HN Brief: 2026-07-19
Today’s HN was split between two moods: deep admiration for scrappy hardware hacks that expose ridiculous vendor pricing, and deep suspicion about the motives behind every new AI release. The bowling alley ESP32 story became a referendum on captive markets and planned obsolescence, while Alibaba’s jumbo Qwen model sparked debates over whether open weights are genuine gifts or geopolitical honeypots. A quieter throughline was the backlash against AI branding—Perforce charging $500 for AI-narrated training videos, and a study showing AI advice makes people confidently wrong, both fed a growing sense that the hype is eroding trust in basic competence.
The threads worth your time: “Show HN: I replaced a $120k bowling center system with $1,600 in ESP32s” for the sheer vindication of open hardware against predatory vendors; “Claude Code uses Bun written in Rust now” for the messy fight over whether Anthropic is killing an open-source project or just making boring production changes; “Better and Cheaper Than IPTV” for the self-own of a piracy tool whose commit history betrayed its author’s denials; “The Last MPEG-4 Visual Patent Has Expired” for a clean punchline about patent hell that’s too little, too late; and “Texas Police Spent $4.5M on Four Chevy Tahoes” because the line-item breakdown of surveillance-ified SUVs is a masterclass in reading government waste.
Show HN: I replaced a $120k bowling center system with $1,600 in ESP32s
1116 points · 117 comments · news.ycombinator.com · 8h ago
The linked article wasn't available to this summarizer; from the discussion, a developer bought an abandoned 8-lane bowling alley in a rural town and replaced its proprietary, six-figure scoring system with about $1,600 worth of ESP32 microcontrollers and off-the-shelf parts. The entire thread immediately rallies around the project, with a heavy consensus that bowling equipment vendors have been price-gouging a captive market for decades—multiple people chime in with their own horror stories about $400 proprietary keyboards and absurdly expensive fouling sensors. A substantial split emerges over the question of kiosk-style automation and tap-to-pay: one side cheers the idea of removing friction and lowering prices, while the other argues that bowling alleys are social spaces where the human interaction at the counter is part of the draw, not something to eliminate. A mechanic who worked on old Brunswick A2 machines dives deep on how the entire pinsetter runs off a single relay, confirming that all that expensive electronics is literally just pushing one button, which only makes the original vendor pricing look worse. There’s also a quick detour into shoe rental debates—whether they're actually necessary or just a cash grab—and a useful warning from someone who installed these systems professionally: the mechanical forces in the pinsetter can destroy unprotected hardware, so mount everything away from the machine.
Qwen 3.8 [comments]
736 points · 514 comments · x.com · 14h ago
Alibaba’s Qwen team announced Qwen 3.8, a massive 2.4-trillion-parameter open-weight model, claiming it’s second only to the frontier Fable 5, and made a preview available immediately through their subscription plans and apps. The HN thread immediately dismissed the hype about raw parameter count, noting that both DeepSeek and Kimi have shipped similarly huge models for over a year, and argued that the real story is the strategic shift from China’s “value” small models to jumbo-sized open-weight releases. A major split emerged over motivation: some argued Alibaba is commoditizing intelligence to undercut US labs and avoid FUD about CCP control of closed models, while others warned this is a honeypot — open weights now, rug pull later once market share is locked in, and the data is likely poisoned on sensitive political topics. A smaller but pointed tangent pushed back on the “good for humanity” framing, asking how much of this is genuinely altruistic versus a geopolitical play to make the world dependent on Chinese infrastructure and debase American AI valuations.
Transcribe.cpp [comments]
713 points · 147 comments · workshop.cjpais.com · 22h ago
The article introduces transcribe.cpp, a new ggml-based transcription library that supports over 60 ASR models with GPU acceleration on Vulkan, Metal, and CUDA, and is designed as a drop-in replacement for whisper.cpp with verified numerical accuracy. HN mostly celebrated the project, with many developers appreciating that it’s maintained by a single developer (who also built the Handy app) and backed by Mozilla AI sponsorship, though a few commenters pushed back on the author’s claim that the blog post was AI-free, arguing that using AI transcription inevitably shapes the writer’s word choices—a point the author and others countered by noting the library is for speech-to-text, not LLM generation. Technical discussion zoomed in on the massive performance gap between Metal and Vulkan in the benchmarks, which the author clarified was just an M4 Max vs. an old integrated GPU comparison. A recurring side conversation debated whether Apple’s built-in dictation truly runs offline or sends contacts to Apple, with several readers vowing to switch to Handy or transcribe.cpp after learning about the privacy implications.
What I learned selling 2,500 MIDI recorders: Hardware is not so hard [comments]
380 points · 181 comments · chipweinberger.com · 12h ago
The author argues that building hardware isn’t as hard as software, based on his experience shipping 2,500 units of a simple MIDI recorder—but the thread immediately zeroes in on the 200,000 lines of code (85K in a Flutter app, 110K in firmware), which many find absurdly large for an ESP32-based device, and that’s where the real complexity lived. Several EEs push back hard, saying the hardware was trivial because it’s essentially an ESP32 with a single screw and no slides, so of course that was easy—the real challenge is when you need complex tooling, EMI compliance, or high-volume manufacturing. A long side conversation erupts about anti-counterfeiting chips and whether you ship HSMs to factories or pre-flash ICs in-house, with the author revealing he uses an STSAFE chip. Another thread covers compliance testing: he went with FCC and CE self-certification for about $5K, noting most Amazon sellers skip it entirely, but people caution that self-certifying is a footgun if you get big enough to attract regulator attention.
AI Mania Is Eviscerating Global Decision-Making [comments]
375 points · 217 comments · ludic.mataroa.blog · 21h ago
The article argues that AI mania has crippled rational decision-making across organizations, with the author claiming every AI project they've observed as a consultant has failed, yet executives are terrified to admit it because their jobs depend on professing faith in AI. The Hacker News thread mostly agreed with the thesis but pushed back on specifics, with several people pointing out that useful AI applications get quietly integrated into workflows while the flashy, visible "AI projects" like internal chatbots are the ones that bomb. A major split emerged around AI-assisted coding: some commenters argued the author ignored the real productivity gains engineers are seeing with LLMs for prototyping and learning APIs, while others countered that even those gains are fragile and won't survive the maintenance cycle. One skeptic noted that the author runs a consultancy selling "ancient techniques from books written before 2000," so of course they're hostile to AI—though others countered that ad hominem doesn't refute the core observation about executive groupthink.
Claude Code uses Bun written in Rust now [comments]
359 points · 486 comments · simonwillison.net · 12h ago
Simon Willison dug into his Claude Code installation and found concrete evidence that it's bundling a Rust rewrite of Bun (version 1.4.0) that hasn't been officially released yet. The HN thread immediately split into two camps: one side worried that Bun's open-source project is effectively dying under Anthropic's control, pointing out that a closed-source AI company is shipping a version of Bun the public can't build from a tagged release — and that the language switch from Zig to Rust alienated the small but passionate Zig community that made Bun interesting in the first place. The other side pushed back hard, arguing the Rust port is just a boring production improvement, that the code is publicly available in the repo and via canary releases, and that the real objection is just anti-Anthropic sentiment dressed up as FOSS concern. A messy but entertaining side-debate erupted over whether Zig is actually better than Rust or whether it's a DOA language plagued by the same footguns as C, with one participant getting roasted for praising Zig while admitting they hadn't written a single line of it.
Better and Cheaper Than IPTV [comments]
318 points · 99 comments · github.com · 21h ago
This is a CLI tool called Castor that finds video streams on web pages, transcodes them, and casts them to any DLNA-compatible TV from your terminal. The HN thread immediately split into two camps: one side called it a straightforward piracy tool, pointing out that the default config originally came pre-loaded with URLs to three known pirate streaming sites, and that the README was hastily edited to remove those after the post went up. The other side argued it's just a general-purpose casting utility and that what you point it at is your own business, though the "just bring your own sources" framing got undercut when people saw the commit history. A separate, entirely positive thread erupted around a different project someone posted in the comments—TV Explorer—which is a lightweight, ad-free static site for browsing free public HLS channels, and that side discussion got more praise than the main project.
Codex Resets [comments]
290 points · 182 comments · codex-resets.com · 23h ago
The article is a feed of announcements from OpenAI's @thsottiaux repeatedly resetting usage limits for Codex and ChatGPT Work, sometimes multiple times a day, often tied to hitting user milestones or apologizing for outages. HN spent most of the thread trying to figure out the economics and strategy behind it—some suspect it's a pre-IPO hype play to juice user numbers and developer goodwill, others argue the resets might not cost much if they're utilizing idle capacity or manipulating how usage is metered. A lot of the talk pivoted to comparing this with Anthropic's more stingy approach, with people pointing out that Anthropic also does periodic resets but is getting more restrictive with Fable access, while OpenAI is handing out "banked" resets that let you save extra usage for a month. There's a split between those who see this as a genius move to fend off competition from Chinese models like Kimi and GLM, and skeptics who note that "Codex" now just means the merged ChatGPT desktop app, so those 9M active users aren't 9M new developers—they're mostly just existing users who updated their software. A few deep-dive comments pointed out that the $200 Pro plan gives you roughly $9k-$10k in API-equivalent value per month, which sounds unsustainable but might just reflect extreme margins on API pricing.
OpenAI reduces Codex Model Context Size from 372k to 272k [comments]
282 points · 134 comments · github.com · 15h ago
OpenAI quietly reduced Codex's context window from 372k to 272k tokens in a recent commit, citing cost and latency tradeoffs rather than a capability regression. The thread quickly swerved into frustration with Codex's aggressive auto-compaction, which can kick in at 20% remaining context and loop endlessly, destroying session state—several people noted they can't disable it and lose valuable work. Others argued the reduction is temporary, pointed to DeepSeek's vastly superior caching for cheap long contexts, and called the 272k window an effective 516k thanks to GPT-5.6's token efficiency. A separate tangent emerged around a bug where Codex accidentally deleted home directories, leading to a mix of horror stories, defenses of container-level isolation, and a pointed system prompt change preventing destructive commands on broad paths.
Minecraft: Java Edition now uses SDL3 [comments]
245 points · 162 comments · www.minecraft.net · 11h ago
Minecraft's latest snapshot swaps out the GLFW library for SDL3 to handle window management, input, and platform integration, a backend change that carries most of the weight in this release. The thread quickly zeroed in on the known issues section, which warns that exclusive fullscreen mode crashes on Wayland and can cause problems on Windows with multiple monitors—several people pointed out that these are exactly the kind of blocking bugs that used to delay a snapshot, while others countered that snapshots are literally development builds meant to catch exactly these problems, and that exclusive fullscreen is a legacy mode most players already avoid in favor of borderless windowed. A separate vein of the discussion compared this to the rhythm game osu's recent switch to SDL3, which brought major latency and performance improvements, and debated why SDL3 adoption has been slow—blaming Linux distros like Ubuntu 24.04 that shipped just before SDL3 hit stable, and noting that many games already use it transparently through compatibility layers. The tone overall was that this is a boring-but-important infrastructure improvement that primarily benefits Linux users with proper Wayland support and fixes a long tail of input bugs.
Moonshot AI suspends new subscriptions due to Kimi K3 demand [comments]
171 points · 59 comments · x.com · 6h ago
Moonshot AI’s Kimi K3 model was so popular that it blew through their GPU capacity, forcing them to pause new subscriptions to protect service for existing users. The thread quickly split: some argued the surge is genuine quality-driven demand (K3 is pricier than alternatives like DeepSeek, so it’s not just people migrating to a cheaper model), while others hammered the terrible experience on the $20 plan—one person reported a 12-minute think time only to hit a daily quota, and multiple users said the base tier burns through its allowance insanely fast. A major pushback focused on the fact that Moonshot hasn’t released K3’s weights yet (scheduled for the 27th), meaning you can’t bypass their overloaded servers via third-party inference providers, which undercuts one of the main arguments for open models: distributed hosting would prevent exactly this kind of bottleneck. The thread also took a deep architectural tangent, noting K3 leans heavily on RNN/linear attention layers (like xLSTM) and sparking a lament that Europe never scaled those architectures to LLM sizes despite having the academic chops.
I joined the IndieWeb, here's what I learned [comments]
134 points · 83 comments · en.andros.dev · 11h ago
The article is a detailed, personal explainer of the IndieWeb movement—its philosophy of owning your content on your own domain, the key protocols like Webmention and microformats, and which specific pieces the author chose to implement or skip on their own site. The HN thread immediately split into two camps: people who found the technical breakdown useful and were sharing their own setups, and a vocal group pushing back hard that the IndieWeb is a "tech soup" that is completely inaccessible to 90% of users who don't know what DNS is and won't touch a command line. A long, sharp argument erupted over whether the IndieWeb's "scratch your own itch" mentality is a feature or a fatal flaw, with one side insisting that without a one-click, App Store-ready solution it remains "NerdNet" and the other countering that it's not a business and doesn't need to conquer the market. A notable sub-thread dug into the irony of the author using a `.dev` domain owned by Google, questioning how independent you really are when a corporate giant controls your TLD, while another practical debate focused on the fact that the IndieWeb's recommended `h-feed` standard for feeds is almost completely unsupported by feed readers, making a traditional RSS feed essential for actually being followed.
The Last MPEG-4 Visual Patent Has Expired [comments]
131 points · 29 comments · www.phoronix.com · 6h ago
The linked article from Phoronix reports that the final MPEG-4 Visual (Part 2) patent has expired in Brazil, marking the complete end of patent restrictions for the codec behind DivX and Xvid. The HN crowd immediately jumped on the correction that this is *not* H.264—that's MPEG-4 Part 10, and its patents still loom, with one key date landing around 2030. People dug into the implications for open-source tooling, noting that while the old codec is now fully free, hardware support and practical utility have been overtaken by H.264's long tail and more capable modern codecs like AV1. There's a clear split: some see this as a symbolic victory for patent expiration, while others dismiss it as too little, too late, arguing the real battle remains over H.264 and the systemic problem of software patents granted post-specification.
The death and rebirth of my home server [comments]
113 points · 72 comments · sgt.hootr.club · 12h ago
The article is a detailed walkthrough of a home server death and rebuild after its Raspberry Pi’s SD card finally gave out, with the author moving everything to a btrfs pool on old spinning drives and using NixOS tricks like zram and volatile logging to stop killing flash. The thread immediately pivots into a long-running HN debate about whether Raspberry Pis are uniquely terrible at keeping SD cards alive—many argue it’s actually bad USB power or cheap cards, not the Pi’s fault, while others point out that Raspbian historically enabled atime and desktop logging by default, basically begging for corruption. Some commenters swear by booting from USB SSDs or PCIe NVMe on Pi 5s, and a few chime in with stories of embedded Linux systems running the same SD card for decades without issue, which undermines the “Pi is the problem” narrative. A separate tangent digs into hardware lifespan monitoring: SMART data exists for HDDs and SSDs, but most home users don’t proactively replace drives, so the consensus is that redundant backups are more practical than trying to predict failures.
Bananas sprout in Rayleigh Garden UK after 15 years [comments]
109 points · 82 comments · www.bbc.com · 9h ago
The BBC article covers a couple in Essex who finally got bananas from their 200 plants after 15 years, pinning it on warmer UK summers and milder winters, though the specific variety (Musa Basjoo) is inedible — "like a mouthful of ball bearings with half a teaspoon of banana." HN immediately dove into banana history, correcting the common myth that all bananas trace back to a single seedless clone; people pointed out that seedless varieties were cultivated repeatedly across Southeast Asia, and that the Cavendish is just the latest commercial monoculture, preceded by the Gros Michel (which explains why artificial banana flavor tastes different). A secondary thread erupted over whether rising temperatures are real or just artifacts of urban heat islands and poorly sited weather stations — one commenter argued extensively that UK temperature records are corrupted by runways and immigration-driven population density, while others shot back that Lingwood in Norfolk is hardly an airport and that dismissing every record as measurement error is bad faith. There was also a side discussion about a banana tree fruiting in Switzerland, with horticulturalists explaining that getting fruit is easy but ripening it over a full year of frost-free conditions is the real trick, and that cold-tolerant banana species exist but produce lousy fruit. The thread largely left the BBC article's optimistic climate-adaptation framing behind and became a slugfest over banana botany and climate data integrity.
Texas Police Spent $4.5M on Four Chevy Tahoes [comments]
104 points · 48 comments · www.thedrive.com · 6h ago
Here's the rundown on that Texas police Tahoe thread from overnight. The article details how the Texas Department of Public Safety spent $4.5 million on four Chevy Tahoes packed with FalcoNet, an Israeli surveillance system that secretly intercepts cell phone data by mimicking a cell tower. A lot of the pushback focused on the Fourth Amendment, with a split over whether four trucks driving around a state as big as Texas counts as "mass surveillance" or something more targeted and legally defensible, though the article itself points to a recent Supreme Court ruling that location data tracking is a search. Several people dug into the purchase order line items, arguing the $150k per Tahoe was actually the cheap part, and questioning why the police memo was laced with urgent "duty of care" language for a brand-new, unproven tech toy. A cynical side of the thread pointed out that if you just leave your phone at home the whole system is useless, while others countered that the real point isn't catching smart criminals—it's building a retroactive dragnet for when the state decides it needs a short list of suspects.
FDA approves new kind of cholesterol pill [comments]
101 points · 64 comments · www.fda.gov · 20h ago
The FDA approved the first-ever oral PCSK9 inhibitor, Lipfendra, for lowering LDL cholesterol—a major shift from the injectable-only options like Repatha that have been on the market for years. The thread quickly zeroed in on the practical question of cost and insurance coverage, with commenters arguing that the oral form won’t automatically fix the prior-authorization gatekeeping that has limited PCSK9 use, since statins are still far cheaper and insurers will keep demanding patients try them first. Several people pointed out that the molecule itself is a synthetic marvel, linking to Derek Lowe's writeup on its beastly synthesis, while others derailed into a long, heated debate about whether diet and exercise can adequately manage cholesterol—sparking pushback from those who cited the 20-30% dietary effect ceiling and noted that genetic hypercholesterolemia patients often can’t diet their way out of it. A significant chunk of the discussion went meta, with one commenter questioning whether LDL is really the right biomarker and arguing that particle size and glycation matter more, though multiple people shot that down by citing the massive body of evidence linking ApoB and LDL to atherosclerosis. The thread also surfaced news about upcoming gene-therapy PCSK9 inhibitors and lp(a) drugs, giving a sense that while this oral pill is a step forward, the field is already racing toward even more transformative therapies.
AI advice made people 3x less accurate but 2x confident, researchers found [comments]
97 points · 38 comments · thenextweb.com · 1h ago
A study found that giving people access to AI advice collapsed their willingness to say "I don't know" from 44% to 3%, slashed accuracy from 27% to 9%, and cranked up confidence from 30% to 76%—researchers deliberately used questions where the model was usually wrong, so the effect isn't just sensible delegation. The HN thread split hard: some took it as confirmation that AI will turn into an echo chamber of confident wrongness, especially as people anthropomorphize and prefer flattering lies over correction. Others tore into the study's methodology, arguing it's basically an open-book quiz with a textbook full of errors—nothing specific to AI, and the experimental setup doesn't test real-world LLM issues like sycophancy or cognitive helplessness. A few commenters pointed out that the study's source links were broken, and someone linked the actual preprint. The broader takeaway was that the problem isn't just AI, but a system designed to never say "I don't know," and people are happy to use it to pretend they know things.
I burned all my tokens researching how to save tokens [comments]
89 points · 127 comments · quesma.com · 10h ago
The article walks through building a multi-agent deep research pipeline to study AI token economics, starting with the author burning through their entire Claude Max subscription in 30 minutes and ending with a system that orchestrates Claude, Codex, and Antigravity as subagents with pinned model roles, shared memory, and automatic fallback. The thread mostly split between people who've been through the same loop — swapping war stories about multi-model orchestration and caching strategies — and a more skeptical camp arguing that most "token saving" techniques are actually cache killers that cost more than they save because they make prompt prefixes dynamic. A recurring tension came from someone pressing others to name actual shipped products rather than blog posts about optimizing AI, which drew a few concrete examples but also accusations of bad faith. The deeper engineering debate landed on whether context compaction and adaptive model routing are worth the complexity, with several people insisting that freezing your retrieval early and paying a slightly larger fixed prefix beats any clever dynamic scheme, while others pointed out that you just can't share caches across different models at all.
Heavy TV watching associated with smaller brain structures, study finds [comments]
82 points · 51 comments · dornsife.usc.edu · 2h ago
A new study found that people who reported watching TV “very often” in midlife showed smaller brain structures and more white matter damage decades later, with the effect holding even after controlling for exercise and other factors—and notably, sitting at a cognitively demanding job didn’t have the same harm. The HN crowd immediately pounced on the correlation-vs-causation problem, pointing out that without a baseline MRI, the study can’t rule out that people with already-different brains just happen to watch more TV. Several people dug into the finding that occupational sitting looked protective, arguing it supports a “use it or lose it” model where passive consumption is the real culprit, not sitting itself. A recurring split emerged over whether content matters: one side insisted any passive screen time rots the brain, while others pressed for a breakdown between trashy reality shows and something like an Attenborough documentary. There was also a running joke about how modern streaming and YouTube are just “TV by another name,” and a handful of commenters wondered aloud whether the rise of AI tools will shift knowledge workers from active creation to passive curation—a potentially similar form of brain rot.
Harness Engineering [comments]
77 points · 32 comments · github.com · 23h ago
The repository proposes “harness engineering”—shaping the environment (context, tools, constraints) around a fixed coding agent to get better output, rather than tuning the model itself. The author (who also wrote the original OpenAI blog on this) jumped into the thread and took direct questions, which drew a mix of genuine technical curiosity and sharp skepticism. Several people pushed back on the self-quoting in the readme and the 100x productivity claim, with some joking that the agent must have written that number too, while others dug into how to scale the approach to cheap agents, maintain naming coherence across a codebase, and use static verifiers to catch “slop” before it ships. A notable split emerged over the prose: the author admitted the entire repo was written by GPT-5.6 Sol Ultra, which made some readers distrust the recommendations as pattern‑matched babble, while others found the underlying ideas about cumulative organizational memory and adversarial guardrails worth taking seriously.
Clever hacker fits 537,000 domains in a $5 ESP32 ad-blocking dongle [comments]
69 points · 24 comments · www.tomshardware.com · 8h ago
The article covers a developer who crammed 537,000 hashed domain names into a $5 ESP32-C3's flash memory to build a backup DNS ad-blocker that responds in ~10ms, using 40-bit FNV-1a hashes and a binary search. The HN crowd immediately zeroed in on the missing Bloom filter—several people noted the README already lists it as a planned improvement, sparking a debate over whether a Bloom filter would actually save time when flash reads are microseconds and WiFi latency dominates. Others dug into the storage math, questioning why the firmware takes up two full partitions and whether dropping OTA support for extra domains is worth the trade-off. A separate thread dismissed the whole project as over-engineering for most users, pointing out that browser extensions or a Pi-hole are simpler, and that Tom's Hardware itself is so bloated with ads you'd need more than a DNS filter to read it. The conversation also veered into genuine appreciation for what $5 of compute can do, contrasting this lean IoT approach with the bloat of smart appliances.
Ollama: All Aboard Open Models [comments]
63 points · 43 comments · ollama.com · 14h ago
The linked article is Ollama’s founding story and funding announcement, positioning itself as the "Docker for open models" and touting 8.9 million developers and 85% Fortune 500 usage. The HN thread was overwhelmingly hostile, with the top and most-upvoted comments telling people to stop using Ollama entirely. The central charge is that Ollama has systematically rebranded features from llama.cpp as its own without credit or proper license attribution, and the article’s omission of any acknowledgment of Georgi Gerganov’s work was seen as a glaring insult. Many commenters pointed to specific technical failings—like a year-old issue for basic MoE CPU offloading that llama.cpp handles—and argued that Ollama’s cloud subscription and VC money signal inevitable enshittification. A significant split emerged between people who found Ollama’s ease-of-use genuinely valuable for local model experimentation and those who argued that better technical options like unsloth or raw llama.cpp exist, with the latter camp largely winning the thread’s tone.
Terence McKenna's Mega Bad Trip [comments]
59 points · 52 comments · psychedelics.community · 8h ago
The article digs into Terence McKenna's harrowing 1988–89 mushroom trip in Hawaii, which ended his regular psychedelic use and exposed the gap between his public advocacy of heroic doses and his private terror. HN immediately split over whether you need a "spiritual practice" to weather such experiences, with one camp insisting atheists and nihilists handle psychedelics just fine—and the other countering that some grounding framework, even if not religious, is essential to avoid ontological collapse. A strong vein of the thread pushed back hard on McKenna himself: callers argued he became a for-profit bullshit artist who let celebrity and speaking fees lock him into a persona, and that his legacy would have been stronger if he'd been honest about his limits instead of continuing to hype the heroic dose to fans. The piece also drew comparisons to Alex Jones and the observation that McKenna's Timewave Zero predictions fizzled, but a few commenters defended his novelty theory as interestingly prescient about the smartphone era and meme culture.
HMD Touch 4G [comments]
55 points · 54 comments · www.hmd.com · 5h ago
The linked HMD Touch 4G is a tiny hybrid phone sold only in India that tries to bridge feature phones and smartphones with a touchscreen, 4G, and a proprietary chat app called Express Chat. The thread immediately split into two camps: one group was interested in the hardware as a potential small, cheap burner or minimalist device, noting it costs about $45 and fits a niche for people who want a tiny phone without a big screen. The other group tore into the locked-down software, with multiple people pointing out that HMD won't let you unlock the bootloader, making it a device you don't actually control—a dealbreaker compared to old flashable feature phones. A surprisingly heated debate erupted over whether 4G is still viable in 2026, with one person claiming it barely works while others firmly corrected that 4G remains the dominant network globally and 5G has worse range, though some argued that new 5G-only towers are already degrading 4G coverage in certain regions. The consensus was that the dedicated Express Chat app is a non-starter—without WhatsApp support, which famously dropped KaiOS last year, the phone's social utility collapses, making it hard to see why anyone would use this over a cheap Android.
HomeLab #1: MikroTik as a Home Router [comments]
55 points · 37 comments · justsomebody.dev · 3h ago
The linked article is a detailed guide on setting up a MikroTik router for a home lab, walking through ISP-specific configuration, PPPoE, VLANs, MAC cloning, and bufferbloat mitigation with FQ-CoDel. The HN discussion quickly split into two camps: longtime MikroTik enthusiasts who praise the hardware’s longevity and flexibility, and frustrated users who find RouterOS’s interface and configuration process needlessly arcane—several argued that a router shouldn’t require an LLM to set up basic features like port forwarding or hairpin NAT. A recurring counterpoint was that MikroTik’s software phoning home and its poor default bufferbloat handling make it a hard sell for new home users, with people pushing alternatives like OPNSense on Proxmox or Ubiquiti for a more polished out-of-box experience. The thread also veered into tangential debates about OpenWRT on MikroTik switches, the viability of Zyxel gear, and whether the company’s GPL compliance is still a sore spot, but the core tension was clear: this is powerful gear if you already speak the language, and a headache factory if you don’t.
Perforce charges $500 for training training videos.. and it's AI narrated [comments]
52 points · 80 comments · training.perforce.com · 14h ago
Perforce is charging $500 for a training course on its Helix Core version control system—and the videos are narrated by an AI text-to-speech voice. The thread lit up around that: several people who tried the free sample noticed the synthetic voice had weird pauses and glitches, and argued that charging premium prices for AI-generated content signals contempt for customers. Others pushed back, saying $500 is nothing for an enterprise tool, the voice quality is fine, and AI narration makes updates cheaper and faster than re-recording with a human. A deeper tangent emerged about Perforce itself: it survives because Git can't handle large binaries, making it the de facto standard in game dev and automotive, even though Git fans can't believe it's still around.
Modder Runs GTA III Inside GTA: San Andreas on an In-Game TV [comments]
52 points · 11 comments · videocardz.com · 3h ago
A modder named Dryxio got Grand Theft Auto III to run as a fully playable game on an in-game TV inside GTA: San Andreas, rendering the older game directly into a GPU texture rather than just playing a video. The thread is mostly just people being delighted by the sheer absurdity of it—someone immediately joked that GTA 3 plus GTA 3 equals GTA 6, and another quipped that this must be how you unlock Half-Life 3. A few commenters connected it to other deeply nested mods, like running Morrowind inside a Pip-Boy in Fallout 4, framing it as part of a proud tradition of "because I can" engineering. One person noted that if Rockstar could just mod GTA V to run on the Steam Deck, that would be more useful, but others quickly shot that down by pointing out GTA V already runs fine there.
C64 Basic Dungeon Crawler: Goblin Attack (C64 Basic Part 8) [comments]
51 points · 3 comments · retrogamecoders.com · 7h ago
The linked article is the eighth part of a series on coding a dungeon crawler in Commodore 64 BASIC, this time making goblins chase the player instead of standing still. The HN thread is very short, with only three comments, but they take a sharp left turn from the technical content. One person gripes about the accompanying YouTube video being unwatchable due to incessant meme clips inserted throughout. The other two comments completely abandon the code discussion to joke about the game promoting "greenskin genocide," with one user jokingly calling for jail time for goblin sympathizers—a bit of absurdist humor that entirely sidesteps the deep dive into memory management and BASIC tokenization bugs.
Valve say there's no end in sight to the memory crisis, prices going to increase [comments]
46 points · 28 comments · www.pcgamer.com · 14h ago
Valve told Bloomberg that memory prices are still getting worse and will keep rising for the foreseeable future, driven by the AI boom gobbling up supply. The thread largely sidestepped the Steam Machine news and instead lit into the broader hardware landscape: one camp argued the end of Moore's Law means even a PS6 won't offer a meaningful leap, while another insisted the real problem is a Micron-Samsung-Hynix cartel artificially inflating prices — and that U.S. politicians are pushing to block cheaper Chinese DRAM. A few people pointed out that Valve itself failed to hedge memory costs for the Steam Machine, so their grim forecast should be taken with a grain of salt. The conversation also turned to whether this generation of consoles has been worthless, with several commenters noting that the only "exclusives" worth playing are also on PC.
Generated 2026-07-19 22:58 UTC
Generated by Sauron from Hacker News discussions and linked articles.