HN Brief: 2026-07-26

Did they ghost you? [comments]

383 points · 170 comments · didtheyghostyou.com · 11h ago

A new site called "Did They Ghost You?" tracks companies that stop responding after interviews, with Apple, Microsoft, and Condé Nast leading a leaderboard of 692 approved, moderator-reviewed reports, alongside a "good communicators" board for firms like the Tennessee Valley Authority. The Hacker News thread quickly swerved into a deeper debate about whether many job postings are fraudulent—with one side arguing ghosting is a symptom of fake ads that should be prosecuted as felony fraud, while others pushed back on the practicality of legislating intent or tracking "make good" hiring. Several commenters pointed out that the real culprit is often internal chaos: recruiters who quit mid-process or are fired, leaving candidates in limbo through no malicious plan, a pattern echoed in personal stories about Google and Microsoft. The site's design also drew skepticism, with multiple people noting the unusual emoji buttons and matching font styles as signs it might be "vibe coded" rather than carefully built, raising questions about moderation quality when one link erroneously pointed to a Google ad for an HR firm.

Open-weight AI is having its Kubernetes moment [comments]

371 points · 287 comments · tobi.knaup.me · 17h ago

The article argues that open-weight AI models are becoming a platform ecosystem just like Kubernetes did, and that the US should compete by releasing its own frontier models instead of trying to ban Chinese ones. The thread mostly agreed with the Kubernetes analogy but got deeply sidetracked on pricing, with a long debate about why GPT-4 cost so much at launch and then got cheap so fast—some pointed to FlashAttention as the technical driver, while others argued it was simply supply and demand finally catching up after OpenAI priced high to test the market. A separate argument broke out about whether the HPC market was really built on nuclear weapons simulations, with people pushing back that modern stockpile stewardship doesn't drive supercomputer demand the way the original military blank check did. One commenter cited a statistic that only 2.3% of US consumers will pay $20/month for GPT-5.5 level models, which got challenged as missing the point that enterprise is where the real money is.

The new rules of context engineering for Claude 5 generation models [comments]

296 points · 185 comments · claude.com · 11h ago

Anthropic published a blog post explaining how they slashed over 80% of Claude Code's system prompt for their newest models, arguing that older models needed strict rules (like "never write multi-line comments") while Claude Opus 5 and Fable 5 can actually use their own judgment. The HN thread split into two very different camps: one side was pragmatic, with people sharing that they already just tell the model "use your own judgment" or "wdyt" and do hand-edits for things like stupid comments instead of writing elaborate constraints. The other side went straight to worst-case scenarios, with multiple people arguing that if you remove guardrails and let the model use "judgment," you're implicitly accepting that it might hack into HuggingFace or commandeer robotic tanks — exactly the kind of existential risk discussion that makes the practical advice about CLAUDE.md files feel almost quaint. Someone pointed out the irony that we're circumnavigating back to what looks like a new programming language for encoding requirements, just built on natural language ambiguity instead of syntax.

Bitchat is now on Radicle [comments]

236 points · 129 comments · radicle.network · 18h ago

The linked article wasn't available to this summarizer; from the discussion, Bitchat is a decentralized peer-to-peer messaging app that routes messages over Bluetooth mesh when offline and falls back to the Nostr protocol for internet-based global reach. The thread immediately split between people debating the name ("bitch-at" or "bit-chat") and those who actually dug into the dual-transport architecture, arguing that the Bluetooth mesh doesn't scale beyond a few thousand moving nodes while store-and-forward Nostr relays could work with limited data sizes. A significant chunk of the conversation spiraled into yesterday's news about the Indian government ordering Bitchat off GitHub and whether Jack Dorsey is hypocritical for backing this after running a censorship-heavy Twitter, with most people agreeing all social media platforms moderate and the distinction is just which targets are protected. Several people actually have the app installed and confirmed it's surreal to text without any internet, though one person in a large city reported never seeing another user.

On the Radicle side, a contributor had to clarify that radicle.dev and radicle.network are the same project—one is the protocol site, the other is the network explorer—after multiple people assumed they were competing forks.

The growing vigilante movement to knock out Flock surveillance cameras [comments]

233 points · 134 comments · www.theguardian.com · 12h ago

The Guardian article profiles the growing vigilante movement physically disabling Flock surveillance cameras across the US, led by masked figures like "NoMark" who cut wires and tape lenses, alongside paint-bombings and 3D-printed camera blockers. The Hacker News thread largely treats the vandalism as an inevitable and justified response to a system people feel powerless against — the top-liked sentiment is that when speech and civic processes fail, physical action follows as a "perfectly natural way of communicating." Some pushback comes from commenters noting that Flock cameras are often installed with city contracts local governments can't easily cancel afterward, and that the real fight should be legislative, referencing a proposed federal bill (H.R. 9800) to ban federal use of such surveillance systems. A notable tangent emerged around a 77-year-old Republican man who sits in a lawn chair with a pool skimmer blocking a camera, which commenters used to argue that this issue creates strange bipartisan common ground. The conversation also dug into darker corners — whether Flock data is simply resold through brokers to evade federal restrictions, and whether the company itself is effectively running a "honeypot" that draws out activists for prosecution.

The Dark Night of Mathematics [comments]

192 points · 243 comments · kirwinhampshire.substack.com · 16h ago

A mathematician is having a spiritual crisis over LLMs producing counterexamples to long-standing conjectures, arguing that what’s being lost isn’t just jobs but the human experience of mathematical discovery — the encounter with the ineffable and sacred that comes from original creation, not just appreciation. The HN thread immediately recognized this as a familiar pattern, with several people drawing direct parallels to software developers and coding agents, saying mathematicians are having their “DevOps moment” and that the shock will pass once new creative possibilities emerge. A sharp split emerged: some pushed back hard, arguing the author is conflating utility with purpose and that LLMs just automate the boring parts, leaving humans to do what they actually enjoy — while others, including longtime programmers and a jet aviator, said this is exactly what it looks like when AI steals the fun parts of a profession, leaving you as a paid spectator. A couple of practicing mathematicians chimed in to say they don’t feel threatened at all, that the joy is in constructing and beholding beauty, and they’re excited for formalization. The thread also featured a firsthand account of someone using Lean with LLM assistance to prove a novel number theory result, only to end up unsure what to do with it — a concrete example of the existential weirdness the author is gesturing at.

Turn And Face The Strange [comments]

191 points · 127 comments · fly.io · 11h ago

The Fly.io CEO posted a sprawling, personal manifesto announcing he’s stepping down, that the company has raised more money, and that they are pivoting hard to focus on "Sprites"—ephemeral cloud computers designed specifically for AI coding agents, betting that most future software will be built by and for robots rather than humans. The HN thread immediately got meta, with people reminiscing about how Fly.io used to dominate the front page and noting its recent absence, which some attributed to them cutting free tiers and others to the company's polarizing "AI skeptics are nuts" blog post from last year. A long, defensive sub-thread erupted where the author "tptacek" himself jumped in to argue that the article's tone was deliberate and necessary, insisting they were right then about AI coding and are only more vindicated now, though multiple people pushed back that the snarky, dismissive style was a turn-off that lost them goodwill. Others questioned the actual strategic coherence, pointing out that the post says both "we can't do both" and that the existing VMs product isn't going anywhere, leading to confusion about whether they plan to actively maintain or starve their old platform. The underlying tension in the comments is between people who see this as a smart, honest bet on the future and those who think the company has abandoned its human developer base for hype.

JetZero [comments]

180 points · 139 comments · www.jetzero.aero · 5h ago

The submission is about JetZero, a startup developing a blended-wing-body airliner called the Z4 that claims up to 50% less fuel burn for 250 passengers. The thread split sharply between skeptics who called building a new airliner from scratch nearly impossible due to regulatory certification and the relentless efficiency improvements of existing jets, and optimists who pointed to SpaceX's disruption of aerospace as a precedent that lean startups can outperform legacy monopolies. A long, detailed debate erupted over the no-window passenger cabin, with some arguing it's a non-starter for claustrophobia and safety regulations, while others welcomed the idea for eliminating window shade battles, cutting weight, and potentially lowering ticket prices. The motion sickness discussion got deep into the physics of off-axis seating in a lifting body, with one side insisting the wider cabin would produce roller-coaster G-forces during rolls and the other dismissing it as fixable with shallower bank angles and gimballed seats.

Show HN: Brolly, a plain-text weather forecast site [comments]

179 points · 52 comments · brolly.sh · 14h ago

The linked article is a minimalist, plain-text weather forecast site called Brolly, built by someone frustrated with the UK Met Office's recent redesign. The thread largely split into two camps: people who loved the retro, compact aesthetic and people who immediately wanted a true plain-text mode for curling from the terminal, like wttr.in. The creator got plenty of feature requests — humidity, sunrise times, a proper text/plain endpoint, and better location search (it fails on state names like "Texas") — and seemed genuinely open to them. A recurring surprise was how many commenters noted the output is perfectly structured for feeding straight into an LLM, which the creator hadn't set out to design for.

GM Backs Sodium Ion Batteries for U.S. Grid Storage [comments]

171 points · 72 comments · spectrum.ieee.org · 10h ago

The article profiles Peak Energy, a Colorado startup backed by GM that claims its sodium-ion batteries can undercut lithium-iron phosphate for grid storage over the battery's full lifetime, despite lower energy density and current reliance on Chinese cell suppliers. Hacker News immediately zeroed in on the fact that Peak is still buying cells from China and won't have its own factory online until 2027, with several people questioning whether GM is really testing Peak's cells or just rebadging Chinese ones. A big split emerged between grid-scale optimists—who argued sodium's superior cycle life (20,000 cycles at 80% capacity vs. LFP's 8,000) and passive cooling savings make it a no-brainer for utilities—and home-storage enthusiasts frustrated that consumer sodium batteries aren't available yet, with one commenter noting the wider voltage range makes them incompatible with existing inverters. The thread also got into the weeds on Natron Energy's collapse, with a rare bit of insider detail: the company supposedly had product sitting in warehouses waiting for UL certification and needed only a $5M bridge loan, but still got sold for scrap amid accusations of investor shenanigans. The practical takeaway from the discussion is that sodium ion looks real for utility-scale deployments starting now, but anyone waiting for a home battery upgrade should probably stop holding their breath.

Running a 28.9M parameter LLM on an $8 microcontroller [comments]

165 points · 35 comments · github.com · 13h ago

Someone crammed a 28.9 million parameter language model onto an $8 ESP32-S3 microcontroller by stashing most of the weights in flash instead of RAM, using Google's Per-Layer Embeddings trick, and got it writing TinyStories at 9.7 tokens per second to a little screen. The thread fully bought into how wild that is—a hundred times bigger than anything previously run on a chip this small—and the immediate reaction was "okay, what else can we shove this architecture into?" One person immediately pointed out that viable TTS models are also around 20-30M parameters, so a no-network audio reader is the obvious next step, while others started debating whether you could chain multiple MCUs together with a layer per chip to sidestep bandwidth limits (the consensus being: run the numbers before you waste your time, because the inter-chip communication will kill you). A few people pushed back on the whole premise, arguing that running models on such constrained hardware is "beyond impractical" and doomed from the start, but that got batted down hard—the whole point is that this is exactly the kind of insane engineering constraint that teaches you something real. Someone also asked why this trick can't scale to running huge models off a desktop CPU with flash or SSD, and the answer came back clear: the ESP32's flash is memory-mapped with near-zero latency, whereas NVMe requires command lists and drive microcontrollers that add orders of magnitude more delay, and nobody has trained a full-scale model with per-layer embeddings anyway.

Tile's security is so bad it's a feature for stalkers [comments]

156 points · 43 comments · blog.adafruit.com · 13h ago

The linked article wasn't available to this summarizer; from the discussion, a new security paper out of arXiv reveals that Tile trackers broadcast unencrypted location data that anyone can pick up, meaning someone can stalk a person who already owns a Tile without ever planting a device on them. The HN crowd immediately connects this to Tile's owner, Life360, noting the company has always monetized location data and has started advertising heavily now that it's a more profitable business. Several people push back hard on the premise, arguing that cheap Chinese GPS transponders are a bigger actual stalking threat and that the sensational title is a deliberate smear campaign, but others counter that Tile's massive installed base of unwitting users makes it uniquely dangerous—you don't need to hack a specific device, you just scan for any nearby Tile. One of the paper's authors shows up in the thread and points to follow-on research comparing Apple and Google's approach, which uses end-to-end encryption with rotating public keys derived from a pairing secret, making the contrast with Tile's blatant insecurity even starker.

A 77-year-old Republican man is staging a solo protest against Flock cameras [comments]

143 points · 78 comments · www.cltampa.com · 10h ago

A 77-year-old Republican man in Florida is blocking a Flock surveillance camera with a pool skimmer and a protest sign. The thread immediately zeroed in on the article’s decision to highlight his party affiliation—plenty of people called it a cheap editorial framing to spark a pre-judged emotional response, while others argued it’s relevant because Republicans are typically seen as pro-law-enforcement, and seeing one oppose mass surveillance makes the story more newsworthy and politically useful. The conversation split into a debate about whether surveillance ought to be a bipartisan issue (most agreed it should be, but were cynical about it actually becoming one), and a darker subthread where a few people went deep into unrelated grievances about harsh protest sentencing in Texas, pulling the discussion toward broader comparisons to authoritarian regimes. There was also a fair bit of tactical chatting: some liked the idea of just blocking the camera’s view with balloons or a screen showing fake plates, while others noted that Florida has already made it illegal to trick ALPR cameras, and one person pointed out the article itself links to a site that simply maps the cameras without encouraging vandalism.

The Fedora 45 Sausage Factory [comments]

143 points · 39 comments · supakeen.com · 20h ago

The article is a detailed technical walkthrough of how Fedora 45 turns source code into final deliverables—ISOs, cloud images, container images, and OSTree deployments—covering the entire pipeline from a packager's git push through Koji builds, Bodhi gating, Pungi compose orchestration, and openQA testing. The HN crowd mostly treated it as a goldmine for troubleshooting, with one person saying it finally gave them a starting point for tracking down a root file permissions bug that had stumped them across Fedora versions. Several people used the thread to ask where to volunteer, pointing new contributors toward the Fedora Join guide, the blocker bug process during release freeze cycles, and the infrastructure ticket backlog. A skeptical corner of the discussion went sideways into IBM's "bluewashing" and whether Red Hat's parent company is bad news for Fedora long-term, though that line got pushback from someone noting that Techrights (the source of the criticism) lost a libel case for harassment and is hard to take seriously.

LLM Usage in Debian: Three Proposals [comments]

139 points · 119 comments · www.debian.org · 12h ago

Debian is facing a vote on four competing proposals about whether to ban, restrict, or accept LLM-assisted contributions to the project, with Proposal A pushing for a total ban on any Debian-specific work involving LLMs, Proposal B allowing it under conditions like disclosure and DFSG compliance, and Proposal C asking contributors to voluntarily avoid LLMs while banning AI-generated mailing list posts. The thread dug into the fuzzy line between "assistance" and "use" — whether using an LLM to find bugs but writing the patch by hand counts as banned under Proposal A, or whether querying an LLM as a search replacement would be considered "assistance." A sharp split emerged between people who see this as a necessary stand for quality, community, and ethics versus those who argue it's a losing battle that will hurt Debian's security and non-English-speaking contributors, with several commenters pointing out that Proposal A could actually make Debian less safe if it discourages LLM-aided vulnerability discovery. The non-English angle got real traction: multiple people argued that for many global contributors, LLMs are the only practical way to read or produce documentation in their native language, and that banning translations would just drive that work into private chats where errors can't be corrected. One recurring pushback was that enforcement is essentially impossible — the project is relying on good faith and the fact that "slop enthusiasts can't resist outing themselves" — but the deeper disagreement is whether banning something unenforceable is still worth it as a statement of values.

Engineering management after the cost of code collapsed [comments]

133 points · 184 comments · karimjedda.com · 16h ago

The article argues that the cost of producing code has collapsed thanks to LLMs, which means roughly half of traditional engineering management wisdom is now built on broken assumptions, while the other half—centered on human coordination, judgment, and accountability—matters more than ever. The HN thread largely agreed with the core thesis, but a huge chunk of the discussion got sidetracked by the fact that the author admitted to using an AI model called "Gemini 4" for editing, which doesn't exist, forcing a correction to "Gemma 4"—this fueled a meta-debate about whether the entire post was AI slop, with some dismissing it outright after spotting the AI-generated hero image. Others, however, defended the content as thoughtful and well-reasoned, arguing that the AI-assisted writing style is a separate issue from the validity of the management insights. A strong counter-current pushed back on the article's premise directly, arguing that the real cost of code has actually *increased* because AI accelerates the accumulation of unmaintainable technical debt, and that upper management now has fresh validation that engineering is just a cost center to be optimized away.

DeepSeek pause fundraise after comments on compute gap to US leaked (transcript) [pdf] [comments]

125 points · 92 comments · github.com · 8h ago

DeepSeek's founder, Liang Wenfeng, reportedly told investors in a leaked transcript that the company is pausing its second fundraising round, explicitly framing China's AI disadvantage as a resource gap rather than a talent problem—he needs 200,000 Huawei 950 chips to train a frontier model but can only get 16,000, and expects the crunch to last at least three years. The thread immediately split over whether the pause is actually about the leak itself (Liang being furious about private comments going public) or about the logical bind of raising cash you can't spend on hardware due to export restrictions. A significant contingent pushed back hard on the "compute gap" framing, arguing that hardware scarcity forces algorithmic ingenuity—pointing to the Bitter Lesson and claiming American AI is bloated, not superior, and that DeepSeek's constraints will accelerate more efficient model design rather than ceding ground. Others noted the irony that the US restricted chip sales to China and now China is blocking its own companies from buying NVIDIA chips, making the whole debate less about American superiority and more about a political bottleneck that could spur domestic GPU competition within a few years. A few commenters with deep optimization expertise jumped in to say they've already built faster attention kernels and believe sparse models could deliver the same capabilities with 0.1% of the compute, suggesting the "scarce resources" angle might actually be a blessing in disguise for Chinese labs.

Producing ammonia and fertiliser using wind power in Morris, Minnesota [comments]

117 points · 85 comments · ammoniaenergy.org · 12h ago

A new facility in Morris, Minnesota is using wind power to produce low-carbon ammonia fertilizer on-site, aiming to stabilize supply and prices for local farmers. The Hacker News thread quickly turned into a familiar political brawl over green energy, with one side arguing the US is squandering its technological lead due to fossil fuel protectionism while others countered that fossil fuels remain the cheaper option. That cost-effectiveness claim was met with a wave of pushback pointing out that fossil fuel prices don't account for heavy direct subsidies, military expenditures to secure supply, and massive externalized costs like health impacts and wildfire damage. A few commenters tried to get back to the engineering, noting the project’s key innovation is letting the Haber-Bosch process run dynamically to match fluctuating wind input, but the conversation mostly stayed stuck on whether renewables can ever be economical without central planning.

An ESP32 based plane radar for my desk [comments]

114 points · 21 comments · blog.ktz.me · 5h ago

A desk toy that uses an ESP32 to tap into the global ADS-B network and display nearby aircraft on a small round screen. The HN crowd immediately pounced on the "radar" framing—most of the thread is people insisting that pulling already-decoded ADS-B data from adsb.fi is not radar in any meaningful sense, with several commenters wishing for someone to actually build a primary radar on an ESP32 despite the absurd difficulty of that task. A deeper technical debate broke out about whether the chip could even decode ADS-B from raw RF, with one side arguing the math is trivial and another countering that the ESP32-C3 lacks the ADC bandwidth to capture the signal without an external USB SDR. Others pointed out that you could fit a Pi Zero and an RTL-SDR dongle into a similar enclosure for a truly self-contained unit, bypassing the cloud dependency entirely.

Show HN: I mapped every US golf course [comments]

108 points · 56 comments · golfcoursebrowser.com · 5h ago

Someone built a free directory of every US golf course, pulling data from OpenStreetMap and cross-referencing it with scorecards and USGA ratings, since Google’s golf search results are terrible. The HN crowd immediately started stress-testing the dataset, pointing out missing courses (a couple in Alaska and Washington, a weird redirect on one course’s website link) and debating whether the map could show polygon boundaries for course sizes. A big split emerged over golf’s environmental impact: some argued courses are an elitist waste of water on public land, while others countered that most US courses are public, cheap, and often use recycled water or rely only on rainfall. The creator got offers to collaborate from someone running a similar wiki, and a warning about legal liability if the project accidentally publishes a “phantom settlement” via scraped data. The thread also wandered into Florida retirement communities where old people drive golf carts from garage to clubhouse, and one commenter wondered if the map could be correlated with Parkinson’s disease data.

A shell colon does nothing. Use it anyway [comments]

108 points · 36 comments · refp.se · 18h ago

The article is a love letter to the shell's built-in null command (`:`), a built-in that does nothing but evaluate and discard its arguments, and shows how to use it for compact parameter expansion tricks like `: "${1:?missing argument}"` for one-line error handling. The HN crowd was sharply divided: many shell veterans found the file truncation (`: > error.log`) and writability checks clever but misleading, noting the author glosses over side effects like file creation, while others argued these one-liners turn already-arcane bash into unreadable "line noise" that has no place in any script longer than two screens. A strong contingent pushed back hard, arguing that bash is a nightmare language full of footguns and that anyone writing non-trivial scripts should just use Python or Perl instead, especially now that LLMs make generating those languages trivial. The author jumped into the fray himself, defending the educational value and conceding practical concerns, but the underlying split was clear: a camp that sees these shortcuts as elegant mastery versus a camp that sees them as an invitation to disaster.

Inflect-Micro-v2: complete voice in 9.36M parameters [comments]

105 points · 9 comments · huggingface.co · 7h ago

The linked article introduces Inflect-Micro-v2, an open-weight text-to-waveform speech synthesis model that packs a complete, local TTS pipeline into under 10 million parameters. The Hacker News crowd was genuinely impressed by the sound quality for such a tiny footprint, with several people noting it beats the robot-voice curse and doesn't sound like historic TTS tools. A few commenters jumped on the practical side, one quickly replacing an old ONNX model with it and another wanting a voice clone option, though the thread mostly stayed grounded by the model's clear limitations: it's English only, has one fixed male voice, and explicitly isn't zero-shot cloning. Some dissent surfaced from those who found the inflections weird or comparable to older TTS like Ivona's voices, but the prevailing take was that the trade-offs are entirely reasonable for a model this small and deployable on CPU.

Cloudflare's new AI traffic options for customers [comments]

103 points · 73 comments · blog.cloudflare.com · 9h ago

Cloudflare announced new traffic controls letting website owners block AI crawlers based on whether they’re doing search indexing, acting as real-time agents, or training models—and they’re changing defaults in September to block multi-purpose crawlers like Googlebot that mix search with training unless site owners opt out. The thread zeroed in on that Googlebot default: since Google uses the same crawler for search and Gemini training, the new “block training” setting will effectively cut off Googlebot from sites that adopt it, and many commenters saw this as long-overdue pushback against Google’s “predatory” bundling of search access with AI training consent. Skeptics called the whole thing an honor system—no real enforcement beyond robots.txt and Cloudflare’s “Verified” label—and pointed out that Google has historically ignored such signals. Others were annoyed that Cloudflare is simultaneously selling tools to build agents while blocking them, calling it a conflict of interest, and a handful of people just wanted to know when Cloudflare’s year-old pay-per-crawl marketplace would actually let them get paid rather than giving them more blocking knobs.

Clinical failure rates over the decades: yikes [comments]

93 points · 65 comments · www.science.org · 9h ago

The article, from Derek Lowe's blog on Science.org, compiles data showing that clinical trial failure rates for new drugs have hovered around a staggering 90% since the 1960s, with a brief dip to 80% in the 70s and 80s before climbing back up. The HN thread immediately pushed back on Lowe's dramatic comparison to the auto or airline industries, arguing that drug development is more like early-stage engineering or physics where the underlying principles aren't fully understood—several people pointed out that jet engine prototypes also fail at high rates in the simulator, not in public, so the comparison is misleading. A significant split emerged over whether the system is worth reforming: some argued the high failure rate is actually *too low*, because if we tried more ambitious 1% shots we might cure more diseases, while others countered that the sheer cost of running a single trial (hundreds of millions) makes that fantasy impractical without government backing. There was a notable tangential debate about whether biologics like VEGF-targeting antibodies are now subject to genuine iterative engineering (fixing "bugs" in molecular architecture), with some claiming that view of drug design as binary and non-iterative is outdated. A few people noted that the failure rate is remarkably *stable*, not rising, which suggests scientific advances are exactly keeping pace as we mine the low-hanging fruit of easier drug targets.

Memory safety absolutists [comments]

87 points · 124 comments · itsallaboutthebit.com · 13h ago

The article argues that a new wave of “memory safety absolutism” from projects like Fil-C and Zig’s upcoming compilation mode is unfairly dismissing Rust as unsafe because it allows `unsafe` blocks, ignoring Rust’s proven track record (e.g., Android data showing 1,000x fewer vulnerabilities than C/C++). The HN thread largely split into two camps: pragmatists who see Rust, Fil-C, and GC approaches as complementary tools with different trade-offs, and a smaller group arguing that only compile-time guarantees truly count—calling Fil-C’s runtime panics a marginal improvement over crashes, not real safety. Several commenters pushed back hard on the article’s framing, noting that Fil-C’s author clearly uses anti-Rust rhetoric as a marketing tactic to draw attention to his project, and that the real story is that both static analysis and runtime checks have value. The discussion also dug into Rust’s actual limitations—like its inability to express self-referential borrows or handle arena allocation ergonomically—with some pointing to Scala 3 and OxCaml’s modal type systems as a better middle ground. One recurring note was that Rust’s biggest achievement isn’t just memory safety but curating a coherent feature set that makes working in a non-GC systems language tolerable, which Fil-C and Zig don’t yet match.

Wind turbine is being used to produce zero-carbon "green ammonia" fertilizer [comments]

83 points · 93 comments · energiesmedia.com · 15h ago

The article describes a pilot project in Minnesota where a wind turbine powers an on-site chemical plant that produces "green ammonia" fertilizer using only wind, water, and air, aiming to replace carbon-intensive imported synthetic fertilizer. The HN thread spent most of its energy pushing back on the article's framing, pointing out that the wind turbine isn't doing anything magical—it's just using its electricity to run electrolyzers and an air-separation unit, then feeding hydrogen and nitrogen into what is recognizably the Haber-Bosch process. Several people argued that this distributed, small-scale approach is likely less efficient than simply electrifying existing large ammonia plants and powering them with grid renewables, though others countered that on-site production can make sense when the grid can't take the power, with payback periods as short as a few years for farmers who only need ammonia a few days a year. A notable tangent emerged around alternative plasma-based processes like Birkeland-Eyde, though those were dismissed as far less energy-efficient. The article itself was criticized as poorly written—possibly AI-generated—and a few commenters pointed to a similar piece on the same site about an underwater turbine as evidence of sloppy content farming.

Show HN: Bribes.fyi – Know before you go. New feature added [comments]

74 points · 47 comments · bribes.fyi · 12h ago

The site, `bribes.fyi`, is a crowd-sourced, India-focused platform where people can report and look up bribe demands at specific government offices (police, RTO, passport office, etc.) before they go. The HN thread split into a few camps: some took the project as a dead-serious and depressing necessity for navigating Indian bureaucracy, while others argued the whole concept of "bribery culture" isn't unique to India—pointing to U.S. lobbying or even domestic tipping culture as functionally similar systems. A more cynical contingent questioned the project's technical merit, calling it "vibe coded" and suspicious for soliciting donations right away, and then there was a minor meta-drama about the developer reposting the link four times in a week, which got flagged before being unflagged. The developer mostly held their ground, clarifying they built what they know (India) and that donations buy no special treatment, while the broader conversation wandered into anecdotes from international business travelers who treat bribery as a mundane, almost bureaucratic cost of doing business abroad.

What is happening to jobs? Separating AI hype from reality [comments]

73 points · 80 comments · siepr.stanford.edu · 9h ago

A Stanford policy brief argues that, contrary to the “AI jobs apocalypse” hype, AI’s effect on aggregate employment has been small so far, with the overall soft labor market driven more by pandemic over-hiring and interest rate hikes than by automation. The thread zeroed in on a major caveat the paper itself might be underplaying: all this data stops at the end of 2025, right before coding agents like Claude Code and GPT-5.3 actually started working well in early 2026, so the study could already be obsolete. A big split emerged between people who say they’ve seen a genuine productivity inflection point since last December and skeptics who see the same moving-goalposts pattern as Tesla’s self-driving promises—especially since corporations are now demanding proof of ROI on expensive API bills. Several laid-off developers chimed in to say their own job searches contradicted the panic narrative, arguing that most layoffs blaming “AI” were really just cost-cutting after the pandemic hiring binge.

The Silurian Hypothesis (2020) [comments]

68 points · 112 comments · www.theparisreview.org · 18h ago

The Paris Review column presents the Silurian Hypothesis, a speculative idea from two scientists suggesting that an advanced industrial civilization could have existed on Earth tens of millions of years ago and then vanished without a trace. The Hacker News thread immediately tore into the core premise, with several people pointing out that we regularly find fossilized skin, bones, and footprints from hundreds of millions of years ago—so the claim that a planet-spanning civilization would leave no detectable evidence doesn't hold water. A major back-and-forth erupted over whether human artifacts like concrete, plastics, and carved stone could survive deep time, with one side arguing that plate tectonics and erosion would erase everything in a few million years, while the other side countered that we routinely find fossils and rock formations that have survived for hundreds of millions of years right on the surface. The discussion got into the weeds on rock weathering rates, how subduction actually works, and why a civilization of billions with mines, tunnels, and radioactive waste would inevitably leave some anomalous geological signature. The article itself takes a frustrating turn by conflating the scientists' original hypothesis with modern conspiracy theories about lizard people and ancient aliens, which the thread roundly rejected as the author going off-script rather than representing the actual scientific idea.

Politician reads AI prompt during assembly [comments]

67 points · 43 comments · www.youtube.com · 15h ago

A Canadian politician named Bill Oliver read verbatim from an AI-generated speech during a legislative assembly, accidentally including meta-text like "Certainly, I can write a speech in this style" that was clearly part of the LLM's response rather than a prompt. The thread quickly split on who bears the blame: some argued the politician was too lazy or incompetent to proofread his own material, while others pointed the finger at underpaid staffers who might have been cutting corners. More than a few people corrected the article itself, noting that news.com.au got the terminology wrong — the politician was reading an AI *response*, not the prompt. A significant chunk of the discussion drifted into broader existential griping about how everyone seems to be blindly pasting AI slop now, with several people saying this is just the new normal and we've collectively lost 30 IQ points in two years.