HN Brief: 2026-06-08
Today’s HN was dominated by a grim reckoning with how AI is restructuring work. The top thread saw a senior backend engineer detail how LLMs have systematically devalued each pillar of his expertise, with the community largely agreeing his hunt for a niche AI can’t touch is futile. This anxiety was mirrored in a thread on the economics of US AI labs, where submarining subscription costs suggest the current pricing is a temporary land-grab. A story of rebuilding a career after addiction, prison, and a felony offered a human counterpoint, but even that thread pivoted to arguments about whether such a path is possible in today’s algorithmic hiring landscape. Elsewhere, a deep dive into Linear’s local-first architecture reignited the old debate over optimistic UI trade-offs, while a paper arguing against LLM anthropomorphism by training a neural net inside *Age of Empires II* drew a sharp philosophical line in the sand.
The threads most worth clicking into: “LLMs are eroding my software engineering career” for the bleakly honest consensus that the commoditization of domain expertise is accelerating faster than the models are breaking; “How’s Linear so fast?” for the sharp pushback on whether “assume it worked” UI patterns are a performance win or a distributed-systems nightmare waiting to happen; “Scientists ejected from diabetes conference” for the absurd, telling moment of an organization ejecting its own editor-in-chief over fear of losing federal funding; “The OnlyFans Economy of American AI” for its culture-war-worthy claim that US labs are a protected cartel losing ground to cheaper Chinese models; and “Algorithmic Monocultures in Hiring” for the damning, RealPage-level structural problem of nearly all employers using the same few vendors to auto-reject the same applicants.
LLMs are eroding my software engineering career and I don't know what to do [comments]
929 points · 892 comments · human-in-the-loop.bearblog.dev · 19h ago
The post is a personal essay from a senior backend engineer with a decade of experience—mostly in finance, payments, and distributed systems—laying out how LLMs have systematically devalued each pillar of his expertise: first domain-specific knowledge, then debugging, and now even code architecture, which the industry is reducing to “taste” and de-prioritizing as codebases get optimized for machine consumption rather than human readability. The HN discussion largely agrees with his diagnosis, splitting over what to do about it: one camp argues that the market simply doesn’t pay for “hand-built” software and that C-grade codebases are now the acceptable default, while a skeptical minority pushes back hard, pointing out that artisan software communities (Handmade Hero, Tsoding) and anti-AI boycotts prove there’s a niche willing to pay for quality—and that the woodworking-as-fallback joke is even less viable because custom furniture markets are tiny, hyper-local, and already saturated with mediocre hobbyists. Others wade into a surprisingly detailed tangent on whether farming or subsistence living is a realistic escape, with most concluding that rejecting industrial society is financially ruinous unless you’re already rich. Underneath it all is a grim consensus that the author’s real problem isn’t wrong—domain expertise is commoditizing fast—but that his hunt for a niche LLMs can’t touch is probably futile because the models keep one-shotting the hard bugs.
Building from zero after addiction, prison, and a felony [comments]
593 points · 266 comments · gavinray97.github.io · 13h ago
The post is a raw, first-person account from a software engineer who spent ages 14–16 in a maximum-security juvenile prison, became a felon at 19, cycled through addiction, and eventually rebuilt his career through a lucky internship, open-source contributions to Hasura, and people who didn't run a background check. The thread largely split into two camps: one side focused on the brutal contrast between his story and today's market, where people noted that his "walked in off the street and got hired" moment would be impossible now thanks to AI resume filters and endless interview hoops. A big, unrelated fight broke out about a fellow user who admitted to working three remote jobs simultaneously via AI-optimized resumes, with some calling it fraud and others saying it's just survival against automated hiring systems. A few commenters who've been through addiction themselves zeroed in on how he actually got and stayed sober, and there was a notable side conversation about the irony of this success story coming from someone now working at an AI company who openly wonders how long his own job will last.
Anthropic, please ship an official Claude Desktop for Linux [comments]
489 points · 279 comments · github.com · 18h ago
A long, detailed GitHub feature request makes the case that Anthropic should ship an official Claude Desktop for Linux, pointing out the irony that Claude Code (the CLI) already runs natively on Linux and that Cowork, the agent mode, actually boots Ubuntu inside itself on macOS—so the Linux code path literally exists inside the product already. HN ran with two main threads: first, a security argument that the whole point of Desktop over CLI is sandboxing, with people sharing their own container solutions (jai, nono, smolvm, zerobox) and pushing back on whether the Desktop app is actually more sandboxed than wrapping the CLI in bubblewrap or Docker; second, a practical debate on what Desktop gives you that CLI doesn't, with defenders listing scheduled local tasks, cross-conversation memory search, and better artifact rendering, while CLI users countered they'd hacked all that together with cron jobs and local SQLite indexes. The thread split cleanly between "just use the CLI, it's fine" and "you're missing the point about sandboxing and credential-free plugin testing," but nobody argued that Anthropic *shouldn't* ship a Linux build—the pushback was more about whether it's a must-fix or a nice-to-have given Linux fragmentation.
How's Linear so fast? A technical breakdown [comments]
391 points · 177 comments · performance.dev · 13h ago
The article breaks down how Linear achieves its snappy performance by inverting the traditional web-app architecture: it runs a database in the browser via IndexedDB, applies mutations locally first, and syncs to the server asynchronously in the background, so the UI never waits on a network round-trip. HN mostly latched onto the "optimistic UI" pattern at the core of this approach, with a strong split forming between developers who love it and those who hate it. Several people pointed out that this is essentially the same idea Meteor.js popularized a decade ago with minimongo, and there was nostalgia for that framework's bold local-first bet. The pushback was sharp and focused on distributed-systems trade-offs: critics argued that "assume it worked" breaks when the network actually fails, leading to phantom saves, lost data when tabs close mid-sync, and a general violation of temporal ordering that users won't notice until they do. Defenders countered that Linear handles the edge cases—persistent retry queues, rollback indicators within seconds, and durable IndexedDB storage—and that for an issue tracker, the trade-off in favor of perceived speed is worth it, but even some sympathetic readers admitted that "the client is really fast" is a rediscovery made by every generation.
Scientists ejected from diabetes conference for distributing journal reprints [comments]
325 points · 209 comments · arstechnica.com · 21h ago
Five senior diabetes researchers, including the editor-in-chief of the ADA's own journal, were physically ejected and had their lanyards taken at the American Diabetes Association's annual conference in New Orleans for handing out reprints of an editorial in that same journal—an editorial criticizing the Trump administration's attacks on biomedical research. The thread immediately zeroes in on the absurdity: it's hard to argue you're violating a conference's code of conduct when you're distributing an article from the organization's flagship publication at the organization's own meeting, which exists specifically for scientific dissemination. Many point out that this only happened because the protest was timed just before an NIH director's speech, proving the administration's political pressure has made the ADA so skittish it will eject its own leadership. A recurring argument is that this isn't really about conference rules but about the weaponization of federal funding—these organizations are terrified of losing NIH dollars, so they preemptively enforce political loyalty. The broader consensus in the thread is that this is a clear sign of a crumbling research ecosystem, and the Streisand effect is already boosting the editorial's readership well beyond what it would have reached otherwise.
Show HN: Lathe – Use LLMs to learn a new domain, not skip past it [comments]
299 points · 55 comments · github.com · 20h ago
This is about a tool called Lathe that uses LLMs to generate hands-on, source-backed technical tutorials you work through by actually typing the code yourself, designed for learning domains where no good human-written tutorial exists. The HN crowd broadly loved the concept and philosophy—many called it a sane antidote to the "LLM does all the work" problem, and several people said they'd been using similar patterns with custom prompts and CLI agents already. A fair bit of the discussion veered into the mechanics of the "CLI + skills" architecture pattern itself, with people trading notes on how to build deterministic tools that hand off specific steps to a coding agent rather than letting the agent run the whole show. A few people pushed back directly, arguing that LLMs are still terrible educators because they hallucinate details the student can't detect and don't build coherent curriculums, and one commenter pointed out that the generated tutorials will still be drawing on human-written content without attribution. Several tangential projects emerged from the thread—a Socratic quiz skill for deeper questioning, and someone building a "terminal ARIA" system to give agents structured workflow control—which suggests the real energy here wasn't just about Lathe but about the broader pattern of making LLMs a disciplined learning aid rather than a crutch.
DeepSeek V4 Pro beats GPT-5.5 Pro on precision [comments]
231 points · 92 comments · runtimewire.com · 6h ago
The article claims DeepSeek V4 Pro beat GPT-5.5 Pro 38.0 to 33.0 on precision in a four-task head-to-head judged by an AI model. The comments immediately tore into the methodology: the judge model was actually grok-4.3, not the retired grok-4-1-fast-non-reasoning the article named, and several people flagged the article text itself as AI-generated based on phrasing like "the matchup feels earned." The real action in the thread was people sharing their own benchmarks showing DeepSeek costs a fraction of GPT or Claude for comparable or better results—one person ran a vulnerability scanner where GPT 5.5 Pro blew through $100 and found 2 bugs, while DeepSeek cost a dollar and found 4, and multiple developers reported swapping Claude Code's API key to DeepSeek and getting similar performance for pennies. The split is pragmatic: cost-wise DeepSeek is a no-brainer, but plenty of people still trust the US labs more for production-critical code and worry about sending data to a Chinese company.
Making peace with your unlived dreams (2023) [comments]
213 points · 115 comments · nik.art · 13h ago
The article is a personal essay about making peace with the dream of becoming a great snowboarder after an orthopedist told the author his knees couldn’t handle it. The HN thread pushed back hard on the framing, arguing that the author never actually snowboarded, so calling it an "unlived dream" dilutes what a real dream is—something you pour years into, not just a passing “that’d be cool.” That semantic debate got heated, with one side insisting a dream that requires zero action is just idle fantasy, while others retorted that plenty of people carry unfulfilled aspirations their whole lives without ever pursuing them, and that’s exactly what makes them dreams. A separate thread went long on whether you can rehab bad knees with monster braces and single-leg squats, while another tangent asked people to interrogate their motivations with “five whys” until they hit rock bottom and realize most dreams are about impressing other people. The most pointed take was that “regret minimization” is a mantra for centibillionaires who’ve wrecked the world chasing more, and it works fine for normal people only if what they’d regret is not spending time with loved ones, not not getting richer.
Dopamine Fracking [comments]
201 points · 62 comments · igerman.cc · 5h ago
The article coins the term "dopamine fracking" to describe extracting the most concentrated hit of pleasure from an activity — like hobbies, relationships, or online culture — while destroying all the nuance and complexity that made it meaningful. Many in the thread found the term itself spot-on, and the strawberry analogy (comparing a real, complex strawberry to its single synthetic flavor compound) got particular praise for nailing what’s lost. But plenty of people pushed back hard: some called the author a luddite, pointing out that humans have always chased dopamine through art and entertainment, and that Marvel movies or modern franchises aren’t fundamentally different from 80s action comedies. Others noted a blind spot — the author rails against dopamine fracking while apparently spending lots of time on Discord and linking to YouTube videos, and a few commenters argued the real driver isn't tech but capitalism and alienation, referencing Marx and a book called *Attensity* that apparently coined a very similar "human fracking" concept. A handful of voice flat-out said the title gives you everything useful and the rest is just another "internet bad for your brain" rant.
APC–2 – A professional record cutter for producing original playback discs [comments]
194 points · 96 comments · teenage.engineering · 6h ago
Teenage Engineering announced the APC-2, a professional-grade vinyl record lathe for cutting custom playback discs in real time. The HN thread immediately zeroed in on the missing price tag and questioned who would actually buy one — the consensus split between those dismissing it as a hype-driven luxury object for wealthy aesthetes and those defending the appeal of cutting your own dub plates or ultra-small-run records for live shows. A long tangent erupted over whether Teenage Engineering is just a design-forward marketing machine (citing their $200-plus desk and a rebadged Gakken toy), with some domain experts arguing the company prioritizes style over actual utility, while others countered that niche passion projects don’t need to be economical. Several people with lathe experience weighed in on the real costs: the skill, maintenance, and consumables make consistent stereo cutting brutally hard, and one noted that a comparable European setup runs around €10,000 plus a multi-day training course in a Swiss forest. The thread also wandered into a heated comparison of box cutters, used as an analogy for how form-over-function products like this APC-2 fit into a broader market of expensive, “tasteful” tools for people who don’t actually do the work.
Vitamin D3 During Pregnancy and Cognitive Performance at 10 Years [comments]
184 points · 82 comments · jamanetwork.com · 16h ago
A study in JAMA Network Open examined whether giving pregnant women high-dose vitamin D3 (versus standard-dose) improved their children’s cognitive performance at age 10. The HN thread took that narrow clinical question and ran far afield, with most of the energy going into a sprawling debate about whether humanity’s modern indoor lifestyle—car travel, air conditioning, sunscreen, office work—is causing widespread vitamin D deficiency and downstream cognitive effects. Several people pushed back hard against any attempt to link cognitive differences to skin pigmentation or latitude, calling that a fraught and oversimplified framing, while others pointed out that the study was done only in Denmark, a homogenous, low-sun population, making it unclear how the results apply to darker-skinned individuals or people in sunnier climates. A practical takeaway that got traction: the modern knowledge-worker life is essentially an unplanned experiment in sun deprivation, and the afternoon brain fog most people accept as normal might just be biology reacting to a total lack of natural light and fresh air.
The Smallest Brain You Can Build: A Perceptron in Python [comments]
174 points · 22 comments · ranpara.net · 7h ago
The article is a tutorial on building a perceptron from scratch in Python, billed as the smallest possible brain you can construct. The Hacker News crowd largely sidestepped the actual code, treating the post as a launching pad for recommending heavyweight textbooks like Chris Bishop’s new Deep Learning book and the Fast.ai course, with one person calling ad-hoc demos insufficient for real learning. A few commenters pushed back on the “smallest brain” framing, pointing out that an IF statement or even a transistor is arguably simpler, and one person joked that a nihilistic function returning zero is smaller. Another thread veered into the philosophy of knowledge—how you can “know” something for years without truly grasping its implications—while one person shared their own JavaScript version (NanoNeuron) and another asked if the perceptron could run Doom.
New drug 'functionally cures' many hepatitis B virus infections [comments]
152 points · 22 comments · www.science.org · 6h ago
The linked article wasn't available to this summarizer; from the discussion, it covers a new drug called bepirovirsen that "functionally cures" a subset of hepatitis B infections by clearing the virus from the blood and making it undetectable. The main pushback came from a domain-aware comment pointing out that the 19% cure rate is real but only applies to a carefully selected trial population—non-cirrhotic patients already on stable therapy—leaving open the question of whether it reduces the 1.1 million annual HBV deaths, which are driven almost entirely by cirrhosis and liver cancer in patients not enrolled here. A side debate broke out about whether treated patients remain contagious, with one side arguing that undetectable viral DNA means no transmission (citing the HIV U=U standard) and others cautioning that dormant viral DNA still encoded in the liver means the risk isn't zero, unlike with HIV. Elsewhere, the thread shifted to a heated comparison of research priorities: the sheer scale of HBV—300 million people worldwide, causing an extremely aggressive liver cancer—was contrasted with HSV, which one commenter argued is vastly overrated as a disease since most carriers are asymptomatic and it's easily managed, though others pushed back on dismissing the millions who suffer painful, visible outbreaks that crush their quality of life.
The OnlyFans Economy of American AI [comments]
141 points · 196 comments · leoveanu.com · 17h ago
The linked article wasn't available to this summarizer; from the discussion, it's a critical take on the US AI industry's economics and quality. The thread immediately splits into a historical analogy about every creative medium having a "slop phase" — early print was full of smut and misinformation like the *Malleus Maleficarum* — but a vocal faction argues that we're past that and the real problem is a government-protected cartel that can't touch cheaper, better Chinese models like DeepSeek v4 Flash, which many engineers say is "good enough" for 90% of daily coding. A whole other branch of the conversation derails into a bizarre, nearly impenetrable rant about AI idolatry and Tolkien fanfic that leaves even daily AI users baffled, with some calling it a sign that heavy AI use erodes coherent thought. Meanwhile, there's a sharp consensus that the big US labs are now in a "too big to fail" phase, angling to use Treasury bonds and 401(k)s as exit liquidity, and the real split is between those who think frontier models like Opus are mandatory and those who see them as overpriced overkill for tasks a five-dollar subscription can handle.
The gamers taking on the industry to stop it switching off games [comments]
136 points · 137 comments · www.bbc.com · 15h ago
The linked article wasn't available to this summarizer; from the discussion, it's about the "Stop Killing Games" movement pushing back against publishers who shut down online servers, making purchased titles unplayable. The HN thread immediately pivoted to a fierce debate about whether the entire PC gaming industry is fatally pivoting toward cloud streaming and hardware-as-a-service, with one side arguing that Nvidia and others are abandoning consumer gaming for datacenter AI profits, leaving us with a rented future. The other side pushed back hard, pointing to Google Stadia's spectacular failure as proof that input lag and user preference for local hardware aren't going anywhere, plus noting that Nintendo and the indie scene aren't going to vanish into a datacenter. The conversation split further over the economics of it—some argued that if you bought a game and played it for a year, you got your money's worth, while others countered that this ignores the cultural and community value of games as permanent virtual spaces, and that the law (UK's six-year consumer guarantee) should simply force publishers to keep the lights on or release the server code.
Office-open-xml-viewer: Office XML document viewer that renders to HTML Canvas [comments]
130 points · 52 comments · ooxml.silurus.dev · 14h ago
A Rust-and-TypeScript library that renders DOCX, XLSX, and PPTX files to an HTML Canvas, headlined by the fact that every line of application code was generated by Claude through iterative prompting — no human wrote any of it. The HN crowd immediately split: many dismissed the project as "vibe coding" that falls apart on real-world documents, with several people testing it against their own complex files and reporting total failures, while others pointed out that pixel-perfect Office rendering is a famously unsolved problem even Microsoft itself hasn't fully cracked across versions. A few commenters defended the approach as genuinely impressive for what it is — the demo loads instantly, the WebAssembly bundles are surprisingly small, and the XLSX preview was described as "mind-blowing" — but the consensus among people who work with legal or enterprise documents daily is that the "last 10%" fidelity gap (tracked changes, table layout inheritance, nested content controls) makes this unusable for serious work. The thread also veered into a meta brawl about whether disclosing AI authorship hurts or helps the project, with some arguing the prompter couldn't even be bothered to verify their own code works.
1k Data Breaches Later, the Disclosure Lag Is Worse [comments]
111 points · 40 comments · www.troyhunt.com · 4h ago
Troy Hunt marks his 1,000th breach loaded into Have I Been Pwned by arguing that disclosure delays are actually getting worse, not better, despite GDPR and CCPA. The thread largely agrees with his diagnosis but digs into the root causes more bluntly than Hunt does. A few people push back on the idea that HIBP is still needed, pointing out that it doesn't show you the actual leaked data, which makes it hard to know if your leaked password was old or reused, or if your address was current. The bigger split in the comments is about incentives: one long, detailed take argues that the only companies with real motivation to disclose quickly are founder-led startups where a breach kills deals, while public companies and B2C firms have zero financial reason to care—stock prices barely move, and GDPR enforcement is so inconsistent it's almost a joke. A tangent pops up blaming "vibe coders" flooding app stores for slowing down hotfix reviews, but nobody bites on that connection.
If LLMs Have Human-Like Attributes, Then So Does Age of Empires II [comments]
110 points · 109 comments · arxiv.org · 13h ago
A new paper argues that attributing human-like qualities like morality or understanding to large language models is logically flawed, and demonstrates this by training a neural network inside *Age of Empires II*. The Hacker News thread largely grappled with the paper’s central philosophical point—that any sufficiently complex substrate, be it a video game or a pile of LEGOs, could theoretically produce the same behaviors, making LLM anthropomorphism a property of the observer rather than the machine. Several commenters pushed back, arguing the paper conflates the ability to *simulate* something with actually *being* it, and that Turing completeness is a red herring—running Doom on a potato doesn’t make the potato a game console. Others found the argument obvious but welcome, noting it cuts through a lot of hand-wavy Substack essays that treat functional equivalence as proof of identity, though a few felt the paper was just rehashing old philosophy-of-mind debates (like the Chinese Room or the China brain) without offering a new way forward. The conversation ultimately split between those who saw this as a necessary methodological correction and those who thought it missed the real question—whether the underlying *process* in a transformer is different in kind from the one in a game engine.
A Matter Wi-Fi Light Bulb in Rust on the Raspberry Pi Pico 2 W [comments]
104 points · 12 comments · github.com · 7h ago
The linked repository provides Rust Embassy examples for the Raspberry Pi Pico 2, including a Matter-compatible Wi-Fi light bulb implementation that uses BLE for initial setup and Wi-Fi for smart home control. The project's author ran into a tricky hardware bug where the RP2350's faster 150MHz core clock was corrupting the SPI bus to the wireless chip during simultaneous BLE and Wi-Fi operation—fixing it required slowing down the SPI clock divider, which sparked a mini-thread on how to diagnose such issues (oscilloscope decoding SPI is the answer). Several people jumped in to argue that the Pico line is underutilized for projects that normally get thrown at full Raspberry Pi boards, with one side championing the simplicity of flashing a binary over maintaining a whole Linux distro for embedded work. The conversation also noted that the GitHub language stats (75% linker script) misleadingly make the project look like more assembly than actual Rust, since most of the heavy lifting is in pre-built libraries like rs-matter.
Firefox Merges Support for Vulkan Video Decoding [comments]
103 points · 17 comments · www.phoronix.com · 9h ago
Firefox is getting native Vulkan Video decoding, meaning it can finally offload video playback to the GPU via a modern, cross-platform API instead of relying on the VA-API stack. The Linux crowd is genuinely relieved, especially NVIDIA users who had to jump through hoops with compatibility layers like nvidia-vaapi-driver—Vulkan Video cuts that out entirely. A lot of the discussion zoomed in on edge-case hardware like Polaris-era Radeons and the Raspberry Pi 5, where Vulkan support is solid but the hardware literally lacks decoders for VP9 or AV1, so people had to be corrected: Vulkan Video doesn't conjure codec support out of thin air, it just exposes what's already there. There's also a split between those who think this is mostly a win for weird embedded devices and people who just couldn't get VA-API working on their netbooks, versus the Chromium crowd pointing out that Chrome already did this.
1worldflag: A blue dot on a transparent background [comments]
97 points · 67 comments · 1worldflag.com · 6h ago
The submission is a project called 1worldflag — a blue dot on a transparent background meant to symbolize unity without replacing national flags. HN immediately split into two camps: people who loved the transparency gimmick as a fresh flag design move, and people who hated it because it ruins contrast and makes the flag visually weak in practice. A huge chunk of the thread devolved into nitpicking the exact shade of blue (it's "Turbo Pascal blue," not the pale blue of the famous photo), with some pointing out the transparent fabric is synthetic and raises microplastic concerns. The comments also went meta — one person called it the "Esperanto of flags" (with the obvious punchline that nobody uses Esperanto), and a long back-and-forth argued whether nation-states should be dismantled at all, with Europeans pushing for it and others pointing out that a world government would just be a bigger bully.
Texas grid flags risks as data centers, crypto sites fail voltage tests [comments]
89 points · 66 comments · www.reuters.com · 5h ago
The linked article wasn't available to this summarizer; from the discussion, Reuters reported that Texas’s grid operator flagged risks after data centers and crypto mining operations failed voltage tests, failing to ramp down load gradually. The thread quickly moved past the specific failure to argue that the real fix isn't better batteries for data centers—those are sized for seconds of backup, not grid stabilization—but rather forcing these facilities off the public grid entirely or onto a power budget. A strong split emerged: one side wants a ban on crypto and AI datacenters drawing from the grid, while the other warns that would just push operators to run unregulated, polluting private gas turbines, citing SpaceX’s portable generators in Memphis as a current example. Someone proposed a $1/W connection fee to make new megawatt-scale builds pay their true infrastructure cost, though the math on how that scales to residential customers drew pushback. A key engineering detail cut through the politics: large industrial users already negotiate grid connections with capacitor banks and resistive load equipment, but the problem here is instant disconnects of massive loads, which no amount of behind-the-meter hardware solves if the facility falls off the wrong side of the disconnect switch.
7.8 magnitude earthquake shakes part of southern Philippines. Tsunami possible [comments]
89 points · 24 comments · www.yahoo.com · 6h ago
An offshore magnitude 7.8 earthquake struck the southern Philippines near Mindanao, killing at least 19 people, injuring over 200, and sending 1-meter tsunami waves into nearby coasts. The discussion focused heavily on the initial magnitude confusion, with people noting that early estimates bounced between 8.0 and 9.0 before settling at 7.8—a common pattern where scientists revise figures while tabloids grab the highest number. Multiple people pointed out that the local agency Phivolcs is the most reliable source but its website keeps getting overloaded ("hugged to death") after major quakes, which several called unacceptable for a life-and-death emergency. A survivor described being on a high floor in a building 600 km away that swayed for five minutes, sparking a debate about whether to evacuate or stay inside during an earthquake, with some arguing that building codes and personal risk calculus matter more than blanket advice.
Powering up a module from the IBM 604: an electronic calculator from 1948 [comments]
85 points · 25 comments · www.righto.com · 14h ago
Ken Shirriff powered up a single thyratron tube module from IBM’s 1944-era 604 calculator, a pluggable gas-filled tube that conducts until you kill the power, and used it to switch a light bulb. The HN crowd immediately went deeper, with the author himself hanging around for vacuum-tube questions and fielding tangents about Philbrick op-amps and the oddball BRLESC computer that put tubes on printed-circuit boards. A long comment from someone who clearly lived this era laid out the full IBM 600-line history, arguing these weren’t really computers—they were hardwired ALUs with plugboard “programming” and no real memory—and that the commercial side of computing (the stuff that actually shipped and made money) gets overlooked compared to the ENIAC-and-Turing stories. Others pushed back on calling plugboards ROM, calling them more like soft-wired CPU logic, while a separate thread marveled at how few tubes these machines used because the Boolean logic was done with germanium diodes. The top takeaway: Shirriff’s blog convinced at least one person not to toss their old IBM modules, which they’ve been meaning to dump for years.
Why isn't the U.S. better at soccer? [comments]
82 points · 235 comments · www.natesilver.net · 12h ago
Nate Silver’s piece runs the numbers on why the U.S. men’s team has never become a soccer superpower, pointing to the rise of American football, baseball, and basketball siphoning athletic talent, plus an isolationist immigration policy that cut off earlier waves of soccer culture. The thread largely ignores Silver’s specific PELE ratings and historical data, diving instead into a familiar but well-argued structural claim: a country can only dominate so many sports, and the U.S. simply isn’t the default path for its most athletic kids. Several people push back hard on the notion that soccer is “competing for the same athletes” as basketball or American football, arguing that most NBA and NFL bodies are wrong for soccer, but the counter is that the funnel matters—kids who’d be elite soccer players instead commit to basketball or football early, hit a height or size wall, and age out before ever trying the right sport. A recurring, more specific complaint is that even where youth soccer is popular (suburbs, Latino communities), the U.S. lacks the deep, single-sport academy system and street-level pickup culture that countries like Brazil, Spain, or Uruguay take for granted, producing players who’ve drilled drills but never learned to play under real pressure.
Flock license plate reader wrongly linked a San Diego man to a violent crime [comments]
82 points · 39 comments · timesofsandiego.com · 11h ago
A Flock license plate reader falsely tied a San Diego man to an attempted carjacking, landing him in jail for nearly a month despite him being five miles away when the crime occurred. The Hacker News thread immediately split into a familiar fight: one camp insists this is a police screw-up, not a tech failure—the cops ignored the Flock data that actually proved the car was elsewhere—while the other side argues the whole system is the problem, pointing to Flock’s CEO reportedly saying a false positive is better than a false negative and that mass surveillance inevitably enables lazy, biased arrests. A few commenters drove the point home by noting that without Flock, this guy never would have been stopped at all, while others dragged in the British Post Office scandal as an analogy for automated injustice backed by institutional denial. The deeper takeaway from the thread is that Flock sells itself as a neutral tool, but in practice it gives cops a shield to stop first and check facts never.
My automated doubt development process [comments]
80 points · 22 comments · www.alexself.dev · 13h ago
The article describes a structured process for using AI coding agents that emphasizes automated "doubt" through multiple specialized sub-agents that audit specs, code, and security from different angles before shipping. The Hacker News thread largely split between people who had independently converged on similar multi-agent audit workflows and those who remain skeptical about the whole premise. Several commenters pushed back on the assumption that the initial spec itself is trustworthy, arguing that the hard part isn't implementation validation but questioning whether the feature should exist at all — with one person pointing out that "agent talked me out of building it" doesn't demo well, so nobody writes that blog post. A more fundamental criticism came from someone who noted that agents are trained on mediocre software and therefore their architectural feedback often amounts to obvious things the engineer already decided to ignore, while another thread deviated into comparing the approach to nuclear safety culture's "questioning attitude" and permit-to-work systems that became absurdly bureaucratic before being reformed. The most concrete objection was that this entire pipeline encourages overengineering, though a counterpoint emerged suggesting you can add a "skeptical pass" agent that downgrades its own findings.
The architecture of the internet creates risks for democracy [comments]
79 points · 119 comments · www.science.org · 11h ago
The linked article wasn't available to this summarizer; from the discussion, it argues that the architecture of the internet—specifically social media feeds and algorithmic engagement—creates risks for democracy. The thread quickly tore into that framing, with many insisting the real root isn't the internet's structure but unregulated greed, advertising-driven attention capture, or capitalism's demand for endless growth. A sharp split emerged over whether the problem is unique to social media or just the latest iteration of mass media (from the printing press to TV), with one side arguing algorithmic feeds let people build alternate realities at scale while the other side counters that before mass media, local gatekeepers like priests had skin in the game. Several people pushed back hard on the phrase "risk to democracy," accusing the article of using it to mean "support for parties I don't like," which sparked a heated back-and-forth about whether content moderation is censorship or whether the real threat is far-right viewpoints being amplified through coordinated messaging and platform control. A minor but persistent tangent argued that the actual internet architecture—HTTP, brittle protocols, and centralization—matters more than any social media feed, with one dismissive comment summing up the frustration: they wanted a piece about BGP and TCP/IP, not another "algorithm feeds are bad" take.
Algorithmic Monocultures in Hiring [comments]
76 points · 31 comments · algorithmichiring.github.io · 6h ago
A large-scale study of 3.4 million applicants across 156 employers found that over 90% of U.S. companies use hiring algorithms, and because most buy from the same few vendors, it creates an algorithmic monoculture that systematically rejects applicants across multiple jobs and produces racial disparities, with 25% of Black applicants and 14% of Asian applicants directed to positions that show adverse impact under Title VII. The HN thread immediately drew a parallel to the RealPage price-fixing scandal, arguing this should be illegal—if your CV gets a bad score from one vendor, that score is cached for months, so you get auto-rejected by every company using that system, making it nearly impossible to get a human review. Some pushed back that hiring has always been a mess of fads and bullshit rubrics, and that measuring “disparate impact” on applications rather than people is flawed because self-selection bias (one group applying to jobs they’re unqualified for) could skew results. Others countered that the study controlled for this by comparing against non-AI hiring data, where the systemic rejection effect disappeared, proving the monoculture itself—not applicant behavior—is the bottleneck. A couple of frustrated job seekers chimed in that they ace HR interviews but then get killed by personality tests or “intelligence” games before ever speaking to a technical person, and that rewriting a résumé to game the AI feels futile when the whole system is rigged.
Anthropic/OpenAI may be spending more than $1000 for every $100 you pay them [comments]
64 points · 72 comments · ea.rna.nl · 19h ago
The article argues that Anthropic and OpenAI are losing enormous amounts of money on subscriptions, spending over $1,000 in compute costs for every $100 in subscription revenue, because serious coding tasks consume vastly more tokens than simple chat queries. The HN discussion largely accepted the core economics but pushed back hard on the methodology, pointing out that the analysis assumed no prompt caching whatsoever — a deeply unrealistic assumption given that Claude Code explicitly uses caching, which one experienced user said makes their team's actual per-task costs align closely with what they pay for a monthly license. Several people noted that even if API pricing is profitable for Anthropic, subscriptions are clearly subsidized, and the real question is whether this is a temporary land-grab strategy or an unsustainable business model. A recurring counterpoint was that local inference on dedicated silicon with baked-in weights could eventually break the subscription model entirely, though others doubted AI companies would ever give up recurring revenue. One user with direct experience building a 40,000-line codebase with Claude Code reported that his $200/month subscription would cost over $10,000 at API rates, calling it the best tool he's ever used despite agreeing the headline math is directionally correct — and that he'd walk away the moment pricing catches up to actual costs.
Generated 2026-06-08 08:36 UTC
Generated by Sauron from Hacker News discussions and linked articles.