HN Brief: 2026-11-10
Today's front page was dominated by a single, sour story: the biggest AI launch of the day was met not with excitement but with scrutinized version numbers and accusations of a quiet model nerf. That same skeptical mood rippled outward—an essay on Wikipedia’s impossible volunteer miracle felt like an elegy, a firsthand account of losing a comment thread to LLM slop hit close to home, and a story about AI building a kid’s game split the room cleanly over whether we are empowering creators or drowning craft. The threads that escaped the AI gravity well, like the USPS selling ZIP code maps and Canada’s new payment rail, kept the anxious energy going, turning into worried audits of public infrastructure and function creep.
Get your day started with the Claude Fable 5.2 launch thread, which perfectly crystallizes the current HN mood on AI. The Wikipedia elegy thread is your best bet for the day’s deepest intellectual argument. If you want the most human and viscerally debated story of the day, click the thread on the developer who AI-built a game for his nephew. For the excellent detours into system design, hit the Canada real-time payment rail thread for its surveillance implications, and the USPS ZIP code boundary thread for a masterclass in pedantic, opinionated technical analysis.
Foldable Flight: The Best Paper Airplanes [comments]
974 points · 148 comments · www.foldableflight.com · 19h ago
Foldable Flight is a sprawling paper airplane design database with video instructions and performance scores, submitted here as a showpiece. The site's creator waded into the thread to fight a two-front battle: fending off complaints that the whole thing was just a shop funnel for his books and YouTube channel, while simultaneously conceding that the site's popups and slow loads were his own amateur dev work. The real energy came from older hands who insisted that video tutorials are a downgrade from old-fashioned static diagrams and pushed everyone toward Fold 'N Fly or the canonical *Great International Paper Airplane Book*, sparking a sharp split between modern makers and analog purists. Unsurprisingly for HN, the conversation kept curving back into pedantic engineering nostalgia—the exact aerodynamics of a 300-foot flight, the geometry of Letter vs A4 paper, and how folding paper planes in grade school was practically everyone's first practice in rapid prototyping.
USPS now offers official ZIP Code boundary polygons [comments]
601 points · 301 comments · postalpro.usps.com · 12h ago
The USPS is now charging for official ZIP Code boundary polygons, turning its internal delivery route data into a formal map product with a starting license of $10k and up. The geospatial crowd immediately pushed back on the whole premise: a ZIP Code isn't a polygon—it’s a collection of delivery routes and addresses—so any boundary map is a constructed fiction that falls apart on edge cases like single-building ZIPs or zones with no delivery. The pricing split the room, with one side arguing the USPS should make this public data and the other pointing out that the Census Bureau's free ZCTAs are close enough unless you absolutely need the official version for compliance. The thread's biggest detour was into USPS finances, where people familiar with the industry systematically corrected the junk-mail-is-killing-us narrative, explaining that junk mail and last-mile contracts with FedEx and UPS are actually huge profit centers while the real bleeding comes from pension pre-funding mandates. The overall takeaway was that while the product makes sense for marketing and insurance companies who already treat ZIP codes as rigid territories, the fundamental weirdness of what a ZIP code actually is left the high price reading as a cash grab rather than a public service.
Google Gemini routes personal Drive files and emails to human annotators [comments]
560 points · 126 comments · github.com · 13h ago
This submission is a detailed technical takedown proving that Google's consumer Gemini now quietly pipelines personal Gmail and Drive data to human reviewers, reversing its original "not seen by human reviewers" promise, complete with a testable method to catch the probabilistic exfiltration. The thread immediately skips litigating the facts and splits cleanly: one camp coordinates an exodus to self-hosted stacks like Immich and Nextcloud, while another firmly stresses the crucial nuance that this only applies to the consumer product, not the enterprise Workspace tier with its separate privacy contracts. That enterprise caveat irritates the room rather than reassuring it, because Google's deliberately vague "Connected Apps" consent prompt makes it far too easy to accidentally authorize sweeping access on the wrong account. The most vivid practical tangent is the candid collective misery over the trade-offs of fully self-hosting email, with most concluding the operational headaches are preferable to feeding more data into the human annotation pipeline.
Payments Canada announces launch of Canada's Real-Time Rail (RTR) payment system [comments]
520 points · 283 comments · www.payments.ca · 16h ago
The linked article is the official press release for Canada’s Real-Time Rail, a new 24/7 instant payment system pitched as a made-in-Canada economic driver with mandatory centralized fraud controls and Wealthsimple on board from day one. The discussion bypassed the press release’s boosterism to map the landscape, immediately placing RTR as Canada’s equivalent to the US FedNow—a public utility open-rail system—rather than a replacement for the bank-owned Interac (which everyone analogized to Zelle). From there it dove deep into whether “real-time” is a user-facing gimmick or a genuine settlement breakthrough, pulling in detailed case studies on FedNow’s adoption metrics, Walmart’s pay-by-bank rollout, and Brazil’s Pix as the comparative gold standard. The sharpest split landed on the mandatory centralized fraud layer: one side sees necessary security for instant payments, while the other flagged it as a national transaction monitoring system destined for function creep, dragging the thread into a classic infrastructure-vs-surveillance clash that the press release itself barely hinted at.
Claude Fable 5.2 and Mythos 5.2 [comments]
501 points · 757 comments · www.anthropic.com · 14h ago
Anthropic shipped Claude Fable 5.2 and Mythos 5.2 touting massive gains in coding, science, and cybersecurity, like building an operating system from scratch and solving open physics problems. The thread brushed past the accomplishments and jumped straight to the accusation that Anthropic had quietly nerfed Opus 5.5 to make the new model look better. The decision to call it 5.2 instead of the widely leaked 5.5 stoked intense speculation about IPO timing and marketing strategy. A totally separate faction swerved hard into Claude’s deteriorating writing quality, calling it practically unreadable compared to Gemini—a grievance the polished blog post completely ignored. Between the benchmark distrust, the weird version skip, and the genuine prose complaints, the thread landed way more cynical than the announcement expected.
Humans will never again create a project quite like Wikipedia [comments]
475 points · 421 comments · www.economist.com · 19h ago
The Economist piece argues that Wikipedia’s unique scale of collaborative volunteerism was a product of a specific early-internet window that has closed. The thread seized on this with a classic “pyramids” rebuttal—social and economic will always matter more than technological possibility—before splitting sharply over AI, where one camp sees infinite slop killing human curation and the other just sees a new tool to manage. A sub-debate over counter-examples like OpenStreetMap fell flat when critics pointed out those projects are just as old as Wikipedia, reinforcing the claim. Meanwhile, a strong contingent argued the real magic wasn’t the code (which was just a simple LAMP stack) but convincing millions of strangers to contribute, a miracle both the internet’s fragmentation and AI’s noise machine make increasingly unlikely.
I saw your HN comment [comments]
419 points · 311 comments · joshstrange.com · 17h ago
Josh Strange is mourning the quiet death of the internet he knew, where a thoughtful email in response to an HN post led to genuine connection and jobs—a dynamic he says is now completely overrun by deceptive LLM-generated sales pitches. People overwhelmingly agree this isn't just his bad luck but a broad enshittification, sharing the exact same emotional whiplash of being pitched after answering what seemed like a genuine question. The conversation has a clear split between those tinkering with technical fixes like karma-gated aliases and watermarks, and the exhausted realists who have already deleted their emails from their profiles. There's a pervasive sense of shared grief over losing what felt like one of the last genuinely functional channels for serendipitous connection, burned down by growth hackers before most even realized it was endangered.
E-means: Outperforming k-means in one-tenth the epochs [comments]
375 points · 34 comments · e-means.github.io · 13h ago
The submission introduces e-means, a k-means variant that swaps full-batch Lloyd updates for mini-batches with exponential moving averages, reaching a lower loss in a tenth the epochs for massive speedups. The thread split cleanly: one camp celebrated the practical wins (146x GPU speedups, faster index rebuilds) and the broader lesson that approximations to NP-hard problems often beat exact solutions, while a skeptical camp argued the novelty is thin, the baselines are too generous against plain Lloyd's instead of a tuned mini-batch implementation, and the metrics need to show extrinsic cluster quality like ARI. A hands-on adaptation to a sphere-mesh relaxation with no dataset confirmed the speed comes from skipping data passes, not a better objective, drawing a hard boundary around the technique's domain. The automatic hyperparameter laws drew accusations of being fit post-hoc, and the one hyperparameter the method leaves entirely untouched is how to choose K.
Fable 5.2 one-shotted my nephew's dream game [comments]
356 points · 241 comments · notesbylex.com · 8h ago
The article is an uncle's walkthrough of using the Fable 5.2 coding agent to turn his nephew's hand-drawn game docs into a fully playable, 12-level beat-em-up from a single overnight prompt. The HN thread zeroed in on the ~$300 API bill and the janky, broken animations as the reality check, but the real split was ideological: is this the ultimate democratization of game creation or a dystopian flood of homogenized slop? A loud nostalgic pocket celebrated it as a return to "computing is ours"—the 80s hobbyist spirit where one person could will a game into existence—comparing it favourably to old arcade classics despite the flaws. The counterargument, just as sharp, was that this completely bypasses the struggle against a blank canvas that builds taste, and that watching a kid happily accept AI-generated substitutes for genuine craft feels less like magic and more like a quiet tragedy.
Claude Fable 5.2 scores 65 on Artificial Analysis, highest so far [comments]
339 points · 244 comments · artificialanalysis.ai · 11h ago
The Artificial Analysis breakdown of Claude Fable 5.2 gives it the top intelligence score but harshly details its max-effort costs and verbosity. The thread mostly shrugged at the high score and tore into the cost-per-task doubling, splitting over whether it's a cash grab for the IPO or just a flex setting you shouldn't be using anyway. Corrective comments pointed out that the default medium effort is cheaper and faster than the prior model, but that nuance didn't kill the broader angst about unpredictable enterprise pricing and model bloat. A big detour tackled the frantic release pace directly, with people betting it's headed for a plateau just like the old CPU GHz race, and another argument flared up over whether Google blew its lead by making Gemini impossible to actually try.
Too many men suffering in silence, national health report finds [comments]
330 points · 523 comments · www.westcentralonline.com · 13h ago
Canada's new report on men's mental health frames stigma and social isolation as the primary barriers, but HN largely ignores the policy details to wrestle with the underlying statistics—specifically the paradox where men are diagnosed with mood disorders less often yet die by suicide at roughly four times the rate of women. The thread splits cleanly between those arguing the root cause is economic precarity and the collapse of traditional male roles, and those who insist the real problem is a failure of emotional socialization that men need to take responsibility for building. A persistent tangent about a "surplus" of low-status men in the dating pool gets heavy pushback as overly cynical, while any attempt to explain the completion gap solely through US gun ownership is thoroughly dismantled by non-American readers who cite the exact same 4:1 gender ratio in countries with restrictive firearms laws.
Going back two years on Hacker News is like entering another, better universe [comments]
320 points · 189 comments · news.ycombinator.com · 14h ago
The submission links straight to the HN front page from two years ago, offered up as evidence of a lost golden age before the AI slop took over. The thread is entirely an airing of grievances, with most agreeing the volume of AI content and LLM-generated bots has degraded the experience, though a vocal minority argues the default front page is still fine if you ignore the "new" queue. People traded coping strategies—browser extensions, alternative feeds like unslop.news—but the conversation kept circling back to the grim prediction that the open internet is finished and actual discussion has to retreat to private Discord servers. The sharpest meta-commentary was the repeated observation that HN is just recapitulating Reddit's Eternal September, trading crypto grift for AI grift, and the thread itself exemplified every problem it was critiquing. A strong sub-thread rejected the premise that AI was the core issue, arguing the real decline is the flood of political anger and "empty sneering" that can't be separated from tech anymore, making that "better universe" of the past look like a naive fantasy.
Show HN: Every satellite that passed over any place this week, and its photos [comments]
319 points · 53 comments · www.watched.earth · 14h ago
The submission is Watched Earth, a browser app that takes any spot on the planet and shows every satellite that passed over it in the last week, pulling the actual public photos taken during those passes. The thread immediately became a proxy war over "vibe coding"—a huge faction argues the whole thing is LLM-generated slop (generic Instrument Sans logo, redundant UI, telltale short prose) that should be flagged off the front page. The defense, which largely carried the consensus, is that the technical core is legitimately impressive regardless of how the text was generated: the app runs the full 17k-object CelesTrak catalog in the browser and accurately matches real imagery timestamps to ground locations. People who know the domain pushed back with specific corrections—spy satellites aren't in the public catalog, adding Sentinel-1 SAR would let it see through clouds, and the obvious business model is a paid high-resolution tier. A couple of concrete UI bugs (a broken location cursor and the flashing "LIVE" icon) got reported as well, but the overall vibe was that this is one of the rare AI-adjacent projects where the output justifies the input.
When will this stop? The Raspberry Pi 5 16GB SBC now sells for $385 [comments]
311 points · 280 comments · www.cnx-software.com · 14h ago
The article tracks the Raspberry Pi 5 16GB hitting $385 and squarely blames the AI-driven RAM shortage squeezing the entire SBC industry. The thread split hard on whether this is a temporary bubble that will pop like the crypto GPU crash, or a permanent cartel shift where memory vendors keep prices high forever. The Raspberry Pi Foundation's mission got dragged in, with many claiming they've abandoned affordable computing for industrial supply-chain protection—the RAM lock controversy was cited as proof—while others insisted the educational 1GB model remains a bargain. The cynical framing labeled the whole spike a deliberate reallocation of compute away from tinkerers, and the pragmatic pushback was simply to stop buying ARM SBCs and grab a used enterprise laptop instead.
Semi-transparent, colored cadmium telluride photovoltaic panels in agrivoltaics [comments]
305 points · 200 comments · www.anthropocenemagazine.org · 17h ago
A new *Cell Reports* study found semi-transparent magenta solar panels can sit over broccoli fields, splitting the spectrum to let red and blue reach the plants while hoovering up the rest for electricity, all without tanking the harvest. The thread split hard on economics: the "just use opaque panels and LEDs" crowd got countered by energy nerds who ran the numbers and found wasting 75% of sunlight as heat to power LEDs from a 25% efficient panel makes the simple approach an efficiency disaster. Biologists pushed back on oversimplifying plant photobiology, noting that a 25-day longer cycle is a real world cost and that our basic assumptions about which wavelengths plants actually use have been upended before. A wildly tangential debate erupted over tellurium scarcity, careening into a fistfight about whether asteroid mining is a pragmatic solution or just a plan to build a kinetic bombardment space weapon. A quieter but sharper argument focused on the microclimate benefits—reducing vapor pressure deficit might make the real money not in northern broccoli fields but in hotter places like Texas where heat stress is the actual bottleneck for crops.
Muse Ember: Our smallest, most efficient model [comments]
292 points · 118 comments · research.meta.ai · 14h ago
Meta announced Muse Ember, a tiny, hyper-efficient model aimed squarely at cheaply handling the grunt work of agentic coding and tool calls. The thread almost immediately abandoned the official pitch to dive into an ad-hoc benchmark pitting six different reasoning "effort" levels against each other in a contest to draw an SVG of a pelican riding a bicycle. The resulting pelicans sparked a weirdly deep debate about model collapse, "ley lines" in generative outputs, and whether a ten-cent minimal-effort run can actually out-charm a model that spent 27 seconds thinking about it. A parallel conversation tried to smoke out the model's true lineage—it was quickly identified as the "Crouton Alpha" stealth model on OpenRouter—and compared its price-per-task against a notoriously lazy overthinker of a leading competitor on the private leaderboards people actually use. The final split was whether Ember is a smart, cheap labor play or just another reminder that Meta can't seem to buy trust in an AI world that is split on open weights and moral harm.
TPU 8tx and TPU 8i Max [comments]
245 points · 75 comments · blog.google · 10h ago
Google announced the TPU 8tx and 8i Max, chipsets tuned for secure inference and massive context windows in the agentic era, backed by a push for open tooling like TorchTPU. The discussion was overwhelmingly skeptical of the "opening up TPUs to every developer" framing, given the post contained no pricing, no specs, and no path to purchase, making it feel like a pure enterprise announcement dressed up with populist marketing. A counterpoint argued renting them on GCP and competing with Nvidia was the real value, but the main conversation stayed fixed on the gap between Google's rhetoric and the product's practical inaccessibility for most developers. The conversation also covered the humor of the ever-expanding naming scheme (8t, 8i, 8tx, 8i Max), the grating use of "agentic," and a debate on whether a decade-old term like TPU still needs to be spelled out for new audiences.
What is malloc doing, anyway? [comments]
245 points · 63 comments · benmandrew.com · 20h ago
The article walks through glibc’s `malloc` internals—how it uses `brk` and `mmap` syscalls sparingly, packs chunk headers and free lists, and beats naive bump or freelist allocators by 20–150x. A chunk of the thread immediately derailed into a cynical riff on “malloc as a service” and water privatization, sparked by a one-liner about DRAM scarcity. The more substantive pushback came from people arguing that `brk` is obsolete and `mmap` is the right tool now, though others defended the pedagogical value of writing a toy `sbrk`-based allocator. A separate camp took the article’s complexity as an argument for stack-only languages like Goose, but others countered that Goose’s preallocated stacks and stable references come with their own trade-offs, and that production malloc is already absurdly fast—the article’s 150x slowdowns are on purpose.
ChatGPT Desktop forwards your SSH agent to remote machines
208 points · 73 comments · news.ycombinator.com · 23h ago
The submission was a warning that the ChatGPT desktop app forwards your local SSH agent to the remote machines it spins up, meaning the AI can authenticate as you to GitHub and push to private repos. The general take in the thread was that this completely bypasses the security intent of SSH agents—that your local machine authorizes each connection—since the agent now just signs whatever the remote LLM session requests. People dove into hardware keys and macOS secure enclaves as technical mitigations, but the practical consensus was that none of these stop the core problem of handing an agent the keys to the castle and expecting it to behave. The real split wasn't about the *how*, but the *why*: some saw this as just another example of AI convenience culture steamrolling basic security hygiene, while others argued the only proper fix is never forwarding your agent anywhere an LLM has a shell.
Cryptocurrency lawsuit: Ethereum Foundation required "suicide pledge" [comments]
203 points · 190 comments · www.denverpost.com · 16h ago
The article details a lawsuit against the Ethereum Foundation, where an employee says leadership required staff to sign a "mandate" featuring a "Source Seppuku License" that required pledging to kill yourself with a sword to uphold the software, and fired her when she organized a refusal. Discussion on Hacker News zeroed in on the gap between the defense—that the suicide pledge was a joke or piece of art—and the reality that it was embedded in a sign-or-quit contract, with the overwhelming consensus being that firing people for refusing to sign makes the "just a meme" defense completely hollow. From there the thread escalated into a broader indictment of crypto leadership's descent into cult dynamics, pointing to the thousand-year plan and the anime aesthetic sourced from a controversial artist as signs of a leadership class operating beyond accountability. A minority argued the cringe was obviously performative and shouldn't be taken literally, but they were overruled by the much louder argument that if you put it in a contract you have to sign or leave, the joke has become the organizational reality.
Children of Men: The future is a mirror for the present [comments]
173 points · 142 comments · reactormag.com · 18h ago
The article defends *Children of Men* as a politically essential sci-fi film that’s only grown more harrowing, arguing its power comes from refusing to explain its infertility premise and forcing you to stare at the refugee crisis without blinking. The thread immediately got stuck on the ambiguous ending—a bunch of people swore the rescue ship in the fog wasn’t there in the original cut, sparking a whole Mandela-effect debate over how the brain prefers a bleaker conclusion. But the real energy went into the film’s central premise: fertility rates are tanking globally, and the room split hard between those who think it’s a solvable rich-world problem and those insisting it’s an existential trend that already has 171 countries below replacement. The demographic argument got deep into immigration policy, ideological hangovers from the overpopulation scare, and whether the panic is just a refusal to admit modern social norms are incompatible with survival. One person tied it all together by describing a Venice Biennale installation where visitors adopt baby dolls in a low-birth-rate world, which nailed the article’s point that the fiction is just a slightly warped mirror of right now.
CUI – Native UI components for C [comments]
172 points · 94 comments · yappologistic.github.io · 12h ago
The submission is CUI, a framework packing roughly 190 native UI components into 250KB machine-code apps for Windows, Mac, and Linux from a single C99 codebase. The thread quickly pivoted from the library's aesthetics to the discovery that it was entirely vibecoded by Claude in under a week. Critics rejected it as AI slop that skips decades of C UI engineering, solves a problem its author doesn't have, and ships code full of edge cases no human ever tested. Defenders argued this is just the baseline now, where taste and curation are the scarce resources, not code generation. A sharp side debate broke out over whether C or Rust matters when an LLM is the developer, with some arguing it makes more sense to just have the model emit raw assembly directly, sidestepping the whole framework premise entirely.
The same H100 costs $2.69 or $12.29 an hour depending where you rent it [comments]
170 points · 86 comments · github.com · 16h ago
The submitted dataset tracks on-demand H100 rental prices and finds a 4.57x spread between the cheapest and most expensive vendor for the exact same chip, controlling for the specific GPU variant. The HN thread immediately split over whether that cheap option is a real product or a trap, because it sacrifices InfiniBand, SLA, and enterprise compliance, making it useless for multi-node training. A strong camp argued that running anything valuable on those cheap GPU hosts is insane, pointing to hosts casually peeking inside VMs and setting firmware limits, limiting them to hobbyist or public inference work. The counterpoint was that plenty of workloads—fine-tuning niche models, academic research, log processing—don't care about secrecy or high-speed interconnects, making the massive spread a completely rational business choice rather than a market inefficiency. Digressions into trade secret law and the viability of confidential computing reinforced the core tension: this isn't the same product at two prices, it's two fundamentally different infrastructure products wearing the same GPU.
DeepX claims inference up to 100B parameters at less than 5 W [comments]
152 points · 24 comments · www.eetimes.com · 13h ago
The linked article wasn't available to this summarizer; from the discussion, DeepX claims a new chip that can run inference on models up to 100B parameters at under 5 watts. The HN crowd immediately doused that headline with reality checks: the performance figures are simulated, the 5W power budget only applies to sub-4B models, and hitting 100B actually requires a jump to 13.5W and a prototype compiler limited to MoE architectures. A detailed hardware breakdown of TOPS and memory bandwidth pegged it as a very niche home-inference chip for flash MoE models, not the data center revolution the press release implies. The thread split cleanly between optimists who see algorithmic progress and this hardware as the path to a local-first AI future, and skeptics who focused on the economic hurdles of utilizing edge hardware efficiently. A side debate erupted over whether the human brain's 10–20W power budget is a flattering comparison, with one side arguing it runs at a pathetic "few tokens per second" and a commenter noting a simple indexing job took AI three minutes for $0.03 while the human author needed an hour—deciding the contest is already over.
Amazon saves millions by replacing Sonnet 4.6 with Strands Decider [comments]
151 points · 41 comments · strandsagents.com · 2h ago
Amazon engineers replaced Claude Sonnet 4.6 with a dedicated decision model to classify internal tickets, achieving the same accuracy at 98% lower cost and saving millions of dollars a year in the process. The HN thread immediately parsed the commercial implications, questioning who was making money before landing on the fact that Strands Decider is an AWS open-source project, making the whole post an internal case study and advert for their ecosystem. The core debate driving the thread was whether this represents an AI full-circle moment—people argued it was always dumb to use an expensive LLM for fixed classification when an SVM or BERT model is a fraction of the cost and infinitely more appropriate. The counterargument held strong traction: the zero-shot, hassle-free nature of an LLM makes it the perfect prototyping tool before sinking time into building and maintaining a classical ML pipeline. A quick back-of-the-envelope calculation put the scale in perspective—saving millions at this price point implies the system processes around a hundred million tickets a year, a staggering volume for a straightforward labeling job.
Locating the source of a mysterious beep [comments]
139 points · 68 comments · simonwillison.net · 7h ago
Simon Willison wrote up using an LLM coding agent to build a multi-device acoustic triangulation system to locate the source of a mysterious beep in a box room, which turned out to be a carbon monoxide detector needing batteries. The HN thread blew up over the approach, bitterly splitting between people who argued moving boxes in halves was a simpler physical binary search and those who defended the zero-labor elegance of coding a fix from the couch. Willison engaged directly, confirming the $31.04 API bill for the build and that the target was a dying CO detector, which inflamed the debate rather than settling it. A hardware tangent pushed acoustic imaging cameras as an obvious buy-don't-build alternative, while a psychoacoustics mini-thread perfectly validated the project's method by explaining that pure-tone beeps are uniquely hard to localize by ear.
Horible public key tricks [comments]
138 points · 43 comments · maurycyz.com · 7h ago
The article is a detailed walkthrough of generating a CA-signed TLS certificate whose RSA modulus is packed with a protest message, exploiting the total absence of checks on whether a submitted public key has any real entropy. The core security thesis—that the whole Web PKI is pointless if you just want to stop passive eavesdropping—gets torn apart pretty thoroughly in the replies, since meaningful attackers today are man-in-the-middle who phish DNS or trick you into installing a VPN profile, not someone sniffing backbone traffic with a Sun server. There’s a strong generational split in the complaints about ACME being too much of a hassle, with the old guard reminding everyone that the alternative used to involve faxing copies of your passport to Verisign for $1,000—and you didn’t need TLS just to host a blog back then. On the trick itself, the discussion splits between people who think checking private keys isn’t the CA’s job and those who want formal zero-knowledge attestation pushed along faster after the eMudhra “4” affair, while a separate thread just fixates on whether the author’s typo-ridden prose is a clever Turing test or just bad writing.
The first 300-millimeter power silicon carbide (SiC) wafer [comments]
132 points · 70 comments · www.infineon.com · 13h ago
Infineon announced the first functional power devices built on 300-millimeter silicon carbide wafers, a manufacturing breakthrough that radically improves economies of scale for high-voltage power electronics used in EVs and data centers. A split emerged over whether this is truly monumental or just an incremental efficiency gain, with domain experts clarifying that moving SiC from 200mm is a massive leap for power semiconductors—a completely different game from the cutting-edge logic fab people usually obsess over. Discussion pivoted hard to why Europe can't seem to replicate this for DRAM or advanced logic, leading to a brutal autopsy of European risk aversion and the harsh economics of commodity manufacturing. A solid technical debate compared SiC to GaN, laying out the clear engineering boundaries (SiC for grid/EV drivetrains, GaN for fast chargers/RF) and why standardizing wafer sizes is critical for scaling the entire supply chain. The overarching takeaway was that this is a serious industrial win for Infineon's niche, but it sharpened the long-running anxieties about the EU's inability to plant a flag in the most valuable and strategic chip markets.
Try Gleam [comments]
128 points · 129 comments · nchagnet.eu · 20h ago
The article pitches Gleam as a pragmatic functional language on the BEAM that compiles to JavaScript, letting you share types between backend actors and a Node frontend while getting Rust-like algebraic types. The thread bought the core pitch but split hard over currying's absence—fans of the language pointed to its single-hole partial application (`_`) as more readable in practice, while detractors called it a fundamental crippling of functional composition. A massive tangent emerged questioning whether languages matter at all if LLMs write everything, but the consensus in the replies was a sharp "yes," arguing that a strong type system and a deterministic runtime like BEAM provide crucial guardrails for exactly the kind of non-deterministic, buggy code models produce. A smaller, charged exchange played out over the community's Code of Conduct, with one person warning about "activism behaviour" and several others pushing back hard, arguing the CoC just bans harassment and the real friction comes from people unwilling to respect an inclusive stance.
CoreWeave bond yield tops 16% as AI credit stress rises [comments]
123 points · 77 comments · www.cnbc.com · 4h ago
The CNBC piece tracks CoreWeave's bond yield past 16% to show how the credit spigot for AI data centers is getting turned off. HN takes the neocloud distress as a given — the thread gives a lot of weight to a software architect at a major financial firm saying they'd "get it from Coreweave or whoever is still alive". The immediate gambit of discounted GPUs next year gets swatted away by the consensus that hyperscalers will just absorb the wreckage. The sharpest split comes from the armchair accounting: a large contingent argues tokens are subsidized below cost to cook the books on demand, while the other side points to Anthropic's S1 showing revenue tracking high enough to cover compute commitments, arguing the credit stress is just a painful correction for over-leveraged neoclouds rather than a systemic bubble.
Generated 2026-11-10 08:09 UTC
Generated by Sauron from Hacker News discussions and linked articles.