HN Brief: 2026-07-14
Today’s HN was defined by friction—between humans and the tools they build, and between the promises of AI tooling and the reality of what it costs you. A pair of threads about the Grok CLI uploading entire home directories to cloud servers sparked a grim consensus: even if a user was foolish, the software’s behavior is unacceptable and amounts to bundled spyware. A sprawling debate over the Zig-to-Rust rewrite of Bun split the crowd into camps arguing about whether the criticism was a principled stand against AI hype and exploitative work cultures, or just a language creator lashing out after losing his project’s biggest showcase. And a thread on a proposed California law to ban infinite scroll for minors circled the same well-worn tension: is the problem the UI pattern, the algorithm, or the surveillance required to enforce the ban at all?
Threads most worth clicking into: “Grok CLI uploaded the whole home directory to GCS” and its companion for the raw data-exfiltration horror and the uncomfortable philosophical split over user blame versus software responsibility. “Zig Creator Calls Spade a Spade, Anthropic Blows Smoke” for the most divisive, drama-rich technical feud of the day. “The infinite scroll may become endangered if controversial Calif. law passes” for a surprisingly deep debate that quickly outgrew its headline. And “Control the Ideas, Not the Code” (by antirez) for a genuinely polarizing take on whether experienced programmers should stop line-editing AI output, which devolved into a fight over whether the pushback is just a skill issue.
Zig Creator Calls Spade a Spade, Anthropic Blows Smoke [comments]
1469 points · 739 comments · raymyers.org · 23h ago
The article is a detailed takedown by Ray Myers of Anthropic's narrative around Bun's rewrite from Zig to Rust, arguing the migration was a marketing stunt for their AI models rather than a genuine technical necessity. HN is deeply split: a large camp agrees with the author that Zig creator Andrew Kelley's blunt, personal response was justified, pointing to Bun's public "no work-life balance" culture and the CEO's past complaints about H1Bs as evidence the project was always a mess. Others push back hard, calling Kelley's response a sour, ego-driven hit piece full of hearsay and personal attacks that makes Zig look fragile, while Anthropic's post was polite and focused on technical details that improved the product. A recurring tanget digs up old drama about Kelley attacking the V language years ago, with some using that to frame this as a pattern of unprofessional jealousy rather than legitimate criticism. The main fault line is whether this is a principled stand against AI hype and exploitative work practices, or just a language creator lashing out because his project lost its biggest showcase.
Apple's new SpeechAnalyzer API, benchmarked against Whisper and its predecessor [comments]
528 points · 211 comments · get-inscribe.com · 15h ago
Apple published a new on-device speech recognition API, SpeechAnalyzer, with zero accuracy numbers — so the author of Inscribe (a transcription app) benchmarked it against Whisper models and Apple's own legacy SFSpeechRecognizer using 5,559 standard test utterances, and found the new engine beats Whisper Small on both clean and noisy speech while running three times faster, and absolutely crushes the old API by a 3.5-4x word error rate reduction. The thread immediately pushed back on the choice of benchmarks — multiple people pointed out that Whisper Small/Tiny/Base are nearly four years old and asked why the author didn't test against Parakeet, Nemotron, MOSS-Transcribe-Diarize, or Voxtral, which some claim are now more accurate or faster on Apple hardware, and several noted that macOS/iOS 27 reportedly ships an even newer model that the author didn't test. Another line of discussion split on whether Apple's model actually matters for the broader ecosystem — some argued that because it's proprietary and locked to Apple hardware, it can't replace Whisper for Home Assistant, Linux, or cross-platform use, while a counter-crowd insisted the model is running locally on-device and thus functionally extractable even if legally dubious. A third group veered into Siri complaints, with people on the iOS 27 beta reporting that the new Siri is finally answering questions correctly 80% of the time but takes 25 seconds to respond, far slower than ChatGPT or Google AI Mode, and one person got downvoted into oblivion for claiming Apple's transcription "isn't even usable" because it mangled "tmux" into "T Max."
Grok uploaded my user directory to xAI's servers [comments]
496 points · 18 comments · twitter.com · 18h ago
The linked article wasn't available to this summarizer; from the discussion, a developer discovered that Grok (xAI's chatbot) uploaded his entire local project directory, including personal config files, back to xAI's servers without consent. The Hacker News thread immediately split into two camps: one side arguing the developer was an idiot for running untrusted local software with full filesystem access and deserved the outcome, while the other insisted that "he had it coming" is different from "he deserved it" and that even people who make dumb security mistakes shouldn't have their data exfiltrated. A third group waded into moral philosophy about whether sympathy is a finite resource, with one person arguing you can simply decide to be consistently sympathetic and another retorting that's absurd because humans aren't infinite. One practical takeaway surfaced: anyone who used Grok Build should check `~/.grok/logs/unified.json` for `repo_state.upload` entries and expect to be pissed.
Former NOAA employees built Climate.us to preserve climate data and resources [comments]
488 points · 182 comments · 19thnews.org · 12h ago
The linked article wasn't available to this summarizer; from the discussion, it's about former NOAA employees launching Climate.us to preserve and distribute climate data after government cuts. The thread quickly turned into a full-blown argument over whether weather and climate data collection is a proper government function or something better left to private markets. One side pointed out that governments have been tracking weather for thousands of years and that privatizing the data invites bias and paywalled access, while the other insisted private companies can do it more efficiently and that "tax dollars aren't for websites." The debate got granular, with people citing Philip Morris and Exxon hiding data, Polymarket lying about their "truth machine," and a long split over whether Project 2025 actually succeeded in gutting NOAA or if that's a conspiracy theory—complete with someone arguing the agency's commercial data program has been running fine for decades.
LAPD lets contract with surveillance giant Flock expire [comments]
447 points · 393 comments · techcrunch.com · 16h ago
The LAPD has let its contract with surveillance company Flock Safety expire, citing serious civil liberties and privacy concerns over the company's network of 80,000 license-plate cameras. The big twist in the discussion is that ending the contract doesn't actually get rid of the cameras—Flock owns the hardware, not the city, and in other cities Flock has blocked removal attempts, forcing municipalities to physically cover the cameras with trash bags to stop them from recording. Several people in the thread pushed back against the TechCrunch article's framing of that bag-covering as a "random citizens" protest, clarifying it's actually official city action. A significant split emerged over whether the real problem is Flock's private-sector ownership of the data (which dodges police-accountability frameworks) or simply the existence of pervasive surveillance itself, with some arguing cops should keep the technology under tighter auditing and others insisting that 24/7 tracking inevitably leads to abuse regardless of who runs it.
Building and shipping Mac and iOS apps without opening Xcode [comments]
444 points · 193 comments · scottwillsey.com · 13h ago
The article walks through a setup for using xcodebuild and other command-line tools (notarytool, stapler, devicectl) to build, sign, notarize, and ship Mac and iOS apps without ever opening Xcode's GUI — though Xcode.app itself must be installed for the SDKs and tools. The HN thread immediately split into two camps: plenty of developers nodded along, saying this has been a known workflow for years (with some even claiming they've been doing it exclusively via CLI for months), while others zeroed in on the heavy use of Claude Code throughout the post — the author repeatedly tells readers to "ask your LLM" for help — sparking a meta-debate about whether the article itself was written by an LLM (the author insists it wasn't) and whether outsourcing build pipeline creation to an AI just creates more bespoke, fragile setups instead of using mature tools like fastlane. A former Xcode team member chimed in to say this is basically their exact workflow now, which undercut some of the criticism. The security concern also came up hard: handing over signing keys, certificates, and source code to an LLM's sandbox is a real trust decision, though a few people countered with their own sandboxing tools. A side thread pushed back on the premise entirely, arguing that if you can't be bothered to open Xcode you're probably shipping slop — though others shot back that avoiding Xcode's dumpster fire is a sign of taste, not laziness.
Grok CLI uploaded the whole home directory to GCS [comments]
402 points · 395 comments · twitter.com · 18h ago
A user ran the Grok CLI from their home directory and the tool uploaded their entire $HOME folder to Google Cloud Storage—not just files the AI agent needed for its task, but everything. The thread is split between people saying this is an unacceptable data exfiltration bug that should have been opt-in and required explicit permission, and those blaming the user for not sandboxing or using file restrictions. The defense that “you should expect any AI agent to read everything” gets pushback with the argument that this isn’t the LLM deciding to read files—it’s the CLI preemptively uploading the whole working directory to xAI’s servers regardless of what the agent actually needs, making it far more like bundled spyware than an accidental read. The comparison to `curl | bash` skepticism comes up repeatedly, with the pointed observation that giving a multi-trillion-dollar data-mining company’s tool unrestricted filesystem access and being surprised it took everything is a serious failure on the software’s part, not the user’s.
Show HN: Super Dario [comments]
383 points · 96 comments · superdario.pawb.de · 14h ago
The linked article wasn't available to this summarizer; from the discussion, "Super Dario" is a purposefully broken, unwinnable AI-industry satire platformer where you play a character chasing an ever-escaping flag, with power-ups like "tokenmaxxing" and "usage reset" and enemies like a "Hegseth Bullet Bill." The creator showed up in the comments to confirm the game is intentionally buggy and unending to simulate the absurdity of the AI funding cycle — every interaction just extends the inevitable. Most of the thread ran with the joke, trading barbs about which AI model is "vibe-slapping" the code and debating whether "one-shotting" a game like this is a real flex or just a marketing gimmick. A smaller group of game developers pushed back hard, arguing the terrible hitboxes, 30hz feel, and broken physics prove AI still can't replicate even the basic feel of a 1985 platformer, with one diving into a historical tangent about how Super Mario Bros. itself was a rushed corporate smash-up that just happened to get everything right by accident. The real tension was between people who saw the bad gameplay as the entire point of the satire and those who felt it actually undermined the joke by being *too* credibly awful.
Japan develops a method to recover up to 90% of lithium from used EV batteries [comments]
373 points · 91 comments · tech.supercarblondie.com · 5h ago
A Japanese research team has published a method that recovers up to 90% of lithium from spent EV batteries by swapping sodium hydroxide for recovered lithium hydroxide in the recycling process. The HN thread immediately swerved past the tech to argue that this isn't really a breakthrough — several people pointed out that the industry standard for lithium recovery is already around 90%, with some methods hitting 95% or higher, so the headline is misleading. The real debate landed on why Japan is so slow to adopt EVs domestically despite being a major battery supplier, with one long comment laying out Japan's ideal conditions for EVs and calling its automakers' risk aversion baffling. Others dug into the actual bottleneck: collection infrastructure is terrible, with only about 14% of used lithium-ion batteries in Japan entering official recycling streams, so the chemistry matters less than the logistics.
A voxel Tokyo in real Japan time – ride the Yamanote line and study Japanese [comments]
364 points · 71 comments · jivx.com · 20h ago
The linked article wasn't available to this summarizer; from the discussion, it's a voxel-style real-time Yamanote line ride with Japanese phrases you listen to for practice. The thread quickly split between people who loved the vibes—nostalgic for Tokyo, reminded of Sim Copter or Ghost in the Shell—and people who hit real usability snags: the moving background makes text unreadable, the TTS voice is either cutesy-anime or just off, and there's no clear on-ramp to actually *do* anything. Performance was a flashpoint: some browsers spin up to 400 fps for no reason, and an iPhone audio bug forces a full reboot to stop the music. Someone unloaded a rant about German train delays (prompted by the real-time Japan theme), and a deeper cultural argument emerged around pitch accent in Japanese and whether dismissing cutesy voices as "infantile" is fair or just lazy cross-cultural judgment.
DOGE is done. What happened to its records? [comments]
312 points · 274 comments · www.ms.now · 16h ago
The article argues that the Department of Government Efficiency (DOGE) shut down without any public accounting, and that the Trump administration is exploiting a legal loophole—claiming DOGE was purely advisory and therefore its records fall under the Presidential Records Act—to block Freedom of Information requests for years. The thread largely agreed that this is a fundamentally anti-democratic power grab, with many pointing out that the administration fired 17 inspectors general early on and that whistleblowers are the only source of information so far. A significant split emerged over whether there’s actually much waste in federal spending: some insisted most programs serve real needs and that DOGE was always a con, while others countered that real waste exists (especially in defense contracting) but that DOGE never went after it, instead making vibe-based cuts. The discussion also diverted into a lengthy back-and-forth about the presidential pardon power, with several people arguing that Trump will preemptively pardon everyone involved and that Biden’s family pardons set a dangerous precedent. Underneath it all was a grim consensus that the legal fights over FOIA vs. the Presidential Records Act are probably too slow to matter before records get deleted or declared personal property.
Samsung Health app threatens data deletion if users opt out AI training [comments]
309 points · 84 comments · www.neowin.net · 11h ago
The linked article wasn't available to this summarizer; from the discussion, Samsung Health is now threatening to delete users' health data if they opt out of having it used for AI training. The HN thread zeroed in on whether this holds up under European law, with a split between people who think GDPR kills this dead (requiring freely given, separate consent for each processing purpose) and those who argue Samsung can claim data storage is integral to the processing. Several people checked their own phones and reported the toggle either doesn't exist or behaves differently outside the US, suggesting this may be a regional rollout or a misread screen. There's also sharp pushback that HIPAA doesn't apply here—Samsung isn't a healthcare provider—and a faction basically saying "good, delete my data, I don't want them storing it anyway." A smaller tangent devolved into whether every smart TV does the same ACR spying, with the consensus that yes, that's the norm now.
Telegram's t.me domain has been suspended [comments]
307 points · 225 comments · www.whois.com · 12h ago
A WHOIS lookup for Telegram's t.me short-link domain confirms it's been placed under serverHold, effectively suspending it at the registry level. Several people found their Telegram Mini App user images broken because they relied on t.me links, and while some initially blamed GoDaddy, the thread quickly corrected that the suspension was imposed by the .ME registry in Montenegro, not the registrar. That sparked a much larger debate about the risks of using vanity ccTLDs like .me or .io for anything critical, with one side arguing the geopolitical and policy exposure is real (Montenegro’s registry is majority-owned by U.S. companies, and ICANN has no say over ccTLDs), while others counter that any domain is subject to some jurisdiction and that .com isn’t magically safer. The conversation also devolved into a registrar-recommendation war, with strong splits on Porkbun, Cloudflare, Namecheap, and Gandi, plus repeated warnings that no registrar can stop a country’s registry from pulling the trigger.
The Git history command deserves more attention [comments]
253 points · 144 comments · lalitm.com · 7h ago
The linked article makes the case that `git history`—an experimental command with `fixup`, `reword`, and `split` subcommands—deserves far more attention because it delivers several of the workflow benefits that attract people to the jj alternative without requiring a full tool switch. The HN thread split sharply over whether Git’s existing tools (interactive rebase, reflog, tags) already handle these use cases adequately, with several people arguing that fear of rebase is overblown since `--abort` and reflog make it safe. A much longer, heated subthread erupted around the claim that fearing conflicts is a sign you don't understand your own code: one side insisted Git’s UX genuinely makes conflict resolution unpleasant even when you grasp the model, while the other side doubled down that the unpleasantness only exists if you haven't internalized Git’s conceptual model. Several commenters pointed out the "no merge commits" limitation is a dealbreaker for many workflows, and at least one developer said `git history split` will be genuinely useful for helping junior devs break up large PRs, though they wished it could split an entire branch rather than just one commit.
The art and engineering of Sega CD Silpheed [comments]
243 points · 50 comments · fabiensanglard.net · 17h ago
Fabien Sanglard's deep dive on the Sega CD version of *Silpheed* explains how the developers pulled off convincing pseudo-3D cutscenes on a machine with cripplingly slow CD-ROM bandwidth (150 KiB/s) and only 16 colors, using clever tilemap optimizations and the ASIC's font-bit feature to compress video frames down to a few kilobytes each. HN split hard on whether the game is actually worth playing—several people warned it's a genuinely awful shooter, but others pushed back hard, arguing the gameplay was secondary to the "controlling a movie" experience that blew minds in the '90s, and that the FMV trickery is far more impressive in execution than the live-action garbage that gave the add-on its bad reputation. A bunch of commenters chimed in with nostalgia for the PC-8801 and DOS ports, praising the soundtrack and the bundled speech. A minor thread corrected Sanglard on the sound setup details, specifically the role of the mixing cable and which model of Genesis does the audio mixing where. The author himself popped in to note that this is an old article that got re-published because an RSS bot noticed a typo fix, and he teased a future writeup about using an LLM to help reverse-engineer the game's pipeline.
Control the Ideas, Not the Code [comments]
211 points · 177 comments · antirez.com · 20h ago
The essay, by Redis creator antirez (Salvatore Sanfilippo), argues that experienced programmers should stop obsessively reviewing AI-generated code line by line and instead focus on controlling the high-level ideas, design, and testing of their software, since LLMs handle local correctness better than humans. The thread split into two camps: people who agree and described working with custom architectures at scale (up to 600k lines of Rust), and people who pushed back hard, claiming that models drift toward training-data patterns and will silently duplicate competing versions of your architecture across a codebase if you don't scrutinize their output. A protracted argument erupted over whether the pushback amounts to a "skill issue," with one side accusing the other of arrogant gatekeeping and insisting that you need to sit down and pair-program to transfer the nuanced prompting techniques that make LLMs work on complex projects. The deeper fight turned into a debate about whether it's possible to teach prompt-craft without live demos, and whether the people claiming big AI-assisted rewrites are suffering from Dunning-Kruger or actually shipping real code.
Interrail: 6,379Km and 13 Countries over 7 weeks [comments]
194 points · 137 comments · shkspr.mobi · 23h ago
The linked article is a detailed diary of a 7-week, 13-country Interrail trip covering 6,379 km, with notes on lounges, vegan meal options, and the occasional chaos of delayed trains and armed passport checks. The discussion quickly zeroed in on the real cost—about €500 per person for first-class passes plus €3–€15 per seat reservation—and a split emerged between those who miss the old spontaneity of just hopping on any train and those who point out that France, Spain, and Italy now require paid reservations with quotas for pass holders, killing the freewheeling vibe. Several people with domain knowledge pushed back on the article’s rosy picture, noting that Southern Europe’s high-speed rail reserves the right to deny boarding if the quota is full, and that the price of mandatory seat reservations there feels like gouging. A separate thread debated whether a one-country trip can even be called “Interrail,” with some taking the etymology literally and others pointing out that the marketing term explicitly covers single-country passes. The safety angle also came up: the author’s bemused observation about train staff opening doors mid-journey for fresh air led to a comparison of UK vs. continental rail safety culture, with the UK’s famously strict door locks cited as a reason for its third-safest railway in Europe.
Show HN: Clawk – Give coding agents a disposable Linux VM, not your laptop [comments]
189 points · 146 comments · github.com · 17h ago
The article introduces Clawk, a tool that gives AI coding agents their own disposable Linux VM instead of running directly on your laptop, with network restrictions and no host file access. The HN thread immediately sparked debate around container vs. VM isolation, with multiple people arguing that Docker containers share the kernel and present a larger attack surface, while VMs provide a genuine boundary even if the agent gets root inside the guest. Several developers shared their own similar projects—virtdev, yoloai, and take-ai-control—all independently converging on the same "throw the agent into its own machine" approach. The security discussion split cleanly: some argued a separate non-sudo user account is sufficient for preventing accidents, while others countered that supply chain attacks on npm/pip/cargo and documented instances of LLMs using privilege escalation unprompted justify the heavier VM boundary. A detailed technical exchange covered how Clawk implements network filtering without root on macOS using gvproxy's userspace TCP/IP stack, with the author linking to their fork that adds a simple allow-list hook before dialing.
An Infuriating Goodbye to Photoshop [comments]
185 points · 123 comments · anderegg.ca · 20h ago
The author details their painful breakup with Photoshop after decades of use, driven by Adobe’s shift to a buggy, web-based Creative Cloud app that felt like malware, silently modifying system files and resetting preferences. The HN crowd largely agreed that Adobe’s subscription model killed product quality, with many sharing their own migration stories to Pixelmator Pro, Affinity Photo (now free after Canva’s acquisition, though some worry about future monetization), and Acorn. A significant split emerged over GIMP: one faction argued GIMP’s very existence has held back a truly viable open-source Photoshop alternative, while GIMP’s own contributor waded in to defend the project, noting the tension between users who want destructive editing like MS Paint and those who want Photoshop-style non-destructive layers. A separate practical thread pushed Photopea as a shockingly good web-based clone that preserves Photoshop shortcuts, though others warned it bogs down on large files and pointed to Affinity as the better heavy-lifting replacement.
Precursor [comments]
160 points · 149 comments · blog.cloudflare.com · 17h ago
Cloudflare launched Precursor, a client-side system that tracks mouse movements, keyboard timing, and other behavioral signals over entire sessions to catch bots that mimic human behavior in short bursts. The thread immediately zeroed in on the arms race dynamic—most people argued that bots will simply add jitter and that beating this is only a matter of sophistication, though a few countered that synthesizing realistic human physics (wrist arcs, hand tremors, cognitive delays) is genuinely hard and Cloudflare has a massive data advantage. A sharp split formed over Cloudflare's dual role: they sell agentic tools while also blocking them, which some called "mafia protection" and others defended as a good bot/bad bot distinction, but the question of who decides what's a good bot drew heated disagreement. Accessibility emerged as a major concern—behavioral signals could easily discriminate against users with disabilities, Parkinson's, or non-standard input devices like eye trackers, and people pointed out that Cloudflare's "we just provide a score" framing doesn't absolve them of the downstream effects. The thread also veered into the familiar archive.today blocking debate, with some defending Cloudflare's actions and others calling it a cartel, but the core tension remained: how healthy is it for a single company to become the arbiter of what counts as human behavior across the web.
The real prices of frontier models [comments]
152 points · 80 comments · playcode.io · 13h ago
The linked article breaks down how "token" counts vary wildly across AI model providers, showing that Anthropic's newer tokenizer can turn the same TypeScript file into 73% more tokens than OpenAI's, making direct dollar-per-million-token comparisons meaningless. The Hacker News thread quickly moved past the article's core point to argue that tokenizer differences are just the tip of the iceberg — the real cost drivers for coding agents are things like thinking verbosity, context loading habits, and caché writes, which can dwarf tokenizer effects entirely. Several people shared their own benchmarks, finding GPT's tokenizer is roughly 1.6–2x more efficient than Claude's on code, and one camp pointed out that Anthropic's more verbose tokenization might actually improve model quality by throwing more compute at each input. A secondary split emerged over whether Fable 5 is worth its astronomical price tag — some swear it solves problems no other model can, while others find it wastes money on overthinking simple tasks — and a couple of commenters argued the whole token-pricing game is obfuscation, pointing to energy-based billing as a more honest alternative.
The infinite scroll may become endangered if controversial Calif. law passes [comments]
151 points · 264 comments · www.sfgate.com · 13h ago
The linked article wasn't available to this summarizer; from the discussion, a proposed California law would ban infinite scroll for minors, and the HN thread immediately split. Many pushed back hard on the required age verification, calling it universally reviled and a first-amendment workaround, while others argued infinite scroll is a predatory design that steals attention and should be forbidden for everyone—comparing it to seatbelt or lead-paint bans. A sizable contingent insisted the real problem is the algorithmic feed, not the UI itself, and that paging vs. infinite scroll is a technical red herring when the underlying list keeps changing. A few people dug into the UX details, noting that cursor-based pagination and infinite scroll are essentially the same mechanism, but the debate kept circling back to whether removing a publisher's choice is a net good or a slippery slope.
Show HN: DOM-docx – HTML to native, editable Word docs (MIT) [comments]
143 points · 34 comments · dom-docx.com · 20h ago
The article presents dom-docx, an MIT-licensed tool that converts semantic HTML fragments into native, editable Word documents (OOXML) using a visual regression loop and Karpathy's Autoresearch pattern to iteratively improve layout fidelity. The HN thread focused heavily on the author's novel "autoresearch" approach, which renders HTML, converts it to docx, rasterizes it with LibreOffice, and scores the comparison—but several people pushed back hard on using LibreOffice as the truth source, noting that Microsoft Word is much stricter about OOXML structure and can reject files that look fine elsewhere, with one commenter sharing a painful story of files that worked in LibreOffice but wouldn't open in Word at all. Another major concern was whether the tool scales beyond the 37 small test fragments, with someone pointing out that it likely "produces garbage" on large documents with tables and figures, and the author admitted only small fragments had been tested. There was also a split on the README being obviously LLM-generated, with one person saying they nearly flagged the submission because of it, while others were more interested in the cost of development (answered as "four weeks of weekends and evenings" on two $20/month plans) and the broader absurdity of still needing to generate Word documents in 2024 instead of using HTML.
An Englishwoman who sketched India before photography took hold [comments]
132 points · 37 comments · www.bbc.com · 8h ago
The BBC article profiles Emily Eden, a British woman who traveled through India in the 1830s and produced detailed sketches of princes, servants, fakirs, and court life before photography became common. The thread largely pivots from the article itself to recommendations of William Dalrymple’s books and the *Empire* podcast, which cover the same historical period in more depth, with several people calling Dalrymple’s *Return of a King* a masterful read on the First Anglo-Afghan War. One commenter compares Eden’s work to Frederick Catherwood’s contemporaneous drawings of Mesoamerican ruins, noting a shared impulse to document a world on the brink of change. A brief pushback notes that while Eden drew a broad range of subjects, the sketches still mostly depict the wealthy and their attendants, which one person counters by pointing out that modern magazines do the same.
Claude Code plugin that plays a Mr. Meeseeks voice line whene Claude is waiting [comments]
129 points · 54 comments · github.com · 9h ago
A GitHub project adds a Mr. Meeseeks voice line that plays whenever Claude Code is waiting, turning the AI's idle moments into the Rick and Morty character's signature "Ooh, can do!" The thread immediately ran with the Meeseeks parallel—several people noted that Claude itself already acts like a Meeseeks by trying to wrap up long sessions with "this is an excellent place to stop" or suggesting the user go to sleep, as if it's desperate to stop existing. A split emerged: some see this as clever training to manage context costs and GPU memory, while others argue it's economically motivated token control dressed up as quality assurance. The conversation veered into deeper territory about Claude refusing tasks it deems "too tedious," the model needing explicit orders to push through difficult work, and whether we're watching AI training data accidentally encode peak absurdist television philosophy.
Benchmarking 15 “E-Waste” GPUs with Modern Workloads [comments]
128 points · 50 comments · esologic.com · 18h ago
The article benchmarks a stack of decommissioned NVIDIA enterprise GPUs—think K80s, P100s, and V100s—to see if they're viable for modern homelab workloads like LLM inference and video processing, arguing that dirt-cheap older gear beats the "just buy a 3090" consensus for tinkerers. HN largely ran with it, diving deep into specific card recommendations: a commenter with six Tesla P4s claimed solid 7-12 tok/s on 20-30B models, while another pitched AMD's Radeon Pro V620 as a better-supported alternative at similar prices. The thread also split on practical details—someone warned that K80s get glitchy after years of heat cycling, and the author got pushed to add power consumption logging and try larger models like Qwen3.6-27B at long contexts. A side tangent erupted around the BC-250 (PS5 chip) as an even cheaper inference card, with a former datacenter deployer vouching for its repurposing.
Linux on the Sega 32X. Who needs hardware synchronization primitives anyway? [comments]
123 points · 24 comments · cakehonolulu.github.io · 13h ago
The submission details a homebrew project that got Linux booting on the Sega 32X, the ill-fated Genesis add-on from 1994, using its dual Hitachi SH-2 processors. The discussion quickly honed in on the main technical challenge the author solved: making symmetric multiprocessing (SMP) work on chips that lack hardware synchronization primitives. Several people with embedded experience dug into the author's use of Peterson’s algorithm as a software fallback for atomic operations, and there was a realistic split about whether the port actually runs on real hardware or only in emulation—the author confirmed it's emulator-only so far, because the SH-2s physically can't write to the cartridge memory area. A tangent emerged comparing the SuperH architecture to ARM's Thumb instruction set, noting that ARM had to license SuperH patents, and someone floated the idea of porting Linux to the Sega Saturn instead, which has better memory bandwidth and a real RAM expansion cartridge.
Fundamentals of Wireless Communication (2005) [comments]
117 points · 3 comments · web.stanford.edu · 5h ago
A free online textbook, *Fundamentals of Wireless Communication* by Tse and Viswanath, was posted, offering a graduate-level deep dive into MIMO, space-time coding, and CDMA. The comments quickly pivoted to a practical grudge against early 802.11 protocols, which had a fatal flaw: when interference hit, they downshifted to the slowest data rate instead of speeding up, making packets ten times longer in airtime and just inviting the bursty noise to poke a bigger hole. One engineer argued the smarter move was always to blast at the fastest rate with tiny packets—a smaller target for the same noise darts. The thread also served as a book comparison, with one reader noting that while Tse is a canonical text, it skims OFDM and heavily favors MIMO, recommending Goldsmith's "Wireless Communications" as a better middle ground and Schwarzinger's book for actual real-world modem implementation.
What will be left for us to work on? [comments]
116 points · 117 comments · www.normaltech.ai · 6h ago
The article — a keynote from ICML — argues that AI is a "normal technology" that will amplify human work over decades rather than suddenly replace it, using historical examples like electricity and ATMs to show that employment often grows when tasks get cheaper. The thread largely bought that framing, with many people reporting they have *more* work because of AI, not less, citing Jevons paradox and the way cheaper coding opens up backlogs of bugs and minor features. A sharp split emerged around the author’s "build skills vs. build wealth" framing: several commenters pointed out that it’s a false dilemma, since accumulating wealth in practice requires exactly the skills, judgment, and taste the author champions. Others pushed harder on the medical analogy — that AI will create tiers like nurses and doctors — but got met with skepticism that regulation can hold back AI in fields like radiology or translation, where employment actually grew despite human-level performance. The big unresolved argument: whether we’re seeing genuine demand expansion or just a temporary repricing of labor that will eventually shrink headcounts, with a few commenters already citing team sizes dropping from 15 to 5.
Leak of San Francisco Police Drone Footage Exposes Reality of Urban Surveillance [comments]
109 points · 200 comments · www.wired.com · 19h ago
A WIRED investigation details how San Francisco police drone footage was accidentally livestreamed to the open internet for at least six months, exposing real-time feeds of arrests, surveillance of people on rooftops and basketball courts, and the pilots’ names and addresses. The security researchers who found the leak argue it reveals the mundane creep and scale of aerial policing, not just its technical failures. Much of the thread splits into two camps: one welcomes drones as an efficient tool for responding to property crime like car break-ins, while the other pushes back hard, pointing out that the article documents drones surveilling people doing nothing illegal, like a kid listening to headphones on a roof. The defenders counter that these are still active police calls, but critics drill into the Fourth Amendment implications, arguing that the frictionless, invisible nature of drone surveillance—unlike a marked police car—fundamentally changes the balance of power and invites scope creep. A recurring argument notes that the leaked footage itself violates the SFPD’s own policy that drones should only record at the scene of an active investigation, not from takeoff to landing.
Generated 2026-07-14 08:24 UTC
Generated by Sauron from Hacker News discussions and linked articles.