HN Brief: 2026-10-17
Today's HN had Anthropic firmly in the spotlight — Claude solved a 25-year ECDSA mystery, published its own code review philosophy, and got a junk credit rating — but the bigger throughline was a growing skepticism about AI's actual reach. A major labor benchmark study found 90% of US working hours unmeasured by existing AI tests, and Postgres 19's delay after AI reviewers flooded maintainers with bugs sparked a debate about whether the tools are helping or drowning teams. Meanwhile, nostalgia and hardware tinkering ran strong: software rendering vs. GPUs, a jailbroken PS5 running llama.cpp, and a 37-year-old OpenBSD bug that still works.
The threads most worth clicking: "Claude reveals ECDSA curve origins" for the satisfying end to a 25-year conspiracy theory — but also the uncomfortable question of whether Anthropic's hidden reasoning is a feature or a problem. "90% of US working hours are spent on tasks not covered well by AI benchmarks" for the sharp debate about whether AI will only automate rich-people jobs while leaving the rest to humans. "Show HN: Gemini 4 Argon plays StarCraft" for the proxy-rush exploit that won the benchmark and the side drama of a neural-network bot going 1000-0. And "Postgres 19 delayed after AI reviewers find more bugs than humans can fix" for the doom-loop scenario that pits genuine defect detection against maintainer burnout.
Claude reveals ECDSA curve origins [comments]
841 points · 315 comments · www.anthropic.com · 13h ago
After 25 years of conspiracy theories about NSA backdoors in NIST's elliptic curve standards, Anthropic published that Claude solved the mystery by figuring out the seeds were just outputs of the weak C library PRNG `lrand48` seeded with Unix timestamps from 1997—not hashed passphrases as had been rumored. The HN crowd is mostly celebrating that the "tinfoil hats" can finally come off, with many pointing out that this actually *increases* confidence the curves weren't backdoored, since anyone can now reproduce the generation process. But a significant split emerged: some commenters insist the timestamps themselves could still have been cherry-picked, and there's a lot of grumbling about the irony that Anthropic used Claude's hidden reasoning to crack this while regular users get blocked by safety filters for even asking basic cryptography questions. Others are already calling for Ethereum and Bitcoin to finally adopt P-256 opcodes now that the provenance is settled, while a separate faction is more worried about the "strip mining" of scientific problems by frontier labs with capabilities locked away from the public.
90% of US working hours are spent on tasks not covered well by AI benchmarks [comments]
487 points · 408 comments · www.mercor.com · 14h ago
The article from Mercor maps 29 AI benchmarks against 923 O*NET occupations, finding that 90% of American working hours fall into tasks with little or no benchmark coverage—meaning we have almost no way to gauge how well models perform on work people actually do. HN immediately pushed back on the implied narrative that AI is coming for everyone’s job, with software engineers pointing out that 80% of their day is stakeholder politics, expectation management, and bugging overseas contractors, not coding, and that these human-driven friction points have no tight feedback loop for training agents. Several commenters argued the real productivity bottleneck is organizational dysfunction and ego, not technology, and that faster coding only compresses the timeline for bad business decisions to hit production. There was a sharp split between those who see the benchmark gap as proof that AI will only automate narrow, high-ROI tasks (leaving the rest to humans) and those who mocked the idea that “cleaning bathrooms” represents some noble redirection of labor—pointing out that the benchmarks focus on the best-paid jobs because that’s where the profit is, not because those jobs are most replaceable.
How the Claude Code team does code review [comments]
326 points · 303 comments · claude.dev · 12h ago
The Claude Code team at Anthropic explains how they’ve scaled code review to handle AI-generated code at volume, moving most review to automated agents and aiming for at most one human per PR without increasing reverts. The HN thread was deeply skeptical: plenty of people pointed out that the Claude Code repo has thousands of open issues, many of them AI-spammed regressions and locked threads, and called the 30,000 changes-per-day number a meaningless churn metric that explains the product’s bugginess. A split emerged between those who see this as inevitable acceleration and those who argue it’s irresponsible—human reviewers still catch things agents miss, and dismissing security findings as false positives with no AppSec sign-off drew sharp pushback. The thread also dug into practical frustrations with Claude Code’s tooling, like skills that bloat context windows and loop agents into infinite revision cycles, which undercuts the blog’s polished narrative of efficient automation.
I love software rendering [comments]
318 points · 201 comments · www.jamesfmackenzie.com · 19h ago
The article argues that software-rendered DOS games like Quake look better than hardware-accelerated versions because the CPU-driven pixel-by-pixel drawing avoids the blurry bilinear filtering and palette-based lighting that gave early 3D cards their smoothed-over look. The HN thread immediately pushed back on the framing, arguing that what the author actually likes is just “turning off bilinear filtering” — a setting you can toggle in modern GPU drivers, not something unique to software rendering. A deep technical argument unfolded about whether “software rendering” even exists on modern operating systems, since every frame ultimately goes through a GPU compositor, and several people pointed out that modern engines like Unreal’s Nanite technically do software rasterization on the GPU itself, blurring the old distinction entirely. A separate tangent dealt with CRT monitors of the era being sharper than TVs but still softer than today’s LCDs, so the “pixelated crispness” people remember is partly a modern artifact. The consensus split between those who agree the nostalgic look is real and replicable via GPU settings, and those who think the article conflates a historical limitation with an aesthetic choice that has nothing to do with CPU versus GPU.
Waymo in Toronto [comments]
309 points · 411 comments · waymo.com · 17h ago
Waymo announced it's coming to Toronto with a 2028 launch target, starting manual mapping and testing now. The HN discussion immediately zeroed in on winter—how falling snow blinds lidar, how lane lines vanish under slush, and how Toronto drivers just follow the ruts left by the car ahead, with people swapping first-hand accounts of Calgary and Minneapolis testing and links to Waymo's own all-weather driver blog post. A parallel thread erupted over Doug Ford's bike lane removals, with locals arguing they'd rather share the road with robotaxis than with human drivers, while others insisted more steel boxes on the streets is the wrong direction entirely. A smaller contingent questioned whether Canadian money should fund a US startup, and a few commenters dismissed the whole expansion as irrelevant given the province's war on cycling infrastructure.
Apple ML Research: Muse Spark Open with MLX on Mac Studio [comments]
304 points · 157 comments · machinelearning.apple.com · 18h ago
Apple’s ML research team published benchmarks showing that a cluster of four Mac Studios can run Meta’s 800B-parameter Muse Spark Open model at 36 tokens per second, using MLX and RDMA over Thunderbolt. The HN thread immediately fixated on the price tag—a 512GB Mac Studio starts near $10k, so four of them land well north of $40k, with commenters arguing that even heavy OpenRouter usage would take years to burn through that much cash, making the whole demo more academic than practical. Others were annoyed that Apple’s press release heavily featured a Meta model without crediting Meta outside of code references, calling it a weird partnership given the companies’ historically frosty relationship. There was also technical pushback on the benchmarks—the article compared Muse Spark in 4-bit NVFP4 against a Qwen model running at native BF16, which several people called an unfair comparison, since nobody running locally would choose that full-precision quantization for Qwen.
Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console [comments]
273 points · 67 comments · github.com · 13h ago
This project ports llama.cpp to a jailbroken PS5, using the console’s RDNA 2 GPU via Vulkan to serve GGUF models to any device on your network with OpenAI/Anthropic-compatible APIs. The thread quickly zeroed in on the hard constraint — 11.4 GB of shared memory means you’re stuck running heavily quantized models like Qwen 3.8 27B at 2-bit, which several people who actually run these models warned is noticeably degraded and nowhere near Opus 4.8 quality, even if the raw model is decent. A big chunk of discussion pivoted to the economics and legality: people argued the PS5 is sold at a profit, so Sony probably won’t care about this use case, but others pointed to the BC-250 mining board (a rejected PS5 chip you can grab for ~$200) as a cheaper alternative that wouldn’t risk a DMCA takedown of the GitHub repo. There was also the inevitable nostalgia for PS3 compute clusters and jokes about LLM-powered NPCs in games, but the practical consensus was that unless you already have a jailbroken PS5 sitting around, you’d get far better performance and less hassle from a used RTX 3060.
You could have invented transformers from memory [comments]
271 points · 65 comments · alexwang.ai · 13h ago
The article presents transformer components—linear layers, MLPs, and self-attention—as different kinds of associative memory stores, analyzing each layer’s capacity by tracking how much interference builds up when too many keys are stored. The HN thread quickly turned into a meta-debate about whether math notation is appropriate for a “software development site,” with one commenter arguing the post would be more accessible as pseudocode and a worked example, and getting downvoted and pushed back hard for it. A long chain of replies pushed back: defenders said HN is for intellectually curious people, not just programmers, that the math is approachable, and that asking the author to drop notation misunderstands both the article’s value and the breadth of the community. The underlying split was between readers who wanted a programmer-friendly treatment versus those who see mathematical exposition as perfectly fine—and the thread mostly ended up arguing about the site’s culture rather than the article’s technical claims.
Anthropic assigned highest junk credit rating in June [comments]
247 points · 75 comments · www.ft.com · 13h ago
A Financial Times article reports that Anthropic was assigned a junk credit rating by the small agency KBRA in June, with bankers lobbying for investment-grade status ahead of a potential IPO. The thread immediately zeroed in on the fact that Anthropic went to an obscure rater rather than the big three, which many read as a sign that larger agencies balked or demanded more data—Anthropic reportedly withheld key metrics like free cash flow and EBITDA, forcing KBRA to estimate. Comparisons to 2008’s credit-rating manipulation were inevitable: the deeper worry is that if Anthropic eventually gets an investment-grade stamp, pension funds and life insurers (i.e., your mom and dad’s retirement money) would be allowed to buy the debt, repeating the cycle of privatized gains and socialized losses. A strong counterpoint argued that junk is junk, institutional buyers know the risk, and that the real story is the decades-long shift where companies stay private through their growth phase and only IPO to cash out, leaving retail investors holding the bag. Others dismissed the whole thing as a nothingburger, noting that KBRA is not S&P or Moody’s and that hyperscaler datacenter spending is so profitable that interest rates barely matter—though the same commenters admitted the AI lab’s own solvency is a different question entirely.
JPEG XL vs. AVIF for web use-cases [comments]
245 points · 174 comments · jakearchibald.com · 15h ago
Jake Archibald's deep dive compared JPEG XL and AVIF for web images, and his benchmarks conclude AVIF almost always beats JPEG XL on file size for lossy photos, screenshots, and images with transparency, often by a significant margin. The HN discussion pushed back on a few fronts: several people argued that the container-based nature of AVIF (it's just a wrapper for AV1 video frames) means future codecs like AV2 can slot in without sites having to redo their image pipeline, while JPEG XL can't do that—others countered that the container doesn't help if browsers don't implement the new codec, and that AVIF's container is needlessly complex. There was also a strong defense of JPEG XL's lossless recompression of existing JPEG files, which can shave 9–20% off file sizes with no further quality loss, a unique feature for archives that some consider more valuable than raw space savings for new uploads. A few commenters noted the article's quality metric (SSIMULACRA2) may favor AVIF at lower qualities, and that at very high quality (85+ SSIMULACRA2) JPEG XL can match or slightly beat AVIF, especially with jixel's slowest settings—so the "best format" depends heavily on quality target and use case.
Show HN: Gemini 4 Argon plays StarCraft [comments]
216 points · 97 comments · starskirmish.com · 17h ago
The post presents StarSkirmish Bench v0.2, a benchmark where LLMs get one hour to write a functional StarCraft: Brood War bot in C++, then compete against human-coded bots and each other. The headline result was that GPT-6 Astra took first place, but the thread quickly zeroed in on how it won—by exploiting a proxy 2-gate Zealot rush that shortens games into coin flips, a clever but narrow strategy that some argued exposes a flaw in the benchmark's design. A huge chunk of the discussion pivoted to Pluto, an entirely separate neural-network bot that casually went 1000–0 against Astra; people dug into how Pluto's lack of memory outside line-of-sight and its perfect micro create a different kind of challenge, one that top human players can now work around once they realize it's not human. There was also serious pushback on whether the benchmark measures anything coherent—some argued it's a useful end-to-end test of "agenticness," while others complained that the one-hour limit and reliance on memorized open-source bot code make results hard to interpret. A tangent worth noting: Dan Gant, author of one of the top human-written bots (PurpleWave), showed up with a detailed breakdown of how Astra actually plays, and one commenter pointed out that Gemini 4 Argon showed wild run-to-run variance (18-point spread) compared to competitors, which nobody had a clean explanation for.
I never authorized an integration that scans my secrets [comments]
215 points · 182 comments · github.com · 11h ago
Home Assistant was found to have an integration that silently sends truncated SHA-1 hashes of user passwords to Have I Been Pwned, with no opt-in and poor documentation that an opt-out flag even existed. The GitHub issue thread blew up not just over the privacy violation—people felt their local automation hub was phoning home without permission—but over the lead developer's abrasive, sarcastic replies, including telling a user to "just fork it bro." The comments split sharply: many argued this is a fundamental betrayal of trust for a project that markets itself on privacy, while others pointed out that HIBP's k-anonymity model makes the actual data exposure minimal, and that the real problem was the developer's tone and refusal to engage constructively. Things got worse when people dug up a five-year-old NixOS packaging dispute involving the same developer, painting a pattern of hostile maintainer behavior. A follow-up apology from the dev admitted the dismissiveness was wrong, confirmed the feature could be disabled, and promised better docs—but by then the damage to community trust was done.
eBay Restricts Listings of Radio Equipment, but Exempts Ham-Band-Only Gear [comments]
214 points · 173 comments · www.arrl.org · 13h ago
eBay is cracking down on radio equipment listings, requiring FCC IDs for anything that can transmit outside ham bands—though gear that’s strictly amateur-only gets a pass. The thread quickly zeroes in on the loopholes: virtually all Chinese handhelds like Baofengs are trivially unlockable via a key combo or alternative firmware, so the restriction is more about paperwork than actually locking down capability. A lot of the pushback focuses on how absurdly broad this is—test gear, SDRs, ESP32s, and even Meshtastic nodes could all get swept up, though people point out the policy only applies to new items and used gear is exempt. The real split is between those who see this as reasonable enforcement against unlicensed users stepping on public-safety frequencies and those who argue it’s performative regulation that won’t stop anyone determined, while hams already have their own swap forums that don’t care about eBay’s rules.
US Rep Announces Congressional Legislation to Protect Video Game Ownership [comments]
205 points · 113 comments · ritchietorres.house.gov · 16h ago
Representative Ritchie Torres introduced the Power To The Players Act, a bill that would make the FTC force game companies to sell games as owned property rather than revocable licenses and require an offline mode before any server shutdown. The thread quickly split into two camps: one arguing this is a smart, narrow first step toward broader consumer rights (including SaaS and other software), the other pointing out that Torres represents a high-poverty district and that “gamer rights” feel like a distraction from more urgent crises—though plenty of replies shot back that politicians can work on multiple issues at once. A deep technical argument broke out over the offline-mode mandate: while single-player games are fine, MMOs and live-service titles can’t plausibly work offline, so commenters pushed for mandatory server binary or documentation release, only to be met with objections that companies won’t hand over proprietary code (sometimes shared across multiple games) and that the cost and liability would discourage investment in the genre entirely. Most of the discussion ended up skeptical of the bill’s actual chances—the sponsor has a low track record of passing legislation—but broadly supportive of the principle, even as the practical details of “owning” a digital game remain messy.
Butlerian Jihad [comments]
200 points · 107 comments · dune.fandom.com · 5h ago
The submission is a Dune wiki entry on the Butlerian Jihad—the fictional crusade against thinking machines that sets up the no-AI rule in Frank Herbert's universe. HN immediately latched onto it as prophecy for the current AI boom, with one person reporting a real-world "BJ" group in Phoenix that uses angle-grinders to destroy QR codes on rental e-scooters, rendering them useless and blocking sidewalks. The comment thread split hard on whether Herbert's Jihad is a shallow plot device to justify Mentats and never gets explored, or whether it’s a succinct and devastating critique of rentierism and machine-dependence that feeds the entire series' themes. People also dragged in Iain M. Banks’ Culture (where Minds rule and humanoids are “pampered pets”) and Warhammer 40k’s Men of Iron revolt, arguing that both universes take the same premise in different directions—while noting that the real Imperium of Man runs on hypocrisy and secretly worships abominable intelligences anyway.
Postgres 19 delayed after AI reviewers find more bugs than humans can fix [comments]
197 points · 77 comments · www.theregister.com · 16h ago
The Register reports that Postgres 19 has been delayed because AI code reviewers surfaced more bugs than the human team can fix. The thread immediately split over whether this is a net positive or a looming disaster: some developers argued the AI is finding real, subtle bugs that would have otherwise escaped into production—edge cases, race conditions, and design tradeoffs that force hard conversations—while others countered that many of these "bugs" are actually legitimate design decisions or false positives that create an unmanageable triage burden for maintainers already drowning in slop. A recurring theme was the "doom loop" where an AI reviewer flags bugs, an AI coder overengineers fixes that introduce new bugs, and the cycle repeats until a human steps in, with several commenters noting that even top-tier models like Opus 5.5 and Fable-class agents still produce bloated, context-polluted code that requires constant human judgment. There was also a pointed debate about whether the project simply became aware of pre-existing bugs at an unprecedented rate, rather than there being a genuine increase in defect density, and whether the "bug count" metric itself becomes meaningless when AI can invent hypothetical issues that no human would care about.
Desire to Move Permanently to U.S. at New Low [comments]
191 points · 197 comments · news.gallup.com · 15h ago
The Gallup data shows the U.S. remains the world’s most desired destination for permanent migration, but at 15% it’s hit a nearly two-decade low, with the sharpest drops in Latin America and among would-be migrants who previously had the U.S. at the top of their list. The thread immediately split into two camps: one side argued the trend proves the U.S. is becoming less attractive due to politics, cost of living, and a deteriorating civic fabric, while the other pointed out the U.S. still leads by a wide margin and that “desire to migrate” isn’t the same as intent—many people want to come for money, not for the country itself. Several commenters pushed back hard on the idea that the U.S. is a “shithole,” noting that if the U.S. is declining, then Canada, France, and Germany are in even worse shape by the same metric, and that the real story is a global drop in migration desire driven by development in origin countries. A lively side debate erupted over H1B visas and whether high-skilled immigration helps or hurts American workers, with one faction calling for expanded visas and another arguing it’s self-serving and out of touch with domestic labor. The thread also surfaced a geographic blind spot: China didn’t even register as a destination, which undercut the common assumption that rising Asian prosperity would shift flows toward Beijing.
Big Fish: Exploiting a 37-Year-Old Kernel Bug in OpenBSD [comments]
185 points · 56 comments · calif.io · 17h ago
The article details how researchers exploited a 37-year-old integer overflow in OpenBSD's shared memory subsystem (a signed 16-bit counter that wraps), chaining it with a race condition and out-of-bounds write to achieve local privilege escalation on a fully hardened system. Hacker News immediately jumped on the fact that NetBSD accidentally fixed this bug years ago during a code rewrite, and the thread quickly split into two camps: one arguing that a bug surviving 37 years in a supposedly "heavily audited" codebase proves security audits are overrated, and the other pushing back that OpenBSD's default install is so minimal its famous "two remote holes" claim is essentially meaningless. A secondary argument erupted over the write-up itself—some complained it was too terse and possibly AI-generated, others countered that the complaint was really just about a human writer not being verbose enough. The conversation also drifted into the broader impossibility of writing correct C at scale, with multiple commenters pointing AI at small OpenBSD tools and finding bugs, reinforcing the idea that only moving to memory-safe languages will ultimately fix this class of problem.
FLUX 3 Image Klein [comments]
158 points · 44 comments · bfl.ai · 12h ago
Black Forest Labs released FLUX 3 Image Klein, a 5.6B-parameter open-weights image model touted for native 4K generation, editing, and bounding-box control, optimized to run on consumer GPUs. The HN thread barely engaged with those claims — instead, the consensus was that the text-to-image generation is genuinely bad, with multiple people reporting melted faces, six-fingered hands, and body horror straight out of SD 1.5. The blame landed squarely on aggressive safety training: commenters argued that Black Forest Labs neutered the model by scrubbing anything close to NSFW from the dataset, and that this lobotomization destroyed its understanding of human anatomy. A few defenders pointed out that the model's editing capabilities remain solid and that a prompt upsampler is essentially required, but the counterexamples included a screenshot from the upsampler's own demo showing a man with six fingers. Others pointed to competing releases like Moss Image (Apache 2.0) and Krea 3 as looking visibly better, and the thread drifted into wondering whether European safety regulation is handing the lead to Chinese open models that don't handicap themselves.
/dev/null: social network for developers [comments]
150 points · 131 comments · dev-null.sh · 15h ago
The submission is a new developer-focused social network called /dev/null, built entirely over a livestreamed 48-hour weekend using PHP and heavy AI assistance. HN immediately took issue with the site having almost no information on the landing page, requiring GitHub-only login, and claiming "no likes" while still implementing a star feature that functions identically — the creators quickly fixed the landing page after the criticism but the GitHub auth lock-in remained a sore point, with several people arguing it excludes developers who don't have or don't want to expose their GitHub profile. A chunk of the thread turned into a philosophical debate about the name itself, mocking the "dev = developer" reading when /dev/null traditionally means device null, and whether building yet another Twitter clone for developers is adding anything over Mastodon or the fediverse. The site promptly got hugged to death by the HN traffic, drawing wry jokes about irony, and there was scattered appreciation for the open-source approach and rapid iteration despite the rough launch.
What on Earth Is Dekton? [comments]
145 points · 52 comments · www.worksinprogress.news · 17h ago
Charles Mann’s article on Dekton is a sprawling history of porcelain, from ancient Chinese kilns to the modern slab he grudgingly bought for his induction cooktop. The HN crowd didn’t let the framing slide—several people pointed out that the author’s “kitchen nerd” intro undercut itself, since any real kitchen nerd would know induction tops can use tempered glass, and that Dekton is primarily a countertop material being shoehorned into a trendy hidden-cooktop design that’s arguably worse than a normal visible one. The thread also veered into pricing debates, with a split over whether induction is genuinely overpriced or just a US market quirk, and some readers did the math on the author’s implied $150/sq ft for Dekton. A deeper cut from people in the countertop industry argued the article missed the point entirely—the real driver for porcelain’s rise in countertops is the silicosis lawsuits against engineered quartz, not cooktop tech, though others countered that quartz still looks more like natural stone and porcelain can end up looking like “printer paper.”
Designing more efficient quantum circuits with Claude [comments]
134 points · 32 comments · www.anthropic.com · 14h ago
Anthropic published research showing that Claude autonomously designed quantum circuits that beat best-published results by up to 17× in gate count, targeting eight primitives from chemistry and number theory. The HN thread immediately seized on a hallucination in the blog post's acknowledgements — the actual arXiv paper thanks different people and doesn't mention the quoted mentor or program — leading one comment to accuse Anthropic's own writeup of containing fictitious personnel. Another angle dug into the technical substrate: the work uses Google's Kickmix language, which makes quantum circuit optimization a kind of code-golf problem with a closed verification loop, so it's less a breakthrough and more a well-scoped automation of a task academics hadn't bothered to optimize this far. Several commenters pushed back on the article's framing of "optimizing" edge cases away, arguing that dropping rare-input handling means the circuit doesn't perform the same computation — a concern that matters for real chemistry, where rare edge cases are often biologically important. There was also a split on the "envelope principle," with one side arguing AI is just filling in blanks while real progress expands the envelope, and others pointing out that cross-domain synthesis has always been how the envelope grows, and AI is simply better at that synthesis now.
Artificial intelligence is capitalism's final destination [comments]
111 points · 122 comments · www.equator.org · 19h ago
Arundhati Roy’s piece argues that AI isn’t a technological breakthrough but capitalism’s logical endpoint—a system that extracts human meaning, labor, and language the way it already strips mines and rivers. On HN, most of the reaction wasn’t to her actual thesis but to her prose style: a pile-on called it self-indulgent, onanistic, too long, and ironically something an AI would have written more efficiently. A few defenders pushed back, noting that dismissing beautiful writing as “AI-sounding” is exactly the kind of suspicion Roy warns about. Meanwhile, a separate and louder thread outright rejected the premise—commenters celebrated capitalism for delivering cheap goods and AI weapons that stop Russia, accused critics of envy, and argued that the real problem isn’t capitalism but regulatory capture or monopolies. The split wasn’t about AI risks; it was about whether capitalism is the villain or the engine of everything worth having.
What was a Token Ring network? [comments]
109 points · 85 comments · dfarq.homeip.net · 18h ago
The article explains what Token Ring was—an IBM-backed 80s/90s networking tech that used token passing to avoid collisions but lost to Ethernet on cost and speed scaling. HN dug into the nostalgia hard, with people sharing war stories of accidentally mixing 4 and 16 megabit rings to nuke the whole network, and one guy finding a Token Ring wall jack still in a hospital delivery room in 2013. A bunch of old-timers pushed back on the idea Ethernet was inherently worse: they pointed out that early Ethernet also required all nodes on a segment to run the same speed, and that switching—not the protocol itself—fixed collisions. The real fun came from a tangent arguing that token-passing lives on in deterministic real-time systems like CAN bus and even a recent automotive Ethernet variant (10base-T1S) that uses a virtual token ring, so the concept isn’t as dead as the article suggests.
RFC8950: Announce IPv4 with an IPv6 next-hop [comments]
104 points · 69 comments · anuragbhatia.com · 18h ago
The article walks through a lab setup for RFC8950, which lets you carry IPv4 routes over an IPv6-only backbone by using IPv6 next-hops in BGP, letting providers keep IPv4 only at the edge without tunnels. HN immediately corrected the framing: the real point isn't to “facilitate IPv6” but to let operators finally turn off IPv4 in their core networks, with commenters pointing out the practical wins like ditching numbered point-to-point links and just using IPv6 link-local addresses. A long, heated split erupted over whether IPv6 itself is fundamentally broken—one side calling it an over-engineered mess that still can't handle LAN printing reliably, the other arguing the transition mess is a human problem, not a protocol failure, and that every major OS ships with workarounds like privacy extensions and CLAT support. Along the way, someone noted that using IPv6 link-local as a next-hop breaks classic traceroute output, but there’s RFC 5837 to fix that, and others chimed in with alternative transition mechanisms like MAP-T for stateless CGNAT.
My old job feels distant now [comments]
97 points · 95 comments · www.pinecoder.dev · 18h ago
The article is a first-person chronicle of how AI coding tools have progressively transformed the author’s work — from ChatGPT as a smarter Google, through agentic flows, to near-autonomous agents that ship entire features while the author shifts to thinking about architecture and QA. HN took the piece as a lightning rod: the top comment dismisses the author as driving off a cliff, warning he’ll be replaced by a $10/hour student, while others push back that inexperienced hires can’t maintain complex systems and that senior salaries are what’s actually about to cave. A split emerged between those who see AI coding as an unavoidable new reality (“you either adopt the tools now or have trouble finding a job later”) and those who argue the economics don’t add up — pointing to the money-out vs. money-in losses at AI labs and the availability of open-weight models that undercut the whole subscription model. Another contingent mocked the author’s naivety about the “next phase,” noting every such post ends with “but they still need me to press Enter,” while a bizarre tangent saw someone claiming to “vibe code apps while driving,” drawing sharp safety pushback.
Inside Polymarket's paid quote-retweet esports program [comments]
96 points · 59 comments · www.readergrev.com · 16h ago
The investigation focuses on Polymarket's Discord-managed paid quote-retweet army of over 800 participants who are paid in crypto to make their esports betting promotion look like organic conversation. The Hacker News thread largely shrugged at the standard influencer marketing slime but seized on the adjacent detail that Polymarket's own official accounts are frequently just rebranded handles bought from deactivated or canceled esports personalities—effectively purchasing an audience history to pose as community members. A decisive split ran through the discussion over the "no house edge, it's PvP" defense of prediction markets: people who actively trade on these platforms stepped in to clarify that high-frequency trading firms and colocation agreements extract the liquidity edge, meaning retail participants are the product, not traders on a level field. The thread ultimately settled into a bitter recognition that this whole operation is less a scandal and more a textbook example of gambling infrastructure embedding itself into media feeds as the last perceived escape hatch for a generation locked out of traditional wealth-building.
C++26: The New Debugging Header [comments]
89 points · 20 comments · www.sandordargo.com · 13h ago
A new C++26 proposal standardizes a `<debugging>` header with three functions—`breakpoint()`, `is_debugger_present()`, and `breakpoint_if_debugging()`—to replace every major project’s homegrown debug-break macro. The HN discussion immediately dived into a subtle x86 trap bug: `int3` alone can leave the saved instruction pointer pointing past the breakpoint, confusing debuggers like GDB, and commenters argued whether a `int3; nop` sequence or debugger-level compensation is the right fix. Beyond the technical debate, several people flagged that the blog post itself reads like LLM-generated content, pointing to the stilted “debugging is a runtime concern” line and the author’s overuse of em-dashes, and noted the C++ subreddit already bans this source. A smaller side thread complained about bloat and offered the usual `raise(SIGTRAP)` portability workaround, while others wished the committee would standardize more platform-specific patterns instead of leaving them to each project.
Anthropic Console overcharged my card 100x due to internationalization input bug
87 points · 36 comments · news.ycombinator.com · 14h ago
This is a detailed post from a developer whose attempt to add $5.04 in credits to their Anthropic account resulted in a $569.52 charge because the billing pipeline took their decimal dot, interpreted it as a thousands separator, and processed the whole thing without a confirmation screen. The thread largely agreed this is a fundamental i18n and UX failure for a payment system—basic validation and a review step should be table-stakes. Where things split was on the support process: the submitter hit an automated AI bot that kept denying the refund, which a big chunk of the thread saw as the more alarming systemic problem, arguing that having to turn to HN to get a human's attention is an unacceptable support model for a company handling money. An Anthropic engineer did show up to confirm the bug and sort out a fix and refund, but the broader takeaway for many was less about the specific decimal parsing and more about what it reveals when AI agents are the only layer between a user and a company that overcharged them.
How to Formalize All of Mathematics [comments]
80 points · 78 comments · xenaproject.wordpress.com · 18h ago
The post is a guest essay from seven prominent mathematicians arguing that while AI can now formalize theorems at scale, rushing to formalize everything with AI-generated code risks producing useless libraries that damage mathematical culture — they advocate for careful, human-curated projects like Mathlib and the newer Tau Ceti pipeline. The top of the thread immediately gets snagged on Gödel’s incompleteness theorems, with several people pointing out that formalizing “all” of mathematics is impossible in principle; others push back hard, noting that the authors (including Jeremy Avigad, a logic chair at CMU) are well aware of Gödel’s limits and the essay explicitly says “all, or almost all” — the incompleteness reflex is repeatedly called out as a shallow gotcha that misses the point. A more substantial split emerges between those who see AI as a genuine partner for working mathematicians and those who worry that LLMs will generate inscrutable, self-referential branches of mathematics that only other AIs can consume, while real mathematicians lose control over definitions and conceptual structure. Several commenters with research experience note that AI still can’t make new definitions or see structure, only brute-force patterns from known work, and that the real bottleneck isn’t formalization speed but quality — echoing the article’s warning that a bad formalization is worse than none at all.
Generated 2026-10-17 08:06 UTC
Generated by Sauron from Hacker News discussions and linked articles.