HN Brief: 2026-05-25
Today’s HN was a ground war over AI coding agents: whether they're a productivity breakthrough or a slop factory that collapses under real architectural constraints. George Hotz kicked off the anti-agent case with a scorching essay, a new paper coined "constraint decay" to name the failure mode, and a guide on not letting Claude play architect drew sharp rebuttals from engineers who say they get great results when they lead, not follow. The other major throughline was hardware cost panic—memory prices for AI chips are spiking, and the hobbyist PC market is feeling the squeeze, with a $280 RAM kit now going for over a grand. A couple of nostalgia pieces on 80s computing and open-sourced Microsoft BASIC offered a quieter counterpoint.
Threads most worth clicking: "Claude is not your architect. Stop letting it pretend" for the live-fire debate over whether the tool is dangerously agreeable or just badly prompted. "Constraint Decay" for the paper that names the phenomenon everyone on the agent threads is experiencing. "Memory has grown to nearly two-thirds of AI chip component costs" if you want the hardware-side panic and the "survival mode" diagnosis for home builders. "Omarchy Is Not A Distro" for the surprisingly substantive fight over what counts as open source when a celebrity bundles dotfiles and calls it a distribution. "Jira Is Turing-Complete" for the grim humor of proving your project management tool is formally as powerful as it is practically unusable.
DeepSeek reasonix, DeepSeek native coding agent with high caching and low cost [comments]
537 points · 225 comments · esengine.github.io · 18h ago
The article pitches Reasonix, a terminal-native coding agent built exclusively for DeepSeek's API, claiming it exploits DeepSeek's byte-stable prefix cache to achieve 90%+ hit rates and collapse input-token costs to a fifth of normal. HN immediately pushed back: several people said they already get excellent caching with generic tools like Opencode or a quick hacked-together bridge to Codex, questioning what Reasonix adds beyond being "DeepSeek native." The discussion split into a technical debate over cache stability—some argued that Opencode's tool-call pruning breaks cache, while others showed 98.6% cache hit ratios with it—and a much larger political tangent about trusting a Chinese company with code, which quickly turned into a proxy war over which government (China vs. US) is more dangerous for espionage, with Europeans chiming in to say they're now more worried about Washington than Beijing.
Memory has grown to nearly two-thirds of AI chip component costs [comments]
373 points · 385 comments · epoch.ai · 15h ago
The piece from Epoch AI breaks down how high-bandwidth memory (HBM) has ballooned from 52% to 63% of total AI chip component costs in just two years, with absolute spending on memory jumping from $12 billion to $32 billion, outpacing every other part like logic dies or packaging. The thread immediately veered into a consumer panic, with users swapping horror stories about DDR5 RAM prices doubling or tripling in a year—one person paid $280 for a 96GB kit in late 2024 and now sees the same kit at $1,050, while others note they bought used machines where the memory alone is now worth more than the whole system. A few pushed back, arguing that current prices just reflect the end of a brutal price-war trough when DRAM makers were bleeding money, and that history shows memory pricing is cyclical rather than permanently inflated. The bigger debate split between those betting that algorithmic efficiency gains will eventually crater demand—invoking Jevons paradox both ways—and those warning that better models will just consume even larger contexts, making the memory squeeze a structural shift rather than a blip. The consensus among hardware hobbyists was grim: rising memory costs are killing the home-built PC market, with one person calling it "survival mode" for component makers, while a skeptical minority predicted the AI bubble itself would pop long before ram prices ever came back down for gamers.
Show HN: Audiomass – a free, open-source multitrack audio editor for the web [comments]
331 points · 67 comments · audiomass.co · 16h ago
A developer built AudioMass, a free, open-source multitrack audio editor that runs entirely in the browser with no backend, weighing in at just 108kb of code. The HN crowd was genuinely impressed by the tiny footprint and the "constrained creativity" philosophy behind it, though several people immediately started pushing for more features like MIDI support, VST compatibility, and tracker module file formats — requests the creator politely deflected by citing the intentional size limits. A significant thread spun off around collaborative music-making and version control for DAWs, with people arguing that while the GitHub analogy sounds great, music stems are the product itself and musicians usually want attribution or royalties, not public forks. Others debated whether this could replace desktop tools like Audacity, with a split between those who find Audacity's UI unbearable even after decades and those who swear by its silence-detection feature for audiobook chaptering — though several commenters pointed to Ardour and Ocenaudio as better desktop alternatives for serious work.
The Eternal Sloptember [comments]
288 points · 210 comments · geohot.github.io · 4h ago
George Hotz’s latest screed argues that AI agents will be one of software engineering’s costliest mistakes, producing broken code that looks right but fails in novel ways, and warns that high performers use AI as a close-read tool while low performers just flood the zone with slop. The thread split sharply: the true believers insist the models hit a new capability ceiling earlier this year and that anyone not doing agentic coding is missing the acceleration of human progress, while skeptics counter that the underlying model hasn’t actually changed—it just got better at autocomplete, and the real breakthroughs are about harness and tooling, not reasoning. A long back-and-forth erupted over whether calling LLMs “sophisticated statistical models” is a tired 2024 dismissal or a necessary reality check, with Hotz himself chiming in to clarify that the distinction isn’t about consciousness but about _process_—models didn’t have a childhood, didn’t learn to program through self-directed experiments, they just learned to mimic the distribution. Several commenters drew explicit parallels to the cryptocurrency hype cycle, predicting AI will settle into something useful but not world-changing, though others pushed back hard saying crypto was never actually _used_ productively while AI already is, every day. The term “Eternal Sloptember” itself got a standing ovation, with veteran netizens connecting it back to the Usenet Eternal September and worrying that HN itself might be suffering a new influx of agentic noise.
Claude is not your architect. Stop letting it pretend [comments]
248 points · 175 comments · www.hollandtech.net · 13h ago
The article argues that letting Claude design software architecture is dangerous because it’s pathologically agreeable, builds generic “Jenga towers” that ignore real-world constraints, and turns senior engineers into rubber-stamping ticket implementers while the tool faces no accountability. HN largely pushed back on the central claim that Claude can’t say “no” — many commenters shared experiences where Claude actually gets snippy and insists on its own approach, and that the key is prompting it to be critical or giving it a detailed spec rather than asking for open-ended design. A strong split emerged: some agree that managers over-trust AI output and cut corners, while others counter that the article underestimates how much an experienced engineer can guide Claude into producing solid results by providing a clear architecture upfront, after which the AI executes nearly one-shot. Several commenters also noted that LLMs are forcing teams back into proper practices like specs and unit tests, though a minority sees it driving even more sloppy “vibe coding.”
The four-day workweek in Australia: insights from early adopters of 100:80:100 [comments]
242 points · 222 comments · scienceaim.com · 13h ago
Australia ran a trial of the 100:80:100 model—full pay, 80% hours, 100% output—across 15 companies, and 14 kept the four-day week afterward with no productivity drop and some gains. The HN crowd immediately split into two camps: one arguing that the results are obvious and that the only reason we don't have a four-day week everywhere is that abusive employers need control, and the other accusing the study of being a glorified opinion survey with a laughably small sample (15 interviews) and no control group, calling it "not science." Several people brought up the Hawthorne effect and novelty effect, pointing out that any workplace intervention tends to boost productivity temporarily just because people feel watched. A more practical thread emerged around consulting firms: how do you bill 80% of the hours and still make the same money? The answer was basically "raise rates," which led to a secondary argument about whether labor law changes or demographic shifts (declining birth rates reducing labor supply) are the only real paths to a shorter week.
Migrating from Go to Rust [comments]
240 points · 225 comments · corrode.dev · 13h ago
The article is a pragmatic migration guide from Go to Rust, focused on where Go’s nil pointers, data races, and GC pauses push backend teams to consider Rust’s compile-time guarantees. The HN thread immediately latched onto dependency bloat in Rust—one person working on both languages said a Rust project using rusqlite, clap, ratatui, and Tauri pulled in over 400 crates, while Go’s stdlib covers half that, sparking a debate over whether Rust’s crate ecosystem is “npm all over again” or a necessary evil of splitting crates for compile parallelism and reuse. The other dominant split was between developers who love Rust’s enforced correctness for production services and an SRE who finds Rust services worse to operate—verbose logs, contextless errors, and “impossible” memory leaks that still happen. People also pushed back on the article’s praise for Rust’s error handling, arguing Rust actually has *three* incompatible error patterns (io::Error, thiserror, anyhow) that make chaining painful, while Go’s simple `if err != nil` is at least uniform and obvious. A surprising tangent: some teams are finding LLMs generate better Rust than Go because the type system gives the AI strong rails, though the flip side is that Rust’s 10x slower compile times become a bottleneck for agentic coding loops.
Constraint Decay: The Fragility of LLM Agents in Back End Code Generation [comments]
227 points · 122 comments · arxiv.org · 19h ago
A new paper from academics introduces the concept of “constraint decay,” showing that LLM-based coding agents fall apart when asked to follow real architectural rules (like ORM conventions or specific framework patterns) rather than just pass functional tests. The HN crowd immediately dug into the methodology, noting the study used GPT-5.2 instead of the code-optimized 5.2-codex variant and relied on open-source harnesses rather than OpenAI’s own agentic tooling, so the specific performance numbers should be taken with a grain of salt. A major thread argued the deeper issue is that you cannot optimize for two objectives at once — RL training works great for verifiable functional correctness, but structural constraints force the model to guess at incomplete specs, which is fundamentally a different problem. Several people with production experience confirmed the effect anecdotally: agents routinely ignore architectural plans, anchor on their own wrong first guesses, and only follow style guides if you provide explicit code examples rather than written instructions.
Childhood Computing [comments]
197 points · 94 comments · susam.net · 19h ago
The article is a personal memoir from Susam Pal about learning Logo programming on hand-me-down IBM PCs in the early '90s, complete with floppy disks, no hard drives, and graph-paper coding at home. The thread immediately latched onto the olfactory nostalgia—several people confirmed that the smell of old computer labs is a real time machine, with one commenter noting they could still draw a map of their high school lab decades later. The discussion then branched into two main directions: a wave of shared origin stories (Apple IIs, TRS-80s, Geocities view-source hacking, QBasic, RPG Maker epiphanies) and a pointed critique of how Microsoft and other vendors locked down programming tools behind expensive licenses, forcing kids toward pirated copies or free alternatives like GNU tools. A long technical tangent emerged when someone posted a geometry problem about nesting circles for a CNC project, and another commenter walked through the full algebra to solve for the center coordinates—a classic HN diversion into applied math that the original poster immediately linked back to their OpenSCAD code.
Greg Brockman interview [video] [comments]
196 points · 206 comments · fs.blog · 23h ago
The thread is about a Greg Brockman interview, but no one really discusses the podcast itself—instead, it immediately detours into a bitter argument over Brockman's leaked diary from the Elon Musk lawsuit, where he wrote “Financially what will take me to $1B?”. Half the thread takes that as proof he's a grifter whose altruistic OpenAI mission was always a cover for cashing out, while the other half shrugs that wanting a billion dollars isn't some scandalous diary confession, and if that's the worst thing exposed, he looks fine. The conversation eventually collapses into a sprawling, furious debate about whether wanting a billion dollars is inherently immoral, whether anyone who says they wouldn't want it is lying, and whether billionaires are a policy failure—with side tangents into whether the US pulling foreign aid counts as engineering deaths, and whether training LLMs on copyrighted books is theft. The core split is between people who see the diary as exposing a hypocrite and people who think the anti-billionaire moralizing is itself a narrow-minded grift.
Usborne 1980s Computer Books [comments]
176 points · 53 comments · usborne.com · 16h ago
Usborne has made free PDFs of its iconic 1980s computer books available, covering BASIC programming and machine code for machines like the ZX Spectrum and Commodore 64. The HN crowd dug into whether these books should be ported to a modern language, and the split was sharp: the anti-Python faction argued that significant whitespace makes printed code listings a nightmare for kids who have to type and edit manually, while the pro-Python camp pointed out that Python is already the educational standard and tab-counting is less painful than old-school line numbers and GOSUB. A few people are actively building BASIC interpreters or porting the games to HTML/JavaScript, but the consensus was that the magic of the originals came from removing unnecessary barriers—and that Python, despite its merits, introduces exactly the kind of incidental complexity the books were designed to avoid. Another tangent landed on just how hard it was to actually build the robot from the Usborne robots book in the pre-Web era, with mail-ordering motors and convincing dubious parents, versus how trivial the same project would be today with Arduinos and STL files.
Omarchy Is Not A Distro [comments]
170 points · 155 comments · abyss.fish · 17h ago
The article argues that Omarchy, DHH's much-hyped "Linux distribution," is really just Arch Linux bundled with his personal dotfiles and a heap of proprietary defaults—keybindings that open his own web apps, preloaded scripts for Brave, Dropbox, and NordVPN, and a 37-line terminal config for a tool that shouldn't need one. HN split hard on whether that matters: plenty of people shrugged, saying Omarchy is just opinionated defaults and filling a real gap for macOS refugees who want something that "just works," while others pushed back that calling a config script a distro is dishonest, especially when it's soaking up conference sponsorships and merchandise money that serious distros like Debian can't get. A strong contingent argued the real story is that DHH's audience and marketing muscle did the work, not technical merit—and that the "is it a distro?" debate is a distraction from the fact that Omarchy is essentially a proprietary onboarding funnel for Arch, not a community project. Others pointed out that the same criticisms could be leveled at Pop!_OS or CachyOS, and that gatekeeping what counts as a "real" distro only discourages new Linux users. The thread also wandered into a surprisingly detailed argument about 1Password's Electron version quality, because of course it did.
'AI washing': firms are scrambling to rebrand themselves as tech-focused [comments]
163 points · 150 comments · www.theguardian.com · 16h ago
The Guardian piece documents how UK companies are contorting themselves to rebrand mundane automation as "AI," with PR executives internally rolling their eyes at the yoga-level stretches required to pitch floor-plan scanners and basketball hoops as artificial intelligence. The thread largely sidestepped sympathy for the PR folks and zeroed in on the investors supposedly buying this—almost no one believes the hype reflects genuine technology shifts, but the consensus was that it doesn't matter because the game is Greater Fool theory. Many argued the stock market is so disconnected from reality that smart money just needs to sell before the other fools catch on, pointing to Allbirds' pivot from eco-shoes to an AI infrastructure shell company as a textbook pump-and-dump executed by Wall Street guys, not a real business strategy. Some older-timers drew direct lines to the "cloud" and "blockchain" rebranding waves, noting that marketing has always been about disguising what you already do as the hottest buzzword, and the only real difference now is that the word "intelligence" is getting co-opted to sell dishwashers.
The seed oil panic is hurting my cardiac patients [comments]
155 points · 220 comments · www.statnews.com · 17h ago
The article is a clinical dietitian’s plea that his cardiac patients are junking olive oil for beef tallow and avoiding vegetables because a steamer bag says “seed oils,” all based on a panic he says isn’t backed by evidence from large randomized trials. The thread largely agrees the anti-seed-oil craze is data-free and dangerous, but it splits hard when the conversation veers into who’s really to blame: some argue it’s a convenient gift to the beef and dairy industries, while others jab that the credentialed experts lost credibility because people weren’t getting better outcomes before RFK Jr. showed up. A handful of commenters push back on the article itself, challenging its claim that ruminant trans fats are harmless and pointing out that the cited studies on vegetable oils are short-term and look at proxy markers, not real heart attacks. The real fight isn’t over the oil—it’s over whether the movement is a populist health revolt or just swapping one corporate grift for another, with people using seed oils as a scapegoat so they can keep eating 12 cookies without guilt.
CBP Directive 3340-049B: Border Search of Electronic Devices [comments]
149 points · 105 comments · www.cbp.gov · 12h ago
CBP quietly published a new directive (effective January 2026) that formalizes how agents can search phones, laptops, and even vehicle infotainment systems at the border, making clear that a "basic search" requires zero suspicion at all. The thread immediately zeroed in on the chilling effect this has on international travel, with people saying they now treat trips to the US the same way they treat trips to China—bring a burner device, expect it to be compromised, and be ready to abandon it at the border. A big split emerged over whether this is actually new or just codifying existing practices (the 2009 directive was cited), and several people pushed back against the "China comparison" by noting that US border agents aren't installing malware, just demanding passcodes and copying data. The GDPR crowd pointed out that carrying a work phone with European customer data across the US border is now a genuine legal liability, and the thread generally landed on grim pragmatism: if you're bringing a device into the US, assume it will be searched, because the directive gives them the authority to do so without cause.
A fundamental principle of aeronautical engineering has been overturned [comments]
149 points · 73 comments · www.wired.com · 12h ago
A team at Tohoku University claims that applying microscopic roughness — glass beads or sandblasting — to a surface can cut aerodynamic drag by up to 43.6%, supposedly overturning the 80-year-old dogma that smoother is always better. HN immediately pushed back: golf ball dimples already prove smooth isn't always optimal, though several commenters carefully spelled out that dimples work by tripping turbulence to reduce pressure drag, while this new DMR technique delays transition to turbulence to cut skin friction — opposite mechanisms. A lot of skepticism centered on real-world durability — can a 38-micron coating survive bug splats, rain erosion, and ice buildup on a commercial wing, let alone pass certification red tape? Others noted sailors have been sanding hulls with fine grit for decades to improve laminar flow, questioning why aeronautics is acting surprised, and pointed to existing products like Lufthansa's AeroShark film as proof the idea isn't novel. The thread split between people impressed by the 43% number (with caveats about the narrow transition-zone measurement) and those demanding to see actual fuel-efficiency results on a real aircraft before getting excited.
Mastering Dyalog APL [comments]
142 points · 36 comments · mastering.dyalog.com · 20h ago
The post highlights that the classic "Mastering Dyalog APL" book, originally published in 2009, is being rewritten as an interactive Jupyter Notebook edition to make learning the language easier, especially for building muscle memory with its notorious symbol set. The HN crowd split into two camps: those who argue APL has massive long-term ROI because it rewires how you think about computational problems (comparing it to moving from 2D drafting to 3D CAD), and skeptics who say the language feels more like a puzzle-solving exercise in cramming operations into terse, often inefficient expressions. A significant pushback came from experienced users who insisted that real mastery requires writing full applications, not just solving toy problems, and that APL can absolutely be efficient when written with the interpreter in mind—one person even demonstrated a proper O(N log log N) Sieve of Eratosthenes in response to a critic who wrongly claimed APL's "hello world" prime generator was just a brute-force divisor count. Several people pointed to BQN as a modern, open-source alternative that deserves more attention, and the thread also saw a strong recommendation for TLA+ as a language that actually delivers the "brain rewiring for clear thinking" that APL enthusiasts often promise but don't always deliver.
Ruby for Good [comments]
133 points · 53 comments · rubyforgood.org · 16h ago
The linked article is about **Ruby for Good**, an annual in-person event where programmers build open-source software for nonprofits, held this year at a retreat center in Maryland. The HN discussion quickly pivoted away from the event itself into two major debates: whether Ruby is a smart choice for new projects in the age of "agentic coding" (with a strong faction arguing Rust or TypeScript yield better results with AI coders), and a deeper, somewhat defensive conversation about the Ruby community's supposed "artisanal" or morally-virtuous self-image—referencing old drama like _why's disappearance and MINASWAN. A minor sub-thread also got briefly sidetracked accusing the event's website of being AI-generated (a claim the founder half-acknowledged, saying a Claude-designed branch was briefly deployed). The overall takeaway is that reading the comments feels less like a summary of the event and more like a long, winding therapy session where Rubyists simultaneously defend their language's elegance and grapple with its declining relevance.
Jira Is Turing-Complete [comments]
126 points · 42 comments · seriot.ch · 4h ago
The article demonstrates that Jira is Turing-complete by building a working Minsky machine out of issue types, linked issues, and automation rules — using bug and task counts as registers and epic statuses as a program counter. The HN crowd largely shrugged at the theoretical result, since most workflow engines can be contorted into Turing completeness, but they absolutely lit into the practical experience of using Jira. The real debate split between those who've automated their way around Jira's horrors (using Python scripts or LLMs to do the drudge work) and those arguing that Jira's API is such a fractal nightmare of undocumented custom fields and broken migrations that automated workflows often collapse under their own weight. A few people pointed out that the more interesting question isn't "is it Turing-complete?" but "can you actually tell if a Jira workflow will halt?" — and the consensus was no, you can't, because Jira's own design makes everything unpredictable.
Defeating Git Rigour Fatigue with Jujutsu [comments]
118 points · 123 comments · ikesau.co · 13h ago
The blog post outlines a workflow for Jujutsu (jj) where you create empty placeholder commits representing your ideal commit structure first, then use `jj squash -i` to interactively sort all your real work into those slots, sidestepping the "git rigour fatigue" of keeping commits clean during development. The thread quickly turned into a proxy war between jj enthusiasts and skeptics, with the main fault line being whether jj's branch (bookmark) model is a dealbreaker. A vocal contingent insists jj’s anonymous branches and manual bookmark advancement are non-issues in real collaborative repos—they push clean commit stacks to git-using colleagues without friction—while several people push back hard, arguing that named branches are essential for keeping track of long-lived work, picking up someone else’s abandoned feature, or just orienting themselves after a week away. The jj defenders counter that commit descriptions serve that purpose better and that the ability to mutate history freely (especially conflict recording without immediate resolution) is a genuinely new capability that git couldn't offer even with perfect discipline. Underneath the tool debate, there’s a deeper philosophical split about whether rewriting history is a productive way to spend development time or an unnecessary polish step that distracts from shipping new work.
Perceptual Image Codec: What Matters in Practical Learned Image Compression [comments]
110 points · 34 comments · apple.github.io · 19h ago
Apple’s new PICO codec is a learned image compression system optimized for human perception, claiming 2-3x bitrate savings over traditional codecs like AV1 and VVC while running in 150ms on an iPhone 17. The Hacker News thread quickly split into two camps: one group zeroed in on artifacts like hallucinated sweater textures and mangled bicycle racks, arguing that PICO’s texture synthesis produces details that look right at a glance but are completely wrong on inspection—raising philosophical questions about whether that’s better or worse than traditional blur or blockiness. A second camp pushed back on the benchmarks themselves, noting that PICO compares against video codecs like AV1 and VVC rather than JPEG or JPEG-XL, and questioning why neural network decode speeds on an iPhone 17 nearly match those on a V100 GPU, suggesting the GPU tests may not be saturating the hardware. A subtler strain of the discussion worried about deterministic behavior and legal consequences, with one comment pointing out that if codecs start hallucinating structural details, someone could get convicted based on generated image content—a problem no one had to worry about with JPEG.
FreeBSD Foundation executive director tries daily driving FreeBSD on laptop [comments]
89 points · 67 comments · www.phoronix.com · 15h ago
The article covers FreeBSD Foundation Executive Director Deb Goodkin's attempt to use FreeBSD on a Framework laptop as a daily driver — but her self-imposed bar was "at least 10 minutes a day." The HN thread largely roasted this, with the dominant take being that "10 minutes a day" doesn't qualify as daily driving anything, and that the head of the foundation not dogfooding the OS on a laptop for two decades undermines credibility. Several people pushed back, noting she's a former embedded firmware engineer, that FreeBSD isn't a laptop OS, and that the Foundation and the project (governed by the elected core team) are distinct entities. Others redirected to deeper frustrations — that Foundation money is burning through reserves on laptop/KDE polish instead of the server features that actually make FreeBSD compelling, and that the presentation's "I created a system for learning and success" language read as empty LinkedIn slop. A few longtime FreeBSD users recalled running it on laptops in the mid-2000s with fewer complaints, but the consensus was that this effort, while symbolically positive, looked more like a PR stunt than a serious laptop push.
Microsoft's 6502 BASIC is now Open Source (2025) [comments]
85 points · 27 comments · opensource.microsoft.com · 18h ago
Microsoft officially open-sourced the 6502 BASIC that launched a generation of home computers—the exact 1978 source that shipped as Commodore PET's BASIC V2 and later became Applesoft BASIC, complete with a Bill Gates Easter egg and a joint fix with a Commodore engineer. The thread quickly pivoted to whether this code will run on Ben Eater's famous breadboard 6502 computer, which uses a reverse-engineered fork, concluding it should adapt with a bit of work. People were more curious about *how* Microsoft even found this source—was it still in their archives, or did they have to OCR old printouts?—and the cynics noted that while releasing ancient BASIC is nice, they're still not opening Windows. Someone also pointed out that a Pascal interpreter wouldn't have fit in the 8K ROM, so BASIC was the only practical choice for 1970s micros, which drove a short tangent on Tiny Pascal on the TRS-80.
Building Pi with Pi [comments]
76 points · 39 comments · lucumr.pocoo.org · 14h ago
Armin Ronacher reflects on the experience of running the Pi open-source project in an era flooded with LLM-generated contributions, describing how "slop issues beget slop code" as AI agents produce plausible but wrong diagnoses and over-engineered fixes that pile on defensive complexity. The numbers are stark: out of 3,145 external issues filed in 90 days, 2,504 were auto-closed, only 17% reopened, and fewer than 10% of PRs got merged. The thread largely split away from the article's core argument about maintainer burnout and foundation-building to fixate on Ronacher's proposed term "clanker" for LLM-based tools—a word that sparked an unexpectedly intense and emotional debate. Multiple people argued the term evokes racial slurs, citing TikTok skits where "clanker" has been used as a stand-in for racist epithets, while others pushed back hard, calling the reaction absurd for a word aimed at machines and pointing out that craftspeople use derogatory jargon for tools all the time. The author himself jumped in, genuinely fascinated by the visceral response and insisting it's just a name for a piece of equipment, but the conversation kept circling back to whether calling an AI a clanker is a harmless joke or carries real baggage.
When (if ever) it's appropriate to make jokes before the US Supreme Court [comments]
63 points · 51 comments · www.scotusblog.com · 16h ago
The SCOTUSblog article walks through the Supreme Court's official etiquette guide for attorneys—don't crack jokes, don't interrupt, address justices as "Your Honor," and definitely don't take selfies. Hacker News largely ignored the actual advice and instead erupted into a heated argument about whether the Court itself has become a joke, with many commenters arguing that the institution has lost all credibility due to partisan rulings, stolen seats, and undisclosed gifts to justices like Clarence Thomas. A handful of people pushed back, claiming the criticism is overblown and that gifts from people not involved in cases don't count as bribery, but that view got drowned out by accusations of corruption and demands for court-packing. The thread also turned briefly into a meta-commentary on how nearly every top comment was saying the same thing about the court being illegitimate, with some wondering if they were all bots.
The political polarization of health outcomes in the USA [comments]
45 points · 24 comments · www.nature.com · 12h ago
A study in *Nature Human Behaviour* found that conservative Americans saw their health worsen and mortality rise relative to liberals during the 2010s, with the gap widening sharply after 2020 even for non-COVID causes like heart disease and cancer. The research splits the cause roughly in half: demographic realignment brought less healthy people into the conservative coalition, while a separate, stubbornly persistent gap emerged from conservatives themselves growing less healthy over time, which the authors link to declining trust in doctors and medical advice on the right. The thread immediately seized on causation, pushing back hard that the mechanism might run the other way—people becoming more conservative after getting burned by a broken, for-profit healthcare system—but others countered that distrust fueled by GOP misinformation, especially during COVID, is a more direct driver of poor health outcomes. A major tangent broke out over the Affordable Care Act: several people argued you can't blame conservatives for distrusting a system "shaped by Democrats" when the ACA was actually built on conservative ideas like the Heritage Foundation's individual mandate and modeled on Romney's Massachusetts plan, while others shot back that this nuance doesn't matter because the Republican voting public didn't know or care about the policy's origins. A recurring undercurrent was the rural-versus-urban healthcare collapse, with one person pointing to Australian data showing the same political-health divide tracks more strongly with resource deprivation and hospital closures than with ideology, suggesting the real story is about who gets left behind by the system, not just who trusts it.
West Coast Cities Turn to Vacancy Taxes to Grapple with Housing Crisis [comments]
38 points · 17 comments · www.theurbanist.org · 12h ago
The article examines how Portland, Seattle, Tacoma, and San Francisco are exploring or implementing taxes on vacant commercial and residential properties to nudge owners into renting or selling, citing Vancouver B.C.'s "Empty Home Tax" as a model that slashed vacancy to 0.49% but failed to lower rents. Several people immediately pounced on the article for conflating sky-high commercial vacancy rates (Portland at 27% for offices) with the much lower single-digit residential vacancy rates, arguing the housing shortage is primarily a supply problem caused by cities blocking new construction. A sharp split emerged: one side insists vacancy taxes are a harmless way to squeeze speculative landlords, while the other retorts that you can't tax your way out of a crisis when cities deliberately make it illegal to build—someone even called out that permitting fees aren't the bottleneck, but rather the arbitrary denial of building permits unless developers include subsidized units that kill project economics. A concrete example surfaced of a developer demolishing a historic building to build prime retail space, then leaving it vacant for years rather than leasing at market rates, which one comment called a "tax write-off squat" rather than a struggling enterprise. A dissenting note from Vancouver warned that the tax, combined with a glut of new supply built on cheap financing, has now killed pre-sales, thrown projects into receivership, and locked the market so tight that a 0.49% vacancy rate means people literally can't move there.
Companies Are Just a Graph of Algorithms [comments]
25 points · 14 comments · danielmiessler.com · 3h ago
Daniel Miessler argues that companies are really just graphs of algorithms, and that AI will soon map and automate every workflow, slashing the need for humans. HN was deeply split: one camp dismissed it as a warmed-over version of business process reengineering, value stream mapping, and other management fads that never delivered, while others pointed out that many companies are already using AI to document and automate routine SOP-driven work like claims processing and procurement. The skeptics hammered the idea that LLMs are probabilistic pattern matchers, not true optimizers, and mocked the claim that AI will make opaque processes “transparent and inspectable” when it’s actually a fog of floating-point numbers. A few commenters noted the irony that AI companies themselves haven’t replaced their own marketing or accounting departments, and one invoked the classic “internet is a series of tubes” to lampoon the reductionism.
Star Citizen game has reached $1B in funding [comments]
23 points · 13 comments · robertsspaceindustries.com · 8h ago
Star Citizen’s funding page now shows it has crossed $1 billion in total backer contributions since 2012. The Hacker News thread mostly treats that milestone as an invitation to mock, with one person linking a chain of previous milestone threads to suggest the next is always just around the corner, and another calculating that a game some estimate is only 40% complete now costs more than any game ever made. A few people push back on the easy jokes: the top-voted substantive take argues that the $1B figure is a sensationalized distraction, since the game actually pulls in about $100M per year — comparable to a modest live-service title, not a blockbuster — and all of it goes back into development rather than finishing a product. Anecdotal defenders show up too, describing genuinely immersive fly-through sessions that make the purchase worth it. But the dominant note is weariness: the funding model gives Cloud Imperium Games every incentive not to ship, and one reader sums it up as a 13-year-old business that keeps growing revenue while staying incomplete, sustained by a fan base that just likes what’s there.
No Juniors Today, No Seniors in 2031 [comments]
20 points · 33 comments · www.fbritoferreira.com · 16h ago
The article argues that junior software-engineering hiring has dropped roughly 40% since pre-2022 levels, and because turning juniors into seniors takes 5–7 years, the industry is setting itself up for a severe senior shortage by 2031 — with AI being the excuse boards use to keep cutting entry-level roles. The HN thread largely split between people who thought the data span (late 2024 to early 2026) was too short to call a structural collapse, pointing out that hiring cycles have always alternated between bulging and contracting, and several commenters noted they’ve worked at companies that weren’t hiring juniors long before AI was a factor. A strong contingent pushed back hard that there *will* be plenty of seniors in 2031 because current mid-career and older engineers aren’t all retiring — one 57-year-old dev directly refuted the “retirement by 50” assumption. The thread also got meta: multiple people called the article itself obvious AI-generated slop, which undercut its credibility for many, though others defended its core premise about the apprenticeship pipeline being broken. A tangential but heated debate emerged around whether the real bottleneck isn’t lack of juniors but a hiring process that rejects experienced seniors too — a commenter with 25+ years described struggling to get past resume screening and being downleveled despite current skills, suggesting the “shortage” is partly self-inflicted by terrible hiring practices.
Generated 2026-05-25 08:20 UTC
Generated by Sauron from Hacker News discussions and linked articles.