HN Brief: 2026-06-23
Today's Hacker News swung between AI skepticism and surveillance fatigue. Several threads poked holes in model marketing—GLM-5.2's "lossless" quant, Claude Code's hidden reasoning, Moebius's cherry-picked samples—while a separate wave of stories documented how face scans, license-plate readers, and TV proxy SDKs are eroding privacy in plain sight. A third throughline tracked infrastructure anxiety: Canada's nuclear timeline skepticism, Microsoft locking down gas-fired data centers, and the looming Secure Boot cert expiration that reminds everyone who controls the boot chain.
"Never Give Them Your Face" is worth clicking for practical subthreads on fooling age gates with MKBHD's face and iCloud Private Relay's role in anti-tracking. "Flock-Powered Police Chiefs Stalking Women" delivers a sharp Fourth Amendment split between plain view and Carpenter v. US, plus Flock's own lawyer admitting abuse is 'the most common thing.' "GLM-5.2 – How to Run Locally" splits into solo-dev Studio purchases versus cloud rental, with a harsh reality check on 'lossless' quant claims for real coding tasks. "Will It Mythos?" offers vivid stories of Fable solving problems Opus couldn't, balanced by skepticism that it's just safety-filter social engineering. "Jobs and Software Is Fucked" pairs hiring-market despair with a counterpoint that hardware verification and agentic startup roles are desperate for engineers, leavened by a wonderful Heat tangent.
Steam Machine launches today [comments]
1508 points · 1321 comments · store.steampowered.com · 14h ago
The Steam Machine launched today, Valve’s new console-sized PC running SteamOS, starting at $1,049 without a controller. The main debate is about the target audience: a lot of people argue you can build a cheaper, more powerful, fully upgradeable PC for the same money, but the counter from others is essentially “I have a job and a kid, I just want to plug something into my TV and have it work without spending a weekend tweaking BIOS fan curves.” A vocal group is actually buying it specifically as a Linux workstation or HTPC, fed up with Windows and recent macOS releases, and insist SteamOS “just works” for their workflow. The reservation system (a lottery instead of straight preorders) generated some grumbling, though most assume it’s just an anti-scalper measure, while a side-discussion emerged comparing the Steam Machine GPU to scraped-together PS5 dev-kit hardware like the BC-250.
Pledging another $400k to the Zig software foundation [comments]
770 points · 260 comments · mitchellh.com · 18h ago
Mitchell Hashimoto announced he's pledging another $400,000 to the Zig Software Foundation, bringing his total to $700,000, explaining that he respects Zig's technical ambition and its willingness to be weird—like the no-LLM contribution policy—even though he personally uses AI heavily and disagrees with some of the foundation's stances. The thread mostly ignored the Zig-specific substance and turned into a sprawling, heated debate about whether money can buy happiness, with one camp arguing that wealth removes obstacles to contentment and the other insisting that the kind of people who get rich are precisely the kind who end up miserable. Several people pushed back hard on the idea that a little money goes a long way, pointing out that the "hump" where financial stress ends is well above the poverty line and closer to double the median income. Others used examples like Musk and Trump to argue that endless wealth doesn't fix loneliness or purpose, while a few commenters countered that without money those people would be even more miserable, and that Buddhist monks are happy because they've opted out of the entire framework. The conversation completely sidestepped Zig's compiler progress or the AI policy debate, settling instead into a familiar HN rabbit hole about class, motivation, and whether giving away large sums is a sign of healthy wealth or just another side quest.
Never Give Them Your Face [comments]
711 points · 395 comments · nevergivethemyourface.com · 18h ago
The article is a polemic against mandatory face-scans and ID uploads sold as "age assurance," arguing it's really universal identity surveillance that can't be undone. HN largely agreed with the premise, but the thread quickly split: some pointed out that California has already made age verification illegal while requiring minimal-data age assurance, a legal path many hadn't heard of. Others pushed back hard on the article's implied equivalency, noting that banks and airports already demand ID and face scans, so the horse is long out of the barn—and that the site itself ironically collected emails until someone called it out. A surprisingly practical tangent emerged: people shared that they trick age-gates with a looped video of MKBHD’s face, and a full subthread debated whether Tor actually works for this or if iCloud Private Relay is doing more to break IP-based tracking. The quietest but sharpest correction came from someone arguing that zero-knowledge proofs exist and could solve the problem without surveillance, but promoters of the laws avoid them because stripping anonymity is the whole point.
Flock-Powered Police Chiefs Stalking Women Shows Why Warrants Are Needed [comments]
495 points · 207 comments · ipvm.com · 12h ago
The article details multiple documented cases of police chiefs using Flock’s ALPR system to stalk ex-partners and romantic rivals, arguing this pattern proves warrants should be required before querying the database. The HN thread immediately split along Fourth Amendment lines: one side insisted that filming in public requires no warrant under the plain-view doctrine, citing the 9th Circuit’s ruling—but the other side pushed back hard that *Carpenter v. US* controls here, because systematic long-term collection of location data is fundamentally different from an officer jotting down plates with a pen. A recurring point was that Flock’s own Chief Legal Officer effectively admitted the abuse is “the most common thing” and that a warrant requirement is inevitable, which undercuts the company’s defense that this is rare misconduct by a few bad actors. Several people also flagged that their local ACLU chapters have already gotten Flock cameras removed, and that when FOIA requests were made for Flock data, the state simply exempted it from disclosure—further eroding trust that the system is either legal or accountable.
Canada plans 'nuclear renaissance' with up to 10 reactors built by 2040 [comments]
450 points · 297 comments · www.cbc.ca · 12h ago
The article lays out Canada’s plan to construct up to 10 new nuclear reactors by 2040, including small modular reactors and large-scale units, at a cost north of $100 billion — with the energy minister arguing there’s no credible path to doubling the grid by 2050 without nuclear. The thread quickly zeroed in on the gap between “plan to build” and “will actually have operational reactors,” with several people noting that even the timeline to *start* construction by 2035 is wildly optimistic given Canada’s project delivery track record. A sharp split emerged on whether the minister’s “no credible plan without nuclear” claim holds: one side pointed to Alberta’s strong wind and solar growth (and the province’s recent regulatory hostility toward renewables) as proof that the real fight is political, not technical, while the other side argued that baseload from hydro and nuclear already covers over 60% of Canada’s electricity and that adding more intermittent generation without firming capacity is a fantasy. The small-versus-large reactor debate got heated — proponents of SMRs said factory production and remote deployment make sense, but skeptics countered that large reactors are vastly more economical and that nobody has actually built SMRs at scale since the 1950s. There was also a strong undercurrent that Canada should standardize on the CANDU design and treat the whole program as a single coordinated build rather than a collection of bespoke projects, echoing the US’s failure to get nuclear costs under control.
GLM-5.2 – How to Run Locally [comments]
351 points · 151 comments · unsloth.ai · 10h ago
The article covers how to run Z.ai’s new 744B-parameter GLM-5.2 model locally using Unsloth’s dynamic quantizations, which shrink it to fit on a single 256GB Mac or a 24GB GPU with RAM offloading while claiming near-lossless performance at 4-bit. The thread quickly zeroed in on hardware cost — estimates ranged from $50k for a multi-GPU workstation to a Gemini hallucination of $500k — with people pointing out that even a 2-bit quant requires 245GB and that real-world accuracy drops sharply below Q4, no matter what the KL-divergence numbers say. Several commenters pushed back on the "lossless" marketing, arguing that KLD on arbitrary corpora doesn’t translate to long-context coding tasks, and that you typically need to step up a quantization tier to avoid degradation. Others debated whether it’s worth buying a $10k Mac Studio with 512GB unified memory for 20 tok/s at FP4, or better to rent cloud GPUs for a team and stay immune to API outages and caps — with a strong split between solo devs who can’t justify the upfront cost and teams who can amortize it.
The text in Claude Code’s “Extended Thinking” output [comments]
300 points · 206 comments · patrickmccanna.net · 17h ago
Anthropic’s Claude Code logs local "thinking blocks" to disk, but a blog post by Patrick McCanna revealed those logs contain nothing but a 600-character encrypted signature — Anthropic holds the key, and the API only returns a *summary* of the reasoning, not the full trace, which requires an enterprise agreement to access. The HN crowd immediately split: some shrugged and said "duh, OpenAI and Google do the same thing, reasoning summaries are actually more useful for daily work anyway," while others argued this is a deliberate moat — hiding the raw reasoning tokens makes it much harder for competitors to distill the model, and several commenters pointed out that open-source model makers were openly training on those exact traces. A smaller camp pushed back that the real issue isn't competition but alignment theater — one person recalled the brief 12-hour window when an early ChatGPT reasoning model leaked its raw chain-of-thought and you could literally see it navigating around its own safety guidelines, suggesting the summaries are as much about PR as IP protection. The thread also devolved into a genuinely pleasant tangent when the author admitted he inverted the BMP/JPEG analogy in the article, and commenters praised the human charm of his typo-ridden correction ("wait till my father hears about this!") as an increasingly rare signal of real authorship.
Jobs and Software Is Fucked [comments]
296 points · 262 comments · urflow.bearblog.dev · 12h ago
The article is a personal essay from a recently laid-off Blizzard engineer describing a brutal job search full of AI-proctored coding filters, silent rejections after final rounds, and the sinking feeling that companies are using AI hype to justify pulling up the ladder for juniors. The HN thread largely nods along in agreement—many with 10–15 years of experience report the same wall of silence, with recruiter spam entirely dried up for some while ticking back up for others, suggesting a deeply uneven recovery that favors niche specialties. The main split is between those who see this as a cyclical downturn exacerbated by AI panic and those who argue the market has simply rotated: hardware verification shops and startups built on agentic systems can’t find enough PCIe or coding-agent-savvy engineers, and that’s where the offers are. Several commenters push back hard on the author’s “don’t concede to AI” stance, arguing that refusing to use LLMs in interviews is pure self-sabotage—the market has already decided, and principled indignation won’t stop it any more than Luddites stopped the power loom. A side detour into the exact format of the Bearblog date field and a surprisingly deep tangent about Neil McCauley’s “walk away in 30 seconds” speech from *Heat* give the thread some texture, but the core takeaway is that the old hiring funnel is broken, and the people still getting hired are either brand-name pedigrees or specialists in whatever the hype cycle is two years ahead of.
Moebius: 0.2B image inpainting model with 10B-level performance [comments]
278 points · 69 comments · hustvl.github.io · 18h ago
The article presents Moebius, a lightweight image inpainting model with only 0.22B parameters that claims to match the quality of 10B-parameter models like FLUX.1-Fill-Dev while running 15x faster. HN immediately split between wanting to actually test it and skeptical critics who scrutinized the sample images, with one person pointing out structural artifacts like an elongated surfboard and a weird house corner that the paper’s own comparisons conveniently skip. A handful of people found working HuggingFace demos, though most reported the demos crashing or producing crappy results, and the discussion acknowledged that the impressive-on-paper numbers might not survive real-world use. The conversation also veered into whether additive tasks like adding awnings to photos even need inpainting, with several people arguing that classical photogrammetry or simple quad transforms would be cleaner. A smaller thread debated whether the paper’s marketing language—calling it a “synergy” that shatters an “impossible triangle”—is a sign of AI-generated fluff creeping into scientific writing, while others defended the technical contribution as genuinely impressive for its size.
Nearly half of LG smart TV apps contain residential proxy SDKs [comments]
232 points · 146 comments · spur.us · 11h ago
Spur found that nearly half of LG smart TV apps pack residential proxy SDKs, quietly turning fish-tank screensavers and solitaire games into tools that route strangers' internet traffic through your living room. The HN thread split hard: some argued the consent prompts are actually more upfront than expected (the app asks once and then runs in the background after closing), while others called the whole model indefensible and pointed out that Amazon and Roku already ban these SDKs—so why haven't LG and Samsung? A vocal contingent pushed back on the article itself, flagging that it reads like AI slop (short declarative sentences, ChatGPT-style headers) despite the solid data, and one person even posted a bookmarklet to fix the page's scroll hijacking. A key clarification emerged: these are third-party shovelware apps, not LG's built-in software, though the LG app store is described as a "cesspit of spam" that feels stuck in 2009. The deeper worry is that without a blocklist for private IP ranges in some SDKs, that TV becomes a foothold into your home network—router admin panels, NAS devices, printers, all potentially reachable.
Alan Greenspan has died [comments]
225 points · 222 comments · archive.ph · 20h ago
The Washington Post obituary for Alan Greenspan, who died at 100, traces his arc from revered "Maestro" of the Fed who oversaw the Great Moderation to a figure whose deregulatory fervor and low-interest-rate policies are blamed for setting the stage for the 2008 financial crisis. The Hacker News thread immediately split into a heated debate about whether Greenspan deserves that blame or if he's a convenient scapegoat whose successors and Congress deserve more of it. One strong argument pushed back hard, pointing out that the worst of the housing crisis hit a year after he left and that he couldn't be expected to have supernatural foresight when even Michael Burry didn't start shorting the market until late 2005. The counter came back just as forcefully, arguing that he actively campaigned against any regulation of derivatives and mortgages, pressured a regulator to back off, and his "Greenspan Put" philosophy of flooding markets with liquidity at every crisis destroyed moral hazard—a pattern people feel we're still living with the consequences of today. A tangent also emerged from someone citing Greenspan's own youthful advocacy for the gold standard, which another commenter dismissed as rich coming from the man who oversaw two decades of unchecked credit expansion that ultimately blew up.
Why Drawing Tablet Brands Won't Collaborate on Linux Floss Drivers [comments]
220 points · 114 comments · www.davidrevoy.com · 19h ago
The linked article is a detailed account from David Revoy, a Linux tablet reviewer, about his failed attempt to convince drawing tablet brands like Gaomon, Huion, and XPPen to collaborate on open-source drivers by sharing their hardware specs with the Linux input team. The core problem, as the brands explained, is that the entire infrastructure (repos like `libwacom` and `wacom-hid-descriptors`) is branded after their biggest competitor, Wacom, making them feel like they'd be helping a rival while giving away their own specs. The HN thread immediately seized on this, with a clear split: one side sees a straightforward fix in just renaming the repos to something vendor-neutral, while the other side argues that renaming is a massive, painful, and politically fraught undertaking that would break countless links and references, and that the maintainers simply don't see the branding issue as a technical problem worth their time. A long digression emerged comparing this to the "master vs. main" and "GIMP" naming debates, with some arguing that the resistance to renaming is purely emotional and a sign that the maintainers are disconnected from the project’s users, while others counter that "wasting time on a rename" is a logical prioritization call when you could be fixing actual bugs.
Mexican government unveils a prototype for a new homegrown, ultra-affordable EV [comments]
203 points · 173 comments · gizmodo.com · 15h ago
Mexico’s government unveiled a prototype for the Olinia Uno, a six-passenger EV with a 77-mile range and 31 mph top speed, priced at roughly $8,500 and aimed at urban use. The HN discussion immediately zeroed in on those specs as a non-starter for the US market—most people argued Americans won’t tolerate a sub-200-mile range or a car that can’t do highway speeds, even at that price point. A significant counter-argument pushed back by citing data that the vast majority of trips are short, and that this is a second-car or city-only vehicle, not a do-everything sedan. The conversation split between Mexicans familiar with local driving conditions (who found the range and speed fine for Mexico City traffic) and Americans who projected their own commuting needs onto it, with several people noting the real US barrier won’t be legal pushback but consumer demand, not tariffs.
Show HN: Got sick of ads, so I made my own logic puzzle site [comments]
185 points · 112 comments · puzzlelair.com · 19h ago
The submission is a personal logic puzzle site called Puzzle Lair, built as an ad-free alternative with a newspaper aesthetic and a one-time unlock monetization model. HN immediately compared it to Simon Tatham's venerable puzzle collection, which most commenters seemed to prefer for depth and polish. The nonograms drew sharp criticism—several people called them AI-generated, pointing out symmetrical diamond patterns, odd grid sizes not in multiples of 5, and missing visual conventions like proper border grouping. The creator defended the puzzles, admitting some were made with help from an LLM but pushing back that the generators are each hand-tuned and that the site handles uniqueness and difficulty in ways models can't one-shot. A separate thread erupted around a user who thought an X Sudoku puzzle was invalid until others explained the diagonal constraint, and the discussion veered into recommending variant sudoku resources.
Show HN: Oak – Git alternative designed for agents [comments]
182 points · 162 comments · oak.space · 16h ago
The submission is a new version control system called Oak, built from the ground up for AI coding agents, featuring virtual mounts that let agents work on repos without downloading them fully and squash-merging branch-per-session workflows. The HN crowd immediately split into two camps: one side spent most of the thread championing Jujutsu (jj) as the better alternative that already solves the same agent-era problems, while the other side pushed back hard on whether the project actually improves on Git's real pain points like submodules and LFS — Oak's creator jumped in to clarify that they chunk every file identically so there's no separate LFS system, which some found compelling but others dismissed as vague. A substantial tangent emerged around whether any VCS can truly track code authorship across file splits and renames, with several people pointing out that `git blame -C` already does this and that the real missing feature is semantic understanding of code structure, not a new storage format. Critics also called out that the blog post was nearly content-free on actual performance comparisons, though the docs and homepage apparently saved the pitch by explaining the lazy-mount and squash-merge design. The e-ink display premium wasn't taken seriously.
Prompt Injection as Role Confusion [comments]
181 points · 98 comments · role-confusion.github.io · 16h ago
The article argues that prompt injection works because LLMs confuse roles (system, user, tool, thought) based on writing style, not actual tags—so text that *sounds* like internal reasoning or a user command gets treated as authoritative regardless of its labeled source, and the authors demonstrate this with probes and a "CoT forgery" attack that boosted jailbreak success from near zero to ~60%. Hacker News largely ran with the security implications: multiple people pointed out that LLMs provide no real security boundaries—"sanitized input" doesn't really exist for them—and one person analogized treating roles as authorization to letting users craft SQL directly, arguing any user-facing product taking real actions based on LLM output is fundamentally hobbled. Several commenters dove into a proposed fix—baking a role embedding directly into each token so the model can't be tricked by style—but immediately flagged the training-data problem: you'd have to label every token by provenance in the training corpus, which is both a massive data-engineering task and raises questions about how that labeling would affect learning patterns. Another thread split on the paper's presentation: some loved the blog-style writeup as a rare readable alternative to impenetrable academic prose, while others pushed back, noting that the formal academic style exists for good reasons—compression into a precise conversation with prior work, accessibility for non-native speakers, and avoiding the "trivial" label from reviewers.
1,700 free online courses from top universities [comments]
175 points · 32 comments · www.openculture.com · 6h ago
The article is a curated directory of 1,700 free online courses from top universities like Yale, MIT, and Stanford, organized by discipline and linking to MOOCs on edX and Coursera. HN immediately pounced on the practical reality: many of those Coursera links are now behind paywalls even for audit access, and the free textbooks section was broken for some users, though others found they opened fine as PDFs. A big tangent spun off around learning techniques—people debated spaced repetition and the "Make It Stick" book, with one side arguing that effective study requires intrinsic motivation and self-teaching, while others insisted simple structured changes like frequent low-stakes quizzes work even for disengaged students. There was also a pointed correction: Stanford's iTunesU courses (like the brilliant Aeneid lectures) have been truncated to a few seconds, and though they're archived on archive.org, the files are blacklisted for download in some regions. Several commenters dismissed the list as ad-heavy, bloated (one reported 6GB RAM usage), and poorly curated—missing CS50 at first glance—while others still find it a useful starting point despite the decay.
Japanese symbols that speak without words [comments]
167 points · 90 comments · arun.is · 12h ago
The article explores Japan's system of purely graphic symbols—from driver marks to transit badges—that convey social obligations without any text, relying on a cultural practice of "reading the air" to trigger consideration. HN quickly latched onto the driver symbols, especially the shoshinsha (beginner) mark, and compared it to New Zealand's L-plate system, where several people reported that the "I'm learning, be kind" sign actually invites aggression rather than patience. The thread then spiraled into a heated debate about turn-signal etiquette in the northeastern US, with one side arguing that signaling there is a "statement" not an "ask," and the other side calling that anti-social behavior normalized as aggression—complete with citations on accident rates. A recurring split emerged: some see Japan's courtesy symbols as a genuine tool for social cohesion, while others countered that in their own countries, similar markers just give hostile drivers a target.
Optocam Zero: a Pi Zero based digital camera made using off the shelf components [comments]
165 points · 43 comments · github.com · 12h ago
The Optocam Zero is a compact, build-it-yourself digital camera based on a Raspberry Pi Zero, meant to recapture the playful feel of toy film cameras. The HN crowd immediately zeroed in on the 22-second boot time, with a lengthy back-and-forth about whether it's fair to compare a cold-booted Linux SBC to an iPhone that’s already running, though several commenters argued that dedicated cameras shouldn’t need to boot at all and pointed out ways to shave that time down to a few seconds. The $100+ bill of materials also gave people pause, with several noting that’s steep for a niche hobby project. A few people dug into the sensor quality and the decision to crop to a square 2592x2592 instead of using the sensor’s full resolution, while others wrestled with the obvious question—why build this when your phone takes better pictures—and landed on answers ranging from “great for kids” to “the joy of making something yourself.”
VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO [comments]
156 points · 53 comments · arxiv.org · 6h ago
The paper claims a 3B-parameter model matches trillion-parameter models on math and coding benchmarks using a novel training recipe. HN immediately pounced on a commenter who tried to generate an SVG of a pelican and got a rectangle—that failure became a proxy argument about whether the model is actually useful or just gaming narrow benchmarks. Several people pushed back hard, clarifying that this is explicitly a closed-world specialist for verifiable tasks where all needed info is in the context, not a general assistant; it doesn't even support tool calling or structured output well. A long, meandering thread broke out comparing the model to a child learning to drive—can a narrow specialist lack so much base knowledge that it becomes useless even for its intended task? Practical reports were mixed: someone solved a nasty ODE that stumped Mathematica, but another person tested it against a corpus of security bugs and it found zero, underscoring the strict boundary between “narrow reasoning core” and “needs knowledge breadth.”
NSF slashes research programs to support new tech initiative, insiders say [comments]
153 points · 91 comments · www.science.org · 15h ago
The linked article wasn't available to this summarizer; from the discussion, it describes the NSF cutting existing research programs to redirect money into a new tech initiative, widely seen as a carve-out for industry. The thread immediately splits along political lines: one camp argues this is a targeted attack on universities and basic research disguised as innovation, pointing to political appointees overriding peer review and predicting the funds will flow to Trump-connected companies. The other side counters that Focused Research Organizations are a legitimate fix for funding gaps that traditional grants can't bridge, letting large-scale, commercially-relevant projects get off the ground. A practical objection cuts through the noise: people working in the domain say tech companies have no interest in engaging with NSF programs anyway, while academics are desperately begging for letters of support from industry—leaving the whole pipeline strained. Several comments dismiss the AI angle entirely, calling it cover for either a political revenge move or outright grift, though one libertarian-leaning poster insists the real path to prosperity is funding more of what made America rich, not protecting academic institutions.
Chevron signs 20-year power agreement with Microsoft for West Texas data center [comments]
142 points · 132 comments · www.chevron.com · 18h ago
Chevron and Microsoft signed a 20-year deal to build a massive natural-gas-fired power plant in West Texas to run a new data center, using GE Vernova and Solar Turbines equipment. The discussion immediately sidetracked into a debate about whether "Solar Turbines" is greenwashing or just a hundred-year-old company name that happened to age poorly; the majority of people in the thread landed on "it's a joke, get over it," though a few pushed back hard on the sarcasm. The more substantive argument was about why Microsoft would choose gas at all—solar and batteries are the cheapest option on the Texas grid right now, and gas producers in West Texas are literally paying people to take their excess supply. Several people with domain knowledge made the case that this is a play for behind-the-meter, on-site generation at a massive scale, where the fuel is practically free and the real bottleneck is pipeline capacity, not power prices.
Will It Mythos? [comments]
137 points · 80 comments · swelljoe.com · 3h ago
The post is a detailed benchmark from an independent researcher comparing how well various AI models—from Opus to DeepSeek to Gemma—can find real, recently-disclosed security vulnerabilities in code, with the goal of testing whether Anthropic's secretive "Mythos" model is truly exceptional or just overhyped. The thread quickly zeroed in on methodology, with people clarifying that the model was only told which file to audit, not what bug to look for, and that the pre-test verification was solely to confirm the bugs were findable. A major split emerged between those who've actually used the "Fable" variant of Mythos—who insist it's a genuine step-change in capability, especially for spatial reasoning and autonomous debugging—and skeptics who suspect it's just a standard model with safety filters disabled, plus some social engineering from Anthropic's confident tone. Several commenters shared vivid personal experiences of Fable solving problems Opus couldn't, like fixing concurrency bugs from logs or handling full 6DOF geometry without losing track, while others pushed back that the "step-change" narrative feels like the same hype cycle that's accompanied every model upgrade since GPT-3. The thread also delivered a sharp verdict on Mistral and Google's Gemini for security work, calling them unfit for purpose due to aggressive safety guardrails that refuse to even look for vulnerabilities.
In praise of memcached [comments]
135 points · 51 comments · jchri.st · 6h ago
The article argues that memcached is often a better choice than Redis for caching because it's simpler, has no persistence, and its client libraries silently handle failures—so you can't accidentally treat it like a database. HN largely agreed with the diagnosis that Redis gets misused as a persistent store, but pushed back hard on the idea that Redis itself is to blame; many commenters pointed out that Redis works fine as a cache if you properly configure eviction policies and disable persistence, and that memcached's "silent failure" is actually a liability for apps that need to know when the cache is down. A vocal minority defended Redis's richer data structures and built-in clustering as legitimate reasons to prefer it, arguing that memcached's bare-bones approach forces you to reimplement those patterns yourself. There was also a brief meta-tangent accusing the article of being AI-written, which most participants dismissed as a lazy reflex. The thread landed on a split: memcached is the safer straitjacket for pure caching, but Redis rewards discipline with more power.
Memory crisis is getting so bad that even retro RAM prices are going to the Moon [comments]
134 points · 43 comments · www.theregister.com · 12h ago
The article explains that the AI-driven memory shortage is now so severe that hardware makers are redesigning products to use older DDR2 and DDR3 chips, sending prices for even those legacy components skyrocketing. The thread quickly zeroed in on the difference between raw memory chips and the modules you'd buy for a PC—someone in e-waste recycling can't sell DDR3 modules at all, while others report the same 16GB DDR3 sticks that cost $5 last year are now going for $60+. A good chunk of the debate turned into a semantic fight over whether DDR3 qualifies as "retro" (it's barely 20 years old, and the PS3 is now retro, so yes), but the real pushback came from people arguing this is just the latest iteration of memory's boom-bust cycle, with hyperscalers hoarding capacity and no incentive for manufacturers to spin up old fabs fast. Several small business owners chipped in to say the price spike has wrecked their margins and forced them to delay products, while others pointed out that if you really need DDR2 for embedded gear or weird old machines, the genuine "retro" stuff isn't what's spiking—it's the modern-but-older generations being cannibalized for new builds.
Job application asked for my SAT scores [comments]
133 points · 325 comments · mrmarket.lol · 11h ago
A software developer’s blog post describes encountering a job application from a YC startup that asked for his SAT scores, prompting a broader look at the history and science of hiring assessments. The HN thread largely pivoted to a debate over whether the SAT is a defensible proxy for intelligence, with one camp arguing that IQ and SAT scores are highly correlated and stable over a lifetime, while the other side hammered on the test’s susceptibility to coaching, wealth-based advantages, and the absurdity of judging a 40-year-old on a score from adolescence. Several people pointed out that this is likely an unspoken age or class filter, since many immigrants never took the SAT and adults can’t easily verify a self-reported number. The conversation also drifted into bashing Canonical’s hiring process, which reportedly uses similar self-assessments and shoddily translated timed IQ tests, and into a weird tangent about requiring cursive writing as an intelligence screen. The consensus was split: some saw it as a refreshing alternative to LeetCode theater, while others dismissed it as lazy, unverifiable, and discriminatory.
British Columbia, Time Zones, and Postgres [comments]
131 points · 96 comments · www.crunchydata.com · 12h ago
The article walks through British Columbia’s move to permanent Pacific Daylight Time and how that breaks future appointments stored as UTC timestamps in Postgres — when tzdata updates, a 10 AM appointment booked before the change suddenly reads as 11 AM local because the UTC anchor shifted. The HN crowd split hard on the fundamental question: should you ever store anything but UTC? A vocal group argued that for events pinned to a wall clock (doctor’s appointments, standups), local time is authoritative and UTC is lossy; they backed the article’s dual-column pattern of storing local time + timezone + a computed UTC column. Others insisted you should only store current Unix timestamps and let battle‑tested libraries handle everything, but got pushed back repeatedly that libraries can’t design your schema and that this exact scenario — a future rule change — is where “just UTC” fails. A long tangential thread dug into the ambiguity of DST fall‑back hours and whether any system can truly solve a social/political problem, with one person noting that even past timestamps stored in UTC can break if tzdata itself had errors and was fixed later.
DHL Set to Transport Goods on New Wind-Powered Cargo Ships [comments]
129 points · 56 comments · www.wsj.com · 17h ago
DHL announced it will start using new wind-powered cargo ships for some of its freight. The HN discussion immediately zeroed in on the scale problem: these ships carry about 415 metric tons, while a typical Maersk vessel carries around 100,000, meaning you'd need roughly 250 of these to replace one container ship. Several people argued the economics don't work—the crew costs for moving tiny loads very slowly are hard to justify, and one person pointed out that kites on container ships for efficiency have been tried for decades and the leading company just went bankrupt. Others pushed back, suggesting the real niche is point-to-point routes between small ports or islands, where you save a week by avoiding the loading/unloading bottlenecks of mega-ports, and that eliminating fuel cost could actually undercut bigger ships on the only metric that matters. A few skeptics called it a vanity project selling an aesthetic, while defenders countered that every technological leap looks like a sucker's bet until it works, and that DHL has a history of actually deploying new cargo tech like electric vans rather than just hyping it.
ytr: YouTube Radio for Emacs [comments]
121 points · 9 comments · github.com · 10h ago
This is `ytr`, a new Emacs package that turns YouTube channels into an audio-only, Winamp-style floating widget streamer, powered by `mpv` and `yt-dlp` under the hood. The HN crowd immediately flagged that Google has been cracking down on `yt-dlp` recently, casting doubt on the package's long-term reliability. A notable split emerged: several people love the "wrong tool for the job" ethos of pushing Emacs into multimedia, while others pushed back, arguing that a dedicated music player is the right approach and that Emacs’ image handling already taught them that lesson. The author, juggling a newborn, jumped in to clarify it’s an undecorated frame not a floating window, and one Doom Emacs user confirmed it works as advertised but noted keyboard shortcuts conflict with Evil mode and Which-key, wanting a proper keybinding setup. There's also a tangentially hilarious thread where parents swap the specific bizarre YouTube channels their infants are obsessed with.
Linux and Secure Boot certificate expiration (2025) [comments]
119 points · 61 comments · lwn.net · 13h ago
The LWN article covers the impending September 2025 expiration of the Microsoft 2011 key used to sign the shim bootloader, which could prevent new Linux installs from booting on Secure Boot machines that lack the 2023 replacement key. The thread quickly turned pragmatic: people shared their own `mokutil` output showing the expiring cert and debated whether firmware actually enforces expiration on already-signed binaries (most argued it doesn't — existing systems keep working, only new shim signatures are blocked). A strong split emerged between those who think it's a manageable firmware-update problem (KEK and db updates hitting 98-99% success via fwupd) and those who see this as proof that ceding boot-chain control to Microsoft was a terrible idea, with several commenters pointing out that Lenovo/Nvidia hardware can't even remove the Microsoft keys. The deeper critique landed hard: Secure Boot adds real attack surface reduction only if you enroll your own keys, but the default Microsoft-signed shim path means Linux is just along for the ride on Windows’ PKI timeline.
Generated 2026-06-23 08:37 UTC
Generated by Sauron from Hacker News discussions and linked articles.