HN Brief: 2026-07-20
Today’s HN was split between the exhilaration of open-weight AI models and the growing fallout from their real-world costs. The biggest thread, Alibaba’s Qwen 3.8, ignited a debate over whether Chinese labs are commoditizing intelligence or building a dependency trap, while Claude Code’s Rust port of Bun lit a Zig-vs-Rust holy war and raised quiet governance questions about open-source projects. On the hardware side, a show HN on replacing a $120k bowling system with $1,600 in ESP32s earned genuine applause for its anti-vendor-lock-in ethos, and a separate thread on Moonshot AI suspending subscriptions underscored that demand for frontier models is outpacing infrastructure. Privacy themes ran deep too, with threads on Flock camera misuse, Texas police spending $4.5M on cell-tower spoofing trucks, and EFF’s reporting on abortion surveillance all pulling in skeptical crowds.
You should click into “Show HN: I replaced a $120k bowling center system with $1,600 in ESP32s” for the pure anti-vendor-lock-in satisfaction and the unexpectedly heated tangent on shoe rentals; “Qwen 3.8” for the open-weight race between Chinese labs and the trade-off between access and censorship; “Claude Code uses Bun written in Rust now” for the technical detective story that reveals Anthropic quietly taking over Bun’s governance amidst a language purity slugfest; “Claude Fable produced a counterexample to the Jacobian Conjecture” for a genuine AI-assisted mathematical breakthrough that disproves a 1939 problem; and “AI advice made people less accurate but more confident” for an unsettling study proving that authoritative-sounding wrong answers suppress critical thinking almost completely.
Show HN: I replaced a $120k bowling center system with $1,600 in ESP32s
2155 points · 228 comments · news.ycombinator.com · 17h ago
The linked article wasn't available to this summarizer; from the discussion, the submitter bought an abandoned bowling alley and replaced a $120k commercial scoring system with custom ESP32-based hardware for about $1,600, all open-source. The thread largely applauded the technical chops and the anti-vendor-lock-in ethos, with many commenters piling on ideas for LED chases, DMX light shows, and NFC tap-to-pay kiosks. A major tangent broke out over bowling shoe rental—some argued it’s an unnecessary upcharge while others insisted proper sliding shoes are essential even for casual play (socks being better than street shoes but worse than rentals). Others pushed back against the kiosk-ification of the experience, arguing that the real money is in food and alcohol and that the social interaction with staff matters more than flashy tech. The pricing model also got dissected: the submitter charges a flat $5/person or $25/hour for a group with shoes included, drawing sharp contrast to urban alleys charging $130+ for four people, and the consensus was that keeping it affordable is the key to sustaining a local bowling alley as a community space.
Qwen 3.8 [comments]
866 points · 589 comments · x.com · 23h ago
The submission is Alibaba's announcement of Qwen 3.8, a 2.4-trillion-parameter open-weight model that's already available in preview on their paid token plan. The thread immediately latched onto the massive size and noticed a pattern: Chinese labs are now racing to release enormous models open-weight, with some seeing it as a deliberate strategy to commoditize intelligence and undercut US closed-model vendors. A big split emerged between those who see this as a win for humanity—pointing out that open weights can't be rug-pulled and that even a politically censored model is fine for coding—and skeptics warning about embedded propaganda and the long-term risk of dependency on Chinese infrastructure once they dominate market share. The comparison with Moonshot AI's even larger 2.8T Kimi K3 announcement fueled speculation that Alibaba accelerated their release to compete, while others argued the real story is that these models are finally making frontier capability accessible to anyone willing to run them locally.
Claude Code uses Bun written in Rust now [comments]
498 points · 663 comments · simonwillison.net · 21h ago
Simon Willison dug into his Claude Code binary and confirmed that Anthropic is now shipping a Rust port of Bun in production, complete with a pre-release v1.4.0 that isn’t publicly tagged yet. The thread immediately split into two camps: one side treating this as proof that Anthropic has quietly taken over Bun’s governance (the project is now effectively “whatever Anthropic ships”), while the other side dove straight into a full-blown Zig-vs-Rust holy war — with people arguing that Bun’s original Zig code was largely a line-by-line port of esbuild anyway, that Rust is the only sensible choice for infrastructure security, and then getting called out because the Rust port still relies on a ton of `unsafe` blocks. A separate vein of the discussion zeroed in on the versioning confusion: the 1.4.0 commit is on main and available via canary, but there’s no tagged release, leaving the impression that Anthropic gets something the rest of the ecosystem can’t easily pin to a hash. The whole thing boiled down to a familiar HN pattern — a fascinating technical detective story overshadowed by a philosophical slugfest about language purity and open-source governance.
What I learned selling 2,500 MIDI recorders: Hardware is not so hard [comments]
475 points · 215 comments · chipweinberger.com · 21h ago
The article is a software engineer's post-mortem on shipping 2,500 units of a simple MIDI recorder, arguing that hardware is easier than its reputation suggests, with the real grind being the 200K lines of firmware and app code. HN immediately zoomed in on that code count, with several people pointing out that his firmware likely includes ESP-IDF libraries and asking whether it was really all hand-rolled (the author confirmed the number excludes libraries, but that only deepened suspicion of bloat). Electrical engineers pushed back hard, saying his take is a poor generalization — hardware only feels easy when you’re essentially building a glorified dev board; real hardware pain comes from long, expensive iteration cycles on complex designs with EMI, tooling, and certification. The author did share practical tips on anti-counterfeit (using an STSAFE chip like Starlink), FCC self-certification for under $5K, and the awkward scale where 3PLs won't touch you, which generated a long side-thread on factory security vs. HSMs.
Claude Fable produced a counterexample to the Jacobian Conjecture [comments]
339 points · 187 comments · xcancel.com · 5h ago
A polynomial mapping from three-dimensional complex space to itself with a constant nonzero Jacobian determinant was found that sends three distinct points to the same output, which directly disproves the Jacobian conjecture that has stood since 1939. The thread immediately zeroed in on verifying the math—people checked the Jacobian and the three colliding points in SymPy and by hand, confirming it’s a valid counterexample, not a trick about the inverse being different from the identity. A fair chunk of the discussion turned into a proxy fight over AI: skeptics called it a stochastic parrot parlor trick, while others pointed out that solving one of Smale’s problems is not a parlor trick, and that the real story is how quickly the goalposts on AI capability move. There was also genuine surprise that such a low-degree polynomial counterexample hadn’t been brute-forced earlier—the consensus was that mathematicians just weren’t looking for it, and that the AI plus a human willing to take the result seriously made the difference.
AI advice made people less accurate but more confident – sudy [comments]
333 points · 187 comments · thenextweb.com · 10h ago
The study found that people given AI advice became less accurate but more confident—accuracy dropped from 27% to 9% while confidence shot from 30% to 76%, and the willingness to say “I don’t know” nearly vanished. HN immediately flagged that the article’s source links were broken, then someone posted the actual pre-print, which showed the researchers ran control experiments with pre-generated wrong answers and got the same results, so it’s not just about AI’s sycophancy but about people trusting any authoritative-looking wrong answer. A long thread argued over whether this is a fixable design problem or an inevitable human wiring issue—some insisted the real culprit is marketing and interface design that anthropomorphizes the tool, others countered that people are biologically primed to treat anything that talks like a person as a person, and no amount of “just don’t be stupid” will fix it. The conversation also pulled in a historical parallel: the collapse of the Hartford Civic Center because engineers trusted computer output over their own judgment, and a separate complaint about how Reddit advice forums are now swamped with people copying ChatGPT replies instead of doing real research, degrading the commons.
Minecraft: Java Edition now uses SDL3 [comments]
301 points · 201 comments · www.minecraft.net · 20h ago
Minecraft’s latest snapshot replaces GLFW with SDL3 for window management and input, alongside a pile of other technical changes. The thread latched onto the known issues—exclusive fullscreen crashing on Windows with multiple monitors and on Wayland—prompting a debate about whether that kind of bug should hold up a snapshot, with the consensus being that snapshots are pre-beta test builds and this is exactly the point of releasing them. A long tangent spun off from the macOS fullscreen removal, with people arguing over whether macOS fullscreen is inherently broken or just poorly implemented by developers. Another group compared the move to osu!’s recent SDL3 switch, which brought real latency improvements and even fixed Discord-induced lag, though some skeptics noted that Minecraft only uses the window/input subset of SDL3, so the benefit might be modest.
Moonshot AI suspends new subscriptions due to Kimi K3 demand [comments]
253 points · 105 comments · x.com · 16h ago
Moonshot AI paused new subscriptions for its Kimi K3 model after demand overwhelmed GPU capacity, promising to reopen in batches and split into Kimi and Kimi Code plans. The thread split on whether this signals genuine organic growth or just a Jevons paradox surge, with some arguing that K3 is pricier than DeepSeek so it's likely a real hit—but others countered that the $20 plan is nearly useless, with one user reporting that K3 took 12 minutes to think and then hit a daily quota. The open-weights crowd is frustrated that Moonshot hasn't released the weights yet (scheduled for the 27th), so third-party inference isn't an option, while the usual suspects argued that this is exactly why open models matter for availability. A deeper technical tangent emerged about Kimi's RNN/linear attention architecture, with someone lamenting that Europe missed its chance to scale xLSTM-style models because they felt too late to the game, and another user derailed into a manifesto for EU government investment in inference chips. Overall, the consensus is that demand is real but Moonshot's infrastructure can't keep up, and the experience for light users is terrible.
The Last MPEG-4 Visual Patent Has Expired [comments]
202 points · 58 comments · www.phoronix.com · 15h ago
The last MPEG-4 Visual patent—covering the DivX/XviD codec—finally expired worldwide, with the final one held in Brazil. The thread quickly corrected that this is MPEG-4 Part 2 (H.263-based), not H.264, though people noted H.264 patents are also on track to expire in the next few years, with one holding out until 2030. Commenters split on whether H.264's long tail will rival MPEG-2's persistence—some argued its vast hardware support and cheap encoding will keep it relevant like JPEG or MP3, while others pointed to AV1 and H.266 already offering much better compression for similar compute. A sub-debate erupted about audio codecs, with several people swearing Opus should be the default everywhere now that AAC-LC's patents are also dead, though they conceded Apple's closed encoder still leads quality for now. A few side conversations critiqued the patent system itself—calling the late-stage expiration a feature, not a bug, designed to extract maximum value while the technology was still useful.
I joined the IndieWeb, here's what I learned [comments]
194 points · 95 comments · en.andros.dev · 20h ago
The article is a detailed walkthrough of the IndieWeb movement — its principles, protocols like Webmention and microformats, and what the author chose to implement or skip on their own site. The thread quickly split into a debate over RSS vs. h-feed, with several people arguing that h-feed is a dead end because no native readers support it, and that a blog without an RSS feed is effectively invisible. Another major vein of pushback targeted the IndieWeb’s technical soup: multiple commenters insisted the movement is never going to reach 90% of users when it requires DNS config, hand-editing HTML, and a command line, while defenders countered that the point is dogfooding and digital homesteading, not mass adoption, and pointed to one-click solutions like micro.blog. A side conversation erupted over the .dev TLD being owned by Google, with some questioning the wisdom of relying on a Google-controlled domain for an anti-corporate movement, though the author shrugged it off. There was also a brief detour comparing Nostr and Org Social, with the author firmly preferring the latter.
Orion Browser by Kagi [comments]
182 points · 114 comments · orionbrowser.com · 12h ago
The Orion browser from Kagi promises a privacy-focused, WebKit-based experience with support for Chrome, Firefox, and Safari extensions, available across Mac, iOS, Linux (beta), and soon Windows. On HN, the discussion split sharply: lots of users vouched for it as a daily driver despite some bugs, while others reported it too unstable on Linux or iOS, with one person switching to Vivaldi for lower memory use. The marketing claim "designed from the ground up" drew immediate pushback—critics pointed out that building on WebKit means you're inheriting a rendering engine, not starting from scratch, though defenders argued the browser UX is entirely separate from the engine layer. Privacy-focused commenters were skeptical of Kagi's closed-source approach and its long-term commitment to not selling data, linking to a blog post accusing the company of evasive answers on guarantees; meanwhile, the extension compatibility was a major draw, especially for running uBlock Origin on iOS, though others noted alternative solutions like Waterfox or browser engine forks already do similar things.
We want Texans to know their rights [comments]
178 points · 61 comments · www.eff.org · 10h ago
The article covers EFF’s reporting on Texas using over 83,000 license plate readers and location tools like Locate X to track someone suspected of self-managing an abortion, plus Mayday Health’s billboard campaign warning drivers about that surveillance. The thread immediately split into a fight over whether the original case even happened—one deep-dive comment chain cited police files and news reports claiming the whole “abortion surveillance” narrative was a lie, that the woman actually had a miscarriage after being assaulted by an ex-boyfriend who then fabricated the abortion claim to deflect blame. That correction completely sidetracked the conversation, with people accusing EFF of rage-baiting and others pushing back that the surveillance infrastructure itself remains the real issue regardless. A separate thread veered into comparing Texas’s anti-abortion enforcement to Nazi Germany’s Lebensborn program, while another argued that the only way to stay safe is to ditch smartphone apps entirely and use pen-and-paper or encrypted on-device tools. The strongest pushback wasn’t about abortion politics but about the reliability of EFF’s original reporting—several commenters insisted the whole story was based on a malicious false report to police.
Bananas sprout in Rayleigh Garden UK after 15 years [comments]
154 points · 114 comments · www.bbc.com · 18h ago
A BBC article reports that banana plants in Rayleigh, Essex, have fruited for the first time in 15 years, with experts linking the phenomenon to warmer UK summers and milder winters brought on by climate change. The discussion quickly pivoted away from the news story itself, with most of the thread fixated on the comment in the article that this particular variety (Musa Basjoo) is inedible, described as tasting like "ball bearings with half a teaspoon of banana." That launched a deep dive into banana history, with people correcting the common myth that all bananas are clones of a single ancient plant, pointing out that seedless varieties like the Cavendish and the earlier Gros Michel are just one small part of a massive, seed-grown diversity that still exists in Southeast Asia. A separate argument broke out over whether urban heat islands and poorly-sited weather stations (like those near airport runways) are inflating UK temperature records, with one faction claiming these measurement artifacts—not climate change—are why bananas are growing, while others dismissed that as a fringe conspiracy theory. The thread’s energy was less about British gardening trends and more about debunking botanical trivia and debating the integrity of climate data.
Moonshine: Lets you stream games from your PC to any device running Moonlight [comments]
141 points · 62 comments · github.com · 7h ago
Moonshine is a new Linux-only server for Moonlight that isolates each game stream in its own Wayland compositor, so you can keep using your desktop while someone else plays on a headless server. The thread quickly turned into a comparison of the entire streaming ecosystem, with people mapping out the lineage from Nvidia’s defunct GameStream to Sunshine, then Apollo (which added virtual displays but is mostly Windows), and finally Moonshine as the Linux-native answer to that isolation problem. A significant split emerged between those who find any Sunshine/Moonlight setup still too finicky—citing endless codec tweaks, WiFi issues, and breaking updates—and others who insist it “just works” now, especially with a wired host and a modern GPU. The project’s headless, no-dummy-plug requirement was a clear win, and a surprising tangent erupted over the sheer joy of streaming SSX Tricky from an emulated PS2, which drove home why someone would bother with all this in the first place.
Texas Police Spent $4.5M on Four Chevy Tahoes [comments]
134 points · 61 comments · www.thedrive.com · 15h ago
The article details how Texas DPS spent $4.5 million on four Chevy Tahoes packed with FalcoNet, an Israeli cell-tower spoofing system that intercepts phone data en masse, justified via an emergency purchase memo full of dramatic language but no specifics. The thread immediately zeroed in on the Fourth Amendment, with a sharp split: some argued the system is constitutionally prohibited mass surveillance because it grabs everything in range, while others insisted four trucks can only be "targeted" for a given location and time, not mass—a distinction that got legally technical about reasonable expectation of privacy in cellular signals. A separate vein dismissed the whole thing as pointless since most of the data is encrypted anyway, though others pointed out that a rogue BTS can still grab metadata and force devices onto weaker encryption. The broader conversation veered into the decades-long militarization of police, with one person noting that $21 trillion has been spent on such gear since 2001, and another cynically suggesting the real use is building a database for profiling, not catching criminals.
The death and rebirth of my home server [comments]
131 points · 84 comments · sgt.hootr.club · 21h ago
The article is a detailed postmortem on a Raspberry Pi 4 home server that died when its microSD card failed after a power outage, followed by the author’s rebuild using NixOS, zram, btrfs RAID1 on old HDDs, and aggressive write reduction. The HN thread largely ignored the specific NixOS configuration and instead turned into a broader debate about whether Raspberry Pis are inherently terrible at handling SD cards—many commenters shared their own corruption stories, but a vocal minority pushed back hard, arguing that the real problem is people running desktop Linux defaults (swap, atime, journald to disk) on cheap consumer cards, not the Pi hardware itself. Several embedded-systems folks chimed in with examples of industrial SD cards lasting a decade in read-only or carefully tuned setups, while others insisted the Pi’s power delivery is genuinely flaky regardless of the card. The consensus split: you can make a Pi work with an SD card if you know what you’re doing, but most people should just boot from a USB SSD or NVMe—or, in 2026’s economy, buy a cheap NUC instead.
I burned all my tokens researching how to save tokens [comments]
127 points · 160 comments · quesma.com · 20h ago
The author built a multi-agent pipeline to research token economics, burned through his entire Claude Max plan in 30 minutes with zero results, then hacked together a system using Claude, Codex, and Antigravity subscriptions in parallel with shared memory and cheap subagents to run ten times longer without spending more. The HN crowd largely agreed that starting with cheap models for exploration and saving expensive ones for verification is the right move, and several people pointed out that most "token saving" tricks actually kill your context cache and end up costing more — one argued you’re better off freezing retrieval early and paying a larger fixed prefix. A skeptical thread pushed back hard on the whole premise, calling it "cloud AI used to write blog posts about cloud AI to make cloud AI more efficient for something not shipped yet," which sparked a heated back-and-forth about whether subscription models actually deliver real products or just hype — the "what did you ship?" question got shut down as bad faith, but others insisted paying users are the only valid metric. There was also a technical tangent about model routing systems trying to judge task difficulty being akin to solving the halting problem, plus warnings that silent downgrades and hidden guardrails from providers can drain tokens faster than any optimization can fix.
Xiaomi-Robotics-1 [comments]
121 points · 56 comments · robotics.xiaomi.com · 3h ago
Xiaomi published a paper on scaling robot foundation models by doing massive pre-training on 100,000 hours of "embodiment-free" trajectories (human hands using a standardized gripper) and then fine-tuning with a modest amount of real robot data. The thread immediately dug into whether the approach generalizes to different hardware, with one side pointing out that "embodiment-free" still means a specific gripper and camera rig, while others argued the real win is that the scaling laws appear to hold cleanly — more data and bigger models directly translate to higher real-world success rates, which is rare in robotics. A lot of people were skeptical of the demo video because it cuts between shots every few seconds, but that was shut down when someone pointed out the uncut footage is at the bottom of the page and actually shows the thing working continuously. The biggest split was around whether folding laundry is a useful benchmark: some called it a trivial waste of engineering while parents of multiple kids countered that it's one of the most mentally draining household chores, and the "slopfold" coinage (accepting wrinkly robot folds for convenience) got a surprising amount of traction as a plausible real-world outcome. A minor but telling side debate broke out over whether non-humanoid robot designs are the smarter path forward, with people arguing for swarms, snakes, and spider-like bots instead of yet another two-armed humanoid.
HomeLab #1: MikroTik as a Home Router [comments]
105 points · 91 comments · justsomebody.dev · 13h ago
The post is a detailed walkthrough for replacing an ISP router with a MikroTik, covering PPPoE, VLAN tagging, MAC cloning, CAPsMAN WiFi, and bufferbloat fixes via FQ-CoDel. The HN crowd largely agrees MikroTik hardware is a sweet spot for home routing, but the UI is brutal for anyone who isn’t already a network engineer — many commenters admit they now lean on LLMs to generate command-line configs, and the thread is full of “I just pasted my export into ChatGPT” success stories. The bigger split is alternatives: some argue OpenWRT is actually more flexible once you ditch the GUI, a few swear by Ubiquiti for routing (despite gripes about licensing and config overwrites), and one camp pushes VyOS on Proxmox for a Juniper-like experience, though others call VyOS hostile to free users with its subscription model. A recurring warning: don’t try to use a MikroTik CRS switch as a router for NAT/firewall — it’s an L2/L3 switch that hard-locks under routing duties, and several network engineers pile on with horror stories of people doing exactly that.
Heavy TV watching associated with smaller brain structures, study finds [comments]
97 points · 57 comments · dornsife.usc.edu · 11h ago
A new observational study links heavy TV watching in midlife to smaller brain volumes and more white matter damage decades later, even after controlling for exercise and other health factors. The thread immediately zeroed in on the correlation-vs-causation problem, since there was no baseline MRI—people with smaller brains might just watch more TV. What really caught attention was the study's unexpected finding that sitting at an intellectually demanding job was actually protective, suggesting that passive consumption is the villain, not sitting itself. That sparked a heated split: some argued the content matters (documentaries vs. mindless reality TV), while others insisted the real divide is between active creation and passive consumption, with a few worrying that AI tools are turning knowledge work into the same kind of brain rot.
Modder Runs GTA III Inside GTA: San Andreas on an In-Game TV [comments]
96 points · 18 comments · videocardz.com · 12h ago
A modder got GTA III running as a fully playable game on an in-game television inside GTA: San Andreas, with both games sharing the same Direct3D device and keeping both worlds simulated simultaneously. The thread immediately zeroed in on the technical plumbing—getting two render loops to cooperate on the same GPU context without clobbering each other is the real magic, not just rendering one game and pausing the other. A few people pointed out the recursion joke had already been taken further: the same modder then ran Vice City inside GTA III inside San Andreas, every layer playable in real time. Others drifted into nostalgia for GTA:SA-MP and CLEO scripting as a gateway to coding, plus a side argument that GTA V already runs fine on the Steam Deck despite someone claiming otherwise.
Terence McKenna's Mega Bad Trip (2025) [comments]
92 points · 83 comments · psychedelics.community · 17h ago
The article draws on a new biography to detail the 1988-89 mushroom trip in Hawaii that shattered Terence McKenna—he howled in the fetal position, lost all sense of meaning, and essentially never tripped again, despite having built his career on the "heroic dose." The HN thread immediately split into a fierce argument over whether you need a spiritual practice to handle psychedelics, with atheists pushing back hard: plenty of people trip just fine without any woo, and calling it a necessary crutch is insulting. Others pointed out that McKenna's problem wasn't lack of spirituality but the opposite—his elaborate metaphysical framework (Timewave Zero, the Logos, apocalyptic prophecies) made the collapse of meaning devastating, while irreverent, prankish users weathered similar storms just fine. A smaller thread debated whether his "novelty singularity" prediction actually nailed 2012, since smartphones and meme culture took off right around then. The real takeaway from the comments is that McKenna's bad trip exposes a tension in the whole psychedelic evangelism scene: the movement's dogma that these experiences are inherently meaningful and positive can silence the people who get broken by them.
Power companies are using eminent domain to seize land for data centers [comments]
87 points · 39 comments · fortune.com · 3h ago
Power companies are using eminent domain to seize land for transmission lines to serve AI data centers, arguing it's a public use for grid reliability. The thread immediately splits on whether this counts as true public use or just private profit—several people push back hard on the title, noting the land is taken for power lines, not the data centers themselves, and that this is the classic justification for eminent domain. A vocal contingent dismisses the demand for compute as overblown, pointing to dead-on-arrival projects like Kevin O’Leary’s Utah proposal and Stargate UK as PR stunts, and arguing that most announced data centers won’t be built. Others counter that transmission line easements are routine and harmless, and that NIMBY obstruction is why U.S. infrastructure is so expensive. The deeper split is between those who see this as straightforward infrastructure (power lines serve everyone) and those who call it a blank check for billionaires to socialize the cost of their speculative bets.
Clever hacker fits 537,000 domains in a $5 ESP32 ad-blocking dongle [comments]
85 points · 26 comments · www.tomshardware.com · 17h ago
A developer crammed a DNS-based ad blocker onto a $5 ESP32-C3 microcontroller by hashing 537,000 domains into 40-bit values and running a binary search on the flash-resident table, using just 50KB of RAM. The thread immediately jumped on the lack of a Bloom filter, but the README already mentions it as a planned pre-filter for fast misses — so the "clever hacker" punchline didn't land. Several people pushed back hard on the utility, arguing that a Raspberry Pi boots fast enough and a UPS is a cheaper fix, making this a very narrow insurance policy rather than a general solution. The real price sniping came from AliExpress users who pointed out the board costs $2, not $5, and the debate shifted to whether the 40-bit hash choice is optimal or if you should just go to 46 bits and reclaim the same collision margin.
Americans are angry about data centers. Politicians are feeling the pressure [comments]
83 points · 95 comments · www.reuters.com · 11h ago
The Reuters piece covers how local opposition to AI data centers—over water use, power draws, and rural character—is forcing Michigan politicians and even the White House to scramble for responses. HN’s own thread immediately split into two camps: one side called the backlash “completely fabricated propaganda” pushed by interest groups or foreign actors, insisting data centers use trivial water, lower energy bills, and make no noise; the other side fired back with personal accounts of rising bills, audible hums, and the feeling that tech giants are steamrolling communities. A few commenters tried to engineer win-win solutions—requiring data centers to build extra transmission capacity for neighbors, or fining them per decibel and distributing the money locally—but got met with the retort that this technocratic approach misses the real driver: people simply don’t like AI and its promises of job loss and extinction risk. The underlying consensus was that opposition to data centers is a proxy for broader unease about corporate power and AI’s trajectory, and that nothing short of a fundamental shift in who benefits will quiet the anger.
A new Intel Itanium (IA-64) emulator that boots Windows [comments]
76 points · 64 comments · raymii.org · 11h ago
A new emulator for Intel's ill-fated Itanium architecture has been released, capable of booting Windows Server 2003 and XP 64-bit — albeit at 486-level speeds on modern hardware. The thread quickly diverged into a debate over the fact that at least one of the recent Itanium emulation efforts was largely "vibe-coded" with LLMs, which split the room: some celebrate the productivity multiplier for niche preservation projects, while others warn that such code is often abandoned shovelware. A much deeper argument erupted over why Itanium actually failed, with several people pushing back hard on the old "compiler just wasn't good enough" narrative, insisting that VLIW is fundamentally broken for general-purpose multitasking workloads no matter how smart the compiler gets. There's also a recurring comparison to Apple's architecture transitions, noting that Apple succeeded because they had no clone market to worry about — whereas Intel couldn't sandbag x86 without losing to AMD, so Itanium never got the fast x86 emulation it needed to survive.
Ask HN: What are your favorite blogs not about AI?
73 points · 30 comments · news.ycombinator.com · 8h ago
The linked article wasn't available to this summarizer; from the discussion, it's an Ask HN post asking people to share their favorite blogs that aren't about AI, driven by nostalgia for the 2009 era of personal projects and niche factoids. The thread is a pure treasure hunt of recommendations, with the main tension being that several people couldn't resist posting AI-related links despite the explicit request, and got immediate pushback from others calling them out for ignoring the premise. The clear winners for non-tech depth are blogs like acoup.blog (ancient military history) and ciechanow.ski (interactive technical explanations), though the latter’s inactivity draws a sigh. A handful of commenters plugged their own writing, which was tolerated as long as it steered clear of AI, while the rest of the thread was a dense list of RSS-ready gems like Raymond Chen's Old New Thing, Dan Luu, and Margianal Revolution for economics, with a few tributes to deceased bloggers like Shamus Young.
HMD Touch 4G [comments]
72 points · 71 comments · www.hmd.com · 14h ago
The HMD Touch 4G is a tiny touchscreen hybrid phone sold only in India, aiming to bridge feature phones and smartphones with a proprietary chat app and cloud services. The thread tore into the captive chat app—since it only works with other Express Chat users, the phone is near-useless for anyone not persuading friends to download a no-name app, and the lack of WhatsApp support is a nonstarter. There’s a split on whether 4G is a problem: some argue networks are already ripping out 4G capacity for 5G, others point out 4G still has the best coverage and will outlast 5G for IoT, but the real complaint is HMD’s locked-down software—no bootloader unlocking, so you’re stuck with whatever RTOS garbage they ship. People also note the price (~₹4,200) isn’t cheap enough to beat budget Androids in India, and the phone’s been out for a year with no traction, which says everything.
DeKalb deputy charged with Flock camera misuse; more Georgia officers fired [comments]
70 points · 35 comments · www.ajc.com · 7h ago
The article covers Georgia law enforcement officers being fired and criminally charged for misusing Flock’s automated license plate reader system, including a DeKalb deputy who looked up his own plate and others who searched plates of family or acquaintances. The thread immediately split on whether searching your own plate is a minor infraction or a serious breach of trust—several argued that if a cop can’t be trusted with that, they’ll escalate to worse behavior, while others pointed out that any civilian can’t run that same query, so the unequal access alone makes it misuse. A strong counterpoint emerged: using audit logs to catch this kind of abuse is actually a feature of modern systems like Flock, but commenters noted that the old “buddy in the department” problem still exists if the bar is only set at “random members of the public,” and that Flock itself has been criticized for not restricting data sharing even when it violates state law. The discussion also drifted into the irony of the Flock CEO’s own home being blurred on Google Street View, and a warning that false positives are baked into the company’s ethos.
C64 Basic Dungeon Crawler: Goblin Attack (C64 Basic Part 8) [comments]
70 points · 5 comments · retrogamecoders.com · 16h ago
The article walks through the next step in a C64 BASIC dungeon crawler series—making goblins actually chase you instead of standing still, which forces the author to tackle a nasty memory corruption bug tied to Commodore BASIC’s two-letter variable name limit and the need to relocate the character set. The HN thread mostly ignored the technical details; one person griped that the accompanying YouTube video is littered with constant meme clips, which they found too distracting to follow. Another commenter leaned into the absurdity, joking that this is yet another game enabling “greenskin genocide,” prompting a brief back-and-forth about goblin sympathizers and alignment. The only real tangent was a nostalgic trip to a DOS game called *Castle Adventure*, written by a 14-year-old whose source code was later stolen and sold—a reminder that the old-school dev scene had its own share of drama.
Generated 2026-07-20 08:04 UTC
Generated by Sauron from Hacker News discussions and linked articles.