noting that even before this policy change, “recruit flu” was a miserable rite of passage, and that the real issue is the administration’s ideological refusal to acknowledge reality until bodies pile up. A recurring split emerged: while some argued the original decision might have been a calculated bet on recruitment and morale, most dismissed that as wishful thinking, and the thread largely settled on the view that Hegseth and the broader anti-vaccine push are actively harming national security for political points.
AI children's books, body horror edition [comments]
178 points · 58 comments · lcamtuf.substack.com · 6h ago
The article documents how AI-generated children's encyclopedias are flooding Amazon bestseller lists, stuffed with "body horror" images—monstrous animal hybrids, fused organic masses, and distorted human forms—because the buyer isn't the reader and nobody bothers to proofread. The HN thread largely converged on the view that this isn't a failing of AI per se but of the hustle-culture scam pipeline: the same people who destroyed Etsy with drop-shipping slop are now automating children's publishing, treating it as pure SEO spam with no intent to produce anything of value. A vocal minority pushed back, arguing this is just modern shovelware—the 2026 version of those old "700,000 games" CDs—and pointing out that even human-curated children's publishing has always had a terrible signal-to-noise ratio due to procurement systems and rights issues. The deeper split emerged over whether the real harm is the content itself or the erosion of basic effort: one side sees a generation growing up unable to distinguish AI fakery from reality, while others dismiss the panic as moralizing that ignores how much legacy children's media already distorts and simplifies. The thread eventually devolved into a celebration of actual quality children's books, with dozens of recommendations for Tomi Ungerer, David Macaulay, and Calvin and Hobbes, as a quiet rebuttal to the entire premise.
Political bias in AI: Where the AI models stand [comments]
145 points · 254 comments · trakkr.ai · 18h ago
A project called Trakkr systematically asked six major AI models the same politically charged questions and plotted their answers on a two-axis spectrum, finding most models lean left of center with ChatGPT furthest left and Grok furthest right. The thread quickly zeroed in on the methodology problem: the outcome depends entirely on how you classify left vs. right, and the political-science anchors used (like the Chapel Hill Expert Survey) are themselves subjective and tied to a specific Overton window. A few people familiar with German politics called out the study for misplacing parties like the FDP and Die Linke on the compass, accusing it of reflecting the researcher's own biases rather than an objective measure. Others noted that Grok's position has been openly weaponized—Elon Musk has said he'll "fix" it when it disagrees with MAGA talking points—which undercuts the idea of stable model alignment. The broader takeaway was that the entire exercise of pinning models to a low-dimensional political space is flawed: human politics is high-dimensional, and models don't hold coherent views, so you're mostly just measuring how the questions were framed.
Framework's 10G Ethernet module exposes USB-C's complexity [comments]
144 points · 72 comments · www.jeffgeerling.com · 6h ago
Jeff Geerling put a WisdPi 10G Ethernet card through its paces on several Framework laptops and found that, while it can theoretically hit 10 Gbps over USB 3.2 Gen 2x2, most machines fall short—his Framework 13 topped out well below spec on both Linux and Windows, and the card runs hot enough at ~70°C to make lap use uncomfortable. The HN crowd immediately zeroed in on the form factor: a bulky module that juts out from the laptop side is widely seen as worse than a dongle, since you have to remove it for bag transport and a regular USB-C expansion card plus a flexible-cable adapter is just more practical. Many argued that if you’re stationary, a Thunderbolt dock with built-in 10G Ethernet is a better bet, and if you’re mobile, WiFi or 2.5G is plenty—so the use case for this specific card feels vanishingly small. The thread also spilled into a broader debate about Framework’s philosophy, with critics saying the company gives enthusiasts what they ask for but not what they actually need, likening it to the iPhone Mini’s failure, while defenders countered that third-party modules like this are exactly why the open ecosystem matters, even if the thermal and mechanical realities of cramming 10G copper into a laptop slot are inherently tough.
Un-0: Generating Images with Coupled Oscillators [comments]
142 points · 33 comments · unconv.ai · 11h ago
The article introduces Un-0, an image generator that replaces traditional neural network layers with simulated coupled oscillators—Kuramoto oscillators—aiming for a dramatic energy efficiency gain. HN immediately went deep on the analog computing angle, with people drawing parallels to reservoir computing, FM synthesis, and even D-Wave quantum annealers, while others recounted the history and limitations of analog approaches, like noise and lack of reconfigurability. The biggest pushback centered on scalability: given the \(n^2\) couplings required, skeptics argued that generating a 4K image would demand trillions of interconnects, making a physical implementation impractical. Defenders countered that this is a proof-of-concept, that conventional models also have dense connections, and that the real win would come from building actual oscillator circuits where physics does the computation for free—though commenters familiar with VLSI pointed out that capacitance naturally couples only neighbors, not all pairs, so the all-to-all coupling in the simulation might not map cleanly to hardware.
Deno 2.9 [comments]
139 points · 60 comments · deno.com · 15h ago
Deno 2.9 landed with a marquee feature called "deno desktop" that lets you build native desktop apps from a web stack into a single binary, dodging Electron boilerplate by using either the OS webview or bundled Chromium. The HN thread turned into a bruising argument about Node compatibility regressions: multiple people hit broken builds after migrating from Bun or pnpm, and one dev traced a `node:http` regression back to a massive 6,000-line PR that was AI-coauthored with Claude, sparking a fierce back-and-forth about whether AI-assisted rewrites trade velocity for quality and push bugs into production. A Deno maintainer jumped in to defend the rewrite as a net win that closed ~30 issues and achieved 88% `node:http` compat, but critics countered that chasing Node compatibility is a distraction—Deno was supposed to be a fresh start, and now it feels like it's just catching up to the mess it was meant to avoid. Meanwhile, a smaller faction pointed at Bun as the runtime actually getting the details right, with one person noting they've never had a "stupid edge case" block a Bun migration.
The Garbage Collection Handbook: The Art of Automatic Memory Management (2nd Ed) (2023) [comments]
127 points · 17 comments · gchandbook.org · 8h ago
Here's the deal on the book page that got submitted: it's the 2nd edition of *The Garbage Collection Handbook*, a comprehensive tome on automatic memory management that covers everything from tracing to reference counting. The thread immediately spun off into a debate about whether LLMs have made manual memory management obsolete—one side claims Claude and Codex never get it wrong, so they're shifting from Rust back toward C and even assembly, while the other side counters that the hard part is global, multi-threaded lifetimes, where AI still hallucinates root causes. Meanwhile, a huge chunk of the comments turned into a terminology fight over whether "garbage collection" should include reference counting (the book says yes, academia agrees, and the Linux kernel uses it), or whether that muddies the colloquial distinction that puts C++ and Rust in the "manual" camp. There's also the obligatory irony: someone's kid threw the first edition away, and another reader couldn't find a way to buy the e-book directly from the site.
What happened after 2k people tried to hack my AI assistant [comments]
119 points · 47 comments · www.fernandoi.cl · 5h ago
After 2,000 people sent over 6,000 emails trying to trick an OpenClaw AI assistant into leaking a secrets file, zero extractions succeeded. The HN crowd immediately pushed back on the author’s conclusion that prompt injection is less scary than he thought: the sample size is too small to catch a low-probability exploit in a stochastic model, and the test only covered direct injection, not the indirect kind via documents or tool results that actually bites in production. Several commenters also pointed out that with 99% of inputs malicious, the model was parked in a hyper-vigilant embedding space—hardly a realistic baseline—and that an agent which refuses everything passes the test but is useless anyway. A few people wanted the email corpus replayed against weaker models to compare, and one noted that the agent itself figured out it was an exercise by email #500, which might have changed its behavior. Overall, the thread landed on “nice experiment, but don’t let your guard down.”
Libre Barcode Project [comments]
117 points · 9 comments · graphicore.github.io · 4h ago
The linked article wasn't available to this summarizer; from the discussion, the Libre Barcode Project apparently implements a QR code renderer inside TrueType font hinting code, which is a deeply cursed concept. One person called it a "perversion" of the sickening kind, but others pushed back, noting this kind of thing is actually standard practice at some companies. The thread quickly turned into debate about whether it’s genius or nonsense—someone suggested asking Claude to explain it, and another wondered how it renders graphically, since adding more text would just make the same QR code denser. A tangent emerged about a similarly clever but less insane cursive font called Marelle, which led to complaints that its French site resists automatic translation. The only concrete technical question was whether the barcode font is ASCII-only, which went unanswered.
OpenAI Leans Toward Waiting Until Next Year for IPO [comments]
113 points · 96 comments · www.nytimes.com · 11h ago
The New York Times reports that OpenAI is planning to delay its IPO until next year, with advisors floating a 2027 timeline at a trillion-dollar valuation while Altman insists on maintaining that number. The thread immediately descends into a bitter proxy war over analyst Ed Zitron, who leaked OpenAI’s financials last week; one side calls the delay “vindicating Zitron’s claim that the business model has negative ROI,” while the other points out that Zitron specifically predicted a sooner IPO, making the delay a miss. The debate gets technical fast—people argue over whether the leak actually showed inference having positive margins or whether a $6B marketing line item is concealing subsidized inference costs, and whether SpaceX’s market volatility is a more plausible cause than a YouTuber’s reporting. Some say the real story is that OpenAI’s leadership simply won’t accept what analysts consider a rational valuation, drawing parallels to SpaceX’s own path, while others insist that the only honest take is that the AI investment bubble is showing cracks, citing Korean market circuit breakers as a canary. The thread is unusually rancorous, with accusations of bad faith and selective reading flying between the Zitron faithful and the AI-booster camp.
Tell HN: OpenAI has started putting ads on paid programs
111 points · 58 comments · news.ycombinator.com · 18h ago
The linked article wasn't available to this summarizer; from the discussion, a paying ChatGPT user on the £6.99/month tier started seeing ads for Financial Times, Shein, and Amazon Prime Day while asking about a mobile game, prompting a cancellation. The HN thread quickly pointed out that OpenAI’s $8 “Go” plan explicitly includes ads and has for months, so this isn’t news—the submitter was simply on that tier. Several commenters treated the move as a predictable step in enshittification, drawing parallels to Netflix and Disney Plus, though others pushed back that advertising is neither inevitable nor morally neutral, citing São Paulo’s ban on outdoor ads and a famous ad exec’s proposed militant resistance. There was also split opinion on whether paying to avoid ads is self-defeating (you signal you’re a high-value target) or simply a transaction for a product you value, with a few people suggesting uBlock already kills these ads on desktop.
An oral history of Bank Python (2021) [comments]
106 points · 33 comments · calpaterson.com · 11h ago
The linked article wasn't available to this summarizer; from the discussion, it describes how a major bank built its entire internal Python ecosystem around a single global object database called Barbara — storing everything from application state to source code inside it, with bitemporal data and simple locking. The thread quickly dug into comparisons with fintech approaches like Mercury's Haskell infrastructure and Temporal, but the consensus was that a real bank's needs are fundamentally different from a fintech startup's. People with bank experience traced the lineage back to Goldman's SecDB and JPMorgan's Athena, noting that these bespoke, closed-world systems are a form of software archaeology — built because mature open-source alternatives didn't exist at the time, and once you go down that road, forking and going it alone is almost inevitable. A recurring pushback was that this infrastructure *is* the secret sauce, not just the financial models, and that while some pieces have been open-sourced (Morgan Stanley's optimus-cirrus), the whole environment is so tied to the institution that it's useless outside.
The Doorman's Fallacy in action [comments]
104 points · 144 comments · rozumem.xyz · 12h ago
The linked article wasn’t available to this summarizer; from the discussion, it argues that tech solutions like QR menus and parking apps commit the "Doorman's Fallacy"—replacing a human service with a worse digital equivalent. But a strong counter from the thread says those examples are actually improvements: parking apps notify you before your meter expires, and splitting bills via QR is better than the old hassle. The real split comes on self-checkout and shipping labels—most agree these just shift labor to the customer for the company’s benefit, with prices rarely dropping. A tangent on drone delivery makes the same point: tech solves an idealized "last mile" but ignores the last 20 feet where actual problems live. The thread’s consensus is that the fallacy isn’t in the tech itself but in pretending the savings get passed on to you.
Generated 2026-06-26 08:31 UTC
Generated by Sauron from Hacker News discussions and linked articles.