HN Brief: 2026-05-20
Today’s HN was split between Big AI moves and the backlash they’re stirring. The Karpathy-to-Anthropic hire sparked a debate about whether AGI is close enough to abandon education, while Google’s Gemini 3.5 Flash launch drew fire over its steep pricing and flaky API reliability. A second throughline was platform risk: Google Cloud nuked Railway’s account without warning, CISA left its keys on GitHub for six months, and Minnesota became the first state to ban prediction markets—all feeding a sense that trust in centralized tech and government is fraying. Meanwhile, the AI watermark wars escalated, with OpenAI adopting SynthID and a new tool emerging to strip it, leaving the community split over motive and effectiveness.
The threads worth clicking into: “I’ve joined Anthropic” for the wild split over whether Karpathy’s move signals AGI’s imminence or just another sport-style free agency. “Gemini 3.5 Flash” for the pricing shock and the buried finding that the serving backend alone swings accuracy by 75 points. “Minnesota becomes first state to ban prediction markets” for the deep clash over whether these markets are useful information aggregation or disguised gambling with corruption risks. “Railway Blocked by Google Cloud” for the grim lesson in single-hyperscaler dependency and automated bans that can kill a business. And “Disney erased FiveThirtyEight” for the corporate neglect story and the link-rot epidemic that erased 200,000 person-hours of journalism.
I’ve joined Anthropic [comments]
1291 points · 533 comments · www.axios.com · 16h ago
Andrej Karpathy announced he's joining Anthropic's pre-training team, a major talent grab for the rival lab. The thread immediately questioned why he'd abandon his educational startup Eureka Labs now, with some arguing that AGI is near enough to make educating people pointless, while others countered that if AGI were truly around the corner, he'd be buying a yacht, not taking a job. Another strong vein compared the hire to sports free agency—calling AI news "interchangeable with ESPN"—and debated whether superstar talent actually hurts team cohesion. The biggest split came over Anthropic's reputation: a vocal group insisted the company is just marketing itself as the safe alternative while quietly backpedaling on red lines and cozying up to the Pentagon, while defenders pointed to those red lines still being in place and argued no large org is truly good.
Gemini 3.5 Flash [comments]
749 points · 525 comments · blog.google · 14h ago
Google announced Gemini 3.5 Flash, a new model focused on agentic tasks and coding performance, positioning it as a major step toward "frontier intelligence with action" and rolling it out across Gemini, Search AI Mode, and developer platforms. The thread instantly latched onto the pricing—$9 per million output tokens, a huge jump from the Flash family's usual cost—splitting between people who think it reflects a larger, more expensive model and those who see it as Google trying to extract ROI while the underlying hardware (new TPUs) makes it faster rather than bigger. A recurring counterpoint from people who regularly use Gemini in production is that the API is flaky, with unreliable caching and random error codes, making it a non-starter compared to OpenAI, Anthropic, or Fireworks despite the claimed benchmark wins. Others dug into the naming confusion, noting that calling this "Flash" while pricing it like a Pro model is deceptive, and several commenters ran their own creative coding tests (animated SVGs) showing that 3.5 Flash actually beats the previous Pro model on token efficiency for the same output quality.
I’ve built a virtual museum with nearly every operating system you can think of [comments]
744 points · 166 comments · virtualosmuseum.org · 16h ago
Someone released a virtual museum that's basically a massive pre-configured VM image with over 570 operating systems from the Manchester Baby to recent betas, all set up and ready to run in QEMU or VirtualBox. The thread immediately dove into nostalgia and obscure OS trivia—people were asking about specific oddities like the Compaq Windows 3.1 with a tabbed paper-folder desktop shell, or whether TempleOS made the cut (it didn't seem to), and a long tangent emerged about Domain/OS from Apollo workstations, including a debate over whether its typeahead buffer feature could be replicated with modern PTYs. The biggest frustration was the lack of any searchable list of included OSes on the website, which several people guessed was intentional to avoid copyright takedowns, but it made evaluating the collection impossible without downloading 120GB—and the download itself was failing repeatedly, with users posting live updates of resume attempts and eventually seeding a torrent themselves. Some pushback focused on notable omissions like Novell Netware, though the project maintainer clarified that NetWare 4.11 and 6.5 are actually included, just not shown in screenshots.
Apple unveils new accessibility features [comments]
667 points · 339 comments · www.apple.com · 20h ago
Apple’s press release announced a bunch of new accessibility features powered by Apple Intelligence — better VoiceOver, natural-language Voice Control, on-device generated subtitles for uncaptioned video, and even eye-tracking wheelchair control for Vision Pro. The HN thread immediately split into two camps: people who cheered genuinely useful LLM applications hitting underserved users, and a louder group pointing out that almost every feature here (screen description, live captions, AI-powered screen reading) has existed on Android and Windows for years, framing this as classic Apple late-to-market marketing rather than true innovation. A recurring sub-argument hammered the bill-reading example in the press release — the assistant tells you to verify the amount with the utility company, which critics called a half-baked cop-out that still forces blind users into phone-tree hell. The thread also spun off into a sharp tangent about screen reader speed: sighted developers marveling at how blind users blast through text at insane rates, while a few engineers with real accessibility testing experience pushed back that Apple’s own slow “Announce Notifications” is infuriating even for sighted users, and that the real test is whether these features survive contact with broken third-party apps.
Minnesota becomes first state to ban prediction markets [comments]
633 points · 192 comments · www.npr.org · 12h ago
Minnesota just became the first state to outright ban prediction markets like Kalshi and Polymarket, and the Trump administration is already suing to block the law on federal preemption grounds. The HN crowd immediately jumped on the enforceability question — the law also targets VPNs and payment providers, which some argue is a mess that could even sweep in the state lottery if someone bets on it out of state. A big split emerged over whether these markets are just rebranded sports gambling or actually provide useful information aggregation, with defenders pointing to crop-hedging carve-outs and the difference between a zero-sum bet and a stock exchange. Skeptics pushed back hard, arguing that 85% of volume is on sports, that "prediction market" is a semantic dodge to bypass gambling regulations, and that letting people wager on things like elections or public figures introduces real corruption risks. The deeper tension was between those who see this as a states'-rights vs. federal turf war and those who think the core issue is simply what kinds of bets should be legal at all, regardless of the platform.
Google changes its search box [comments]
505 points · 690 comments · blog.google · 13h ago
Google is rolling out a massive redesign of its search box, turning it into an AI-powered chat interface with agents that can monitor the web, book services, and even make phone calls — the biggest shift to Search in over 25 years. The HN crowd is largely hostile, seeing this as a forced migration into a chatbot UI that nobody asked for, with many comparing it to Apple removing the headphone jack: a change for the worse driven by sunk costs and a misguided belief that AI chat is the future. Several people point out that the real motive is monetization — a chat UI opens up much richer avenues for advertising and sentiment manipulation than plain links, and the long-term risk is that Google will bake ads directly into model responses where ad blockers can’t touch them. Others argue it’s a prisoner’s dilemma: if Google didn’t jump on AI, ChatGPT and slop generators would destroy the open web anyway, so disruption is inevitable. Meanwhile, a vocal minority is already migrating to Kagi, DuckDuckGo, and even Bing, betting that Google is handing its competitors a market opening — though cynics note Microsoft will likely ruin Bing the same way.
Show HN: Gaussian Splat of a Strawberry [comments]
496 points · 190 comments · superspl.at · 21h ago
The submission is a detailed 3D Gaussian splat of a strawberry, reconstructed from roughly 7,900 focus-stacked macro photos shot from 90 angles. The thread quickly turned into a meta-joke about the strawberry’s rotten-looking underside, which is actually just missing splats from its mounting setup — the submitter explained it's resting on nails and pins, and he can’t quite get the camera underneath. People dug into the technical craft: the Nikon Z8’s burst speed means the whole capture takes only 20 minutes, and there’s serious debate about whether focus stacking is even necessary for 3DGS given that blur holds depth cues, with links to DoF-Gaussian papers as a better approach. A lighter side emerged around the “dreamy” degradation splats show when you zoom in — described as reality unraveling into fog and northern-lights slivers — and one person suggested this is what Google Street View should feel like, though others noted splats are too huge to scale.
Tesla's lithium refinery discharges 231,000 gallons of polluted wastewater a day [comments]
463 points · 222 comments · www.autonocion.com · 12h ago
The article reports that a routine inspection by a Texas drainage district discovered a pipe from Tesla’s $1 billion lithium refinery discharging dark, polluted wastewater—231,000 gallons a day—into local waterways, with independent testing finding hexavalent chromium, arsenic, and other contaminants not covered by the plant’s permit. The HN thread largely split over whether the detected levels are actually dangerous: one camp pointed out that the hexavalent chromium concentration (0.0104 mg/L) is just above the reporting limit and comparable to levels in the Erin Brockovich case, while others argued the numbers are low enough that single doses wouldn’t impair health and that the permit limits weren’t technically violated. A significant chunk of the discussion turned into a debate about dilution—whether adding more water to stay under reporting limits is a valid workaround, with deep tangents into smog pumps and catalytic converter regulations from the 1980s. The bigger tension was between those who see the situation as proof that Tesla’s “acid-free clean process” marketing was always bunk and those who think the outrage is overblown given that even the independent test results don’t clearly prove acute harm—though several people pushed back by stressing bioaccumulation and the permit’s failure to monitor for heavy metals or lithium itself.
Railway Blocked by Google Cloud [comments]
455 points · 211 comments · status.railway.com · 7h ago
The article is Railway’s status page showing fully operational services, but that’s just the backdrop—the actual incident was Google Cloud abruptly blocking Railway’s account, taking them down for hours. HN lit into GCP for this pattern, with multiple people citing past instances of Google nuking startups without warning, and a clear consensus that AWS and Azure handle abuse flags far better (usually contacting you first). The split came when some argued GCP has fewer major outages than AWS/Azure, while others shot back that GCP just has fewer customers left to notice because they keep banning them. A tangent emerged on domain registrar risk and multi-cloud failover, with the strong takeaway that if you’re serious, you need completely independent billing and hosting across providers so one account block doesn’t sink you.
Show HN: Forge – Guardrails take an 8B model from 53% to 99% on agentic tasks [comments]
439 points · 171 comments · github.com · 19h ago
This is a Show HN for Forge, an open-source reliability layer that wraps local LLMs with guardrails to dramatically boost their performance on multi-step agentic tasks—the author reports pushing an 8B model from roughly 53% to 99% accuracy without retraining, just by catching and nudging the model when it fails to call tools correctly. The HN crowd dug into the structural insight behind it: the framework treats tool-calling failures as a mechanical reliability problem (compounding small per-step errors) rather than a capability gap, and solves it with domain-agnostic retry nudges, error recovery loops, and VRAM-aware context compaction to keep small models from silently falling back to CPU. Several people pressed on the eval methodology, and the author clarified that the benchmarks are deliberately scoped as a stress test of the recovery loop rather than a measure of end-to-end agentic quality—for instance, if a model tries to analyze sales data without first fetching it, Forge catches the prerequisite violation and injects a corrective message before the bad call ever reaches the tool. A significant tangent emerged around the author's writing style; a vocal minority accused the submission and all his replies of being AI-generated slop, which he denied while acknowledging he uses LLMs to help draft text, and that pushed back-and-forth dominated a chunk of the thread. One of the more technically interesting points was the finding that the serving backend alone can swing accuracy by 75 points on the same weights—llama-server versus Llamafile gave wildly different results for Mistral-Nemo—which the author says standard benchmarks miss because they don't control for infrastructure.
Disney erased FiveThirtyEight [comments]
407 points · 215 comments · www.natesilver.net · 13h ago
Nate Silver’s post details how Disney/ABC News quietly deleted the entire FiveThirtyEight archive, redirecting all URLs to the ABC homepage and effectively erasing roughly 200,000 person-hours of work—a move he frames as a business-school case study in corporate neglect. The HN thread largely swerved into a heated debate about media bias: one early skeptic claimed errors in Silver’s models always leaned one way, which sparked a full-blown argument over whether the mainstream press has actually moved left or right over the last decade, with multiple commenters pointing to the New York Times and Washington Post shifting rightward and others insisting the Overton window has simply moved right relative to steady coverage. A separate, more productive vein acknowledged that link rot is an epidemic—a commenter who worked as a tech journalist shared that almost none of their pre-2008 work survives online, and several people urged donating to the Internet Archive, which still holds the erased content. A final tangent likened the situation to Disney’s treatment of Star Wars, where early IP was purged and the brand degraded, though others pushed back that Disney squeezed every dollar from that franchise whereas FiveThirtyEight was simply starved.
OpenBSD 7.9 [comments]
392 points · 282 comments · www.openbsd.org · 18h ago
The article is the OpenBSD 7.9 release announcement, packed with new features like CPU core management, delayed hibernation, 802.11ax support, and a new song — the first since 7.3, which immediately drew cheers. The thread quickly turned into a full-blown BSD family debate: the usual "FreeBSD for utility, OpenBSD for security, NetBSD for portability" shorthand got shot down hard by people pointing out the real cross-pollination and nuance, especially around desktop usability and driver support. Several veterans chimed in with real-world use cases — one team runs OpenBSD on bare metal for security-conscious clients precisely to avoid the constant Linux sysadmin treadmill, another uses it as a hardware diagnostic because it kernel-panics on flaky gear. The "Canadian OS" quip led to a tangent about crypto export restrictions, which was corrected as largely a 90s/early 2000s issue, though red-flag countries technically remain under embargo.
GitHub is investigating unauthorized access to their internal repositories [comments]
392 points · 206 comments · x.com · 8h ago
GitHub announced on X that it’s investigating unauthorized access to its internal repositories; the attackers (TeamPCP, behind the Shai-Hulud malware) claim to have copied all repos and are selling them. The HN thread immediately zeroed in on a split over AI’s role: one camp argues that increasingly capable LLMs are fueling a wave of new exploits and are now a standard tool in vulnerability research, while the other side counters that the real problem is defenders being forced to merge mountains of AI-generated, poorly reviewed PRs that quietly break security boundaries. Several people with deep CI/CD experience warned that GitHub Actions workflows are especially vulnerable to template injection from PR titles and descriptions, pointing to the `zizmor` static-analysis tool as a must-run. A significant group pushed back on GitHub’s choice to break this news only on X, not on their status page or blog, calling it a poor channel for paying customers who don’t have Twitter accounts, though others noted that regulatory obligations (DORA, NIS2) likely required the rapid notification regardless of platform.
Gemini Omni [comments]
306 points · 128 comments · deepmind.google · 14h ago
Google DeepMind released Gemini Omni, a video generation and editing model that can create or modify footage from text, images, and audio with multi-turn editing and physics-aware realism. The thread quickly split: some predicted Hollywood’s imminent disruption, but others pushed back hard—one VFX practitioner wrote a lengthy, detailed correction debunking the "no CGI" myths around *Project Hail Mary* and *Top Gun*, arguing that studio PR spins "no green screen" into "no CGI," which deprives actual VFX artists of credit. Meanwhile, several commenters tried the model and found it lagged behind Seedance 2, and the demo page itself drew complaints for crashing browsers with too many autoplaying videos. A separate concern was that AI video has already destroyed trust in footage, with SynthID watermarking seen as a too-late fix, and one self-described AI optimist admitted video generation depresses him because knowing something is AI-made instantly kills any visual awe.
OpenAI Adopts Google's SynthID Watermark for AI Images with Verification Tool [comments]
270 points · 140 comments · openai.com · 12h ago
OpenAI announced it's integrating Google DeepMind's SynthID watermark into its image generation tools, alongside C2PA metadata and a public verification tool, to make AI-generated images traceable. The HN crowd immediately split: some argued the watermark is actually robust and hasn't been reliably broken in practice, while others claimed a simple Diffusion-based denoising loop at low strength can strip it trivially—several said they'd built working removers but never published them. A strong undercurrent of suspicion emerged that the real goal isn't stopping misinformation but letting OpenAI and Google filter out AI-generated content from their own training data to prevent model collapse. Meanwhile, a vocal minority saw the closed-source, persistent watermark as a dystopian step toward normalizing indelible tracking on all digital media, not just AI images—comparing it to printer serial number tracking and warning it will eventually be applied to camera output. Overall, most agreed the technique is a temporary speed bump for determined bad actors, but the debate was less about the article's stated benefits and more about motive, effectiveness, and the slippery slope toward mandatory content pedigree.
Remove–AI–Watermarks – CLI and library for removing AI watermarks from images [comments]
257 points · 142 comments · github.com · 9h ago
The post is an open-source CLI and library that strips both visible watermarks (like Gemini's sparkle logo) and invisible ones (SynthID, C2PA metadata) from images generated by most major AI models. HN immediately split into two camps: those who see watermark removal as a way to protect against false-positive accusations that real photos are AI-generated, and those who call that justification bullshit—arguing the tool's real purpose is to help people pass off AI slop as human-made. A technical comment landed hard, pointing out that for Gemini the tool only properly removes the visible watermark, and the “invisible” removal requires regenerating the image with SDXL at low noise, which destroys small details and won't work at higher resolutions like 4K. Another thread spiraled into a broader debate about the hacker ethos: whether fighting watermarks within the corporate system tacitly accepts the whole barcode surveillance regime, or whether the right move is to just run open models locally and ignore the arms race entirely.
Mistral AI acquires Emmi AI [comments]
244 points · 69 comments · www.emmi.ai · 12h ago
Mistral AI is buying Emmi AI to build a combined platform for physics-based industrial simulation — think real-time mold flow analysis for injection molding, power grid stabilization, and digital twins for aerospace and automotive. The HN thread immediately zeroed in on ASML as a major Mistral investor, with some arguing that ASML’s backing adds credibility to Mistral’s industrial pivot, while others countered that it’s a politically nudged “sovereign EU AI” investment with no real technical synergy. Skepticism ran deep: several people pointed out that Emmi has been around only 18 months, has no publicly visible product or customer list beyond a partnership with a German simulation company, and that the acquisition looks more like a talent grab for its 30-person team than a strategic product move. A separate debate erupted over whether Mistral can compete at all — most commenters rated it below the Big 2 (or 3) and Chinese labs — but others argued it doesn’t need to be best-in-class because European governments and defense buyers will prioritize a homegrown, sovereign stack over raw performance.
Dumb ways for an open source project to die [comments]
182 points · 118 comments · nesbitt.io · 12h ago
Andrew Nesbitt's post catalogs dozens of ways open source projects die, from the classic "maintainer vanished" to obscure failure modes like build archaeology and sanctions-stranded packages. The thread mostly ran with it, adding its own pet death modes—the "overconfident fork" that never gains critical mass, and the rage-fork that actually saves the project, with people citing io.js and EGCS as happy examples of the latter. A substantial chunk of the discussion split off into a heated debate about whether we should ditch dependencies entirely for a "function hub" of immutable, never-updated code blocks, with some calling it a natural evolution and others pointing out that lockfiles already solve the same problem. Another argument broke out over the "ungrateful user" pattern, where one side insisted open source is pure altruism and maintainers have no right to expect contributions, while others pushed back hard that the real answer is just paying people.
Google I/O [comments]
179 points · 257 comments · io.google · 15h ago
The linked article wasn't available to this summarizer; from the discussion, Google I/O 2026 is apparently an AI-only keynote, and the HN thread is basically one long, weary joke about how every product announcement is just another rebranded AI model — "AGI now! … ASI now! … Super Duper AI!" — with real complaints buried underneath. There's a split between people mocking the hype cycle and those genuinely frustrated that Google's coding agents (Gemini CLI, "Antigravity") are still unreliable, breaking basic file edits and burning through quota while marketing hypes a "vibe code a whole OS" demo that costs $1,000 in tokens and produces unmaintainable slop. Others defend Google's trajectory, pointing out that Flash 3.5 is now topping benchmarks and the stock doubled, while critics counter that the benchmarks don't measure real agent quality and that forcing "AI Mode" onto search doesn't prove genuine demand — it just hides the ads better. Also, nobody could get the livestream button to work in Firefox.
Gemini CLI will stop working from June 18, 2026 [comments]
179 points · 91 comments · developers.googleblog.com · 14h ago
Google is shutting down its Gemini CLI on June 18, 2026, and rolling everyone into a new, unified product called Antigravity CLI — which launches today. The HN crowd is mostly furious, seeing this as another entry in the Google Graveyard playbook: killing a working open-source tool (Apache 2, 100k+ GitHub stars) in favor of a closed-source replacement that, by many accounts, is buggy, half-baked, and already riddled with opaque quotas and regional lockouts. The short one-month transition window and the loss of agent-client-protocol support in the new CLI are drawing particular ire, though a few people argue that consolidating Google's fragmented AI offerings was overdue. A recurring split: some say Gemini CLI was genuinely slow and bad, others say they relied on it daily and now have no clear paid alternative.
FiveThirtyEight articles on the Internet Archive [comments]
167 points · 39 comments · fivethirtyeightindex.com · 6h ago
This is an index of 21,350 FiveThirtyEight articles that ABC News deleted from the internet, now searchable via the Internet Archive. HN immediately recognized this as a response to Disney’s erasure of roughly 200,000 person-hours of journalism, with several people linking to Nate Silver’s own piece calling it “stratospheric negligence.” A chunk of the thread veered into a familiar argument over Silver’s 2016 election predictions—some claiming his models were sound (30% chance means it happens three times out of ten) while others argued his tone and framing showed he fundamentally misunderstood the political reality. There was also a nitty-gritty technical sideline about whether the Internet Archive actually ignores robots.txt for preservation, with people digging into the archive’s 2017 policy statements and noting that past interactive visualizations are still broken in the archived copies.
AI is too expensive [comments]
137 points · 144 comments · www.wheresyoured.at · 16h ago
Ed Zitron's article argues that the AI industry is economically unsustainable, citing over $800 billion in hyperscaler capex, plummeting margins at Anthropic and OpenAI, and case studies like Zillow burning through a year's token budget in five months. HN mostly agreed the current cheap pricing is a bait-and-switch, comparing it to Uber's subsidized rides before jacking up prices, but many countered that compute costs don't follow the same trajectory—open-weight models like DeepSeek already undercut frontier providers, and hardware improvements keep driving inference costs down over the long run. Some pushed back on Zitron's credibility, saying his dismissal of AI agents as "chatbots plugged into an API" ignores genuinely useful business automation, while others doubled down that the real endgame is embedding undisclosed advertising or influence into model outputs. The split was between those who see a classic VC-funded bubble destined to pop and those who argue the tech's long-term value—or the geopolitical imperative to own the "nukes"—justifies the present losses regardless of unit economics.
Copy Fail, Dirty Frag, and Fragnesia kernel vulnerabilities [comments]
123 points · 50 comments · www.gentoo.org · 16h ago
The linked article is a Gentoo Linux advisory on three new kernel privilege escalation bugs—Copy Fail, Dirty Frag, and Fragnesia—noting that Gentoo’s binary kernels had fixes from day one while upstream remained vulnerable. The thread immediately turned into a debate about whether any distro can realistically handle the accelerating pace of kernel vuln disclosures, with one side arguing that Gentoo's sprawling menu of kernel packages (gentoo-sources, vanilla-kernel, git-sources, etc.) actually makes the workload worse, not better, because each variant needs backporting and testing. A strong strain of commenters pushed back against the idea that live-patching or automated kernel updates are the answer, pointing out that the real problem is the “allow all” default kernel configuration that loads every rarely-used module, and that the whole stack was never designed with security first—unlike projects like OpenBSD that moved slower but didn't dominate. Someone also challenged the notion that users are “getting hacked left and right” from these local-privilege-escalation bugs, arguing that without real-world exploitation reports, the call for radical changes like mandatory live-patching is overblown, while others countered with the xz backdoor near-miss as proof that the current model is fragile even if nobody's been caught yet.
Why is almost everyone right-handed? A new study connects it to bipedalism [comments]
121 points · 190 comments · journals.plos.org · 17h ago
A new Oxford study in PLOS Biology argues that human right-handedness emerged in two stages: bipedalism first freed the hands and created pressure for lateralization, then larger brains hardened the bias toward the right. Hacker News immediately pointed out that the headline promises a "why" but the paper only delivers a "how" and "when," with several people noting that the Bayesian modeling simply made humans no longer an outlier after controlling for brain size and bipedalism — which one person called "not obviously interesting." A long subthread spun off on the old hypothesis that the heart sits left, sparking a surprisingly detailed debate about situs inversus, lung lobes, and whether stab wounds through the sternum actually work, with anatomy folks dismantling the premise. The most substantive correction came from someone who clarified that handedness is two separate traits — strength of preference and direction — and that the study actually found Australopithecus was already strongly lateralized (committed handers) long before the rightward consensus emerged, so most coverage conflates two things that evolved millions of years apart. A side discussion about mixed-handedness and whether it's pathologized as a mental illness got a few grounded responses reframing it as cultural suppression.
Everything in C is undefined behavior [comments]
116 points · 83 comments · blog.habets.se · 1h ago
The blog post argues that essentially all nontrivial C/C++ code runs afoul of undefined behavior, citing examples like casting pointers in ways that allegedly trigger UB on creation. The thread immediately pushed back hard on that specific claim: several people pointed out that casting a `uint8_t*` to an `int*` is itself explicitly allowed by the standard—only dereferencing an unaligned pointer is UB, and the compiler is free to adjust the pointer's bit representation. The debate turned into a meta-example of the article's own thesis, with both sides citing spec paragraphs and accusing the other of misreading, while others noted that even if the cast is safe, the real problem is modern compilers using UB to justify aggressive, time-traveling optimizations that eliminate code paths. A side discussion questioned whether the term "C/C++" makes sense anymore, with some arguing that modern C++ with ranges and smart pointers avoids most footguns, while others retorted that borrows via `span` are still unsafe and that the culture war with Rust inevitably sneaks in.
Graduates are booing pep talks on AI at college commencements [comments]
114 points · 174 comments · apnews.com · 14h ago
The AP article reports that graduating students at several universities are loudly booing commencement speakers who pivot their addresses to AI, with former Google CEO Eric Schmidt facing particular jeers at the University of Arizona after telling students AI “will touch every profession.” On HN, the thread largely sidesteps whether the booing was justified and instead drills into the deeper panic: new grads are entering the worst entry-level job market in a dozen years, and they see AI as a final nail, not an opportunity. The dominant take is that the speaker class is cosmically out of touch—nobody is answering the “okay, then what?” question about rent and food, and the executives stand to profit while the audience gets replaced. A few commenters push back, arguing that booing is useless and students should instead organize politically or learn to use the tools, but that view gets met with scorn from people who note that even senior engineers at Google have zero say in how AI is deployed. The conversation also veers into darker territory, with some linking the tech elite’s worldview to the “Dark Enlightenment” and the recent Zizian violence, suggesting that the disconnect isn’t just tone-deafness but a genuine ideological contempt for the lives of ordinary workers.
Colonization of Venus [comments]
114 points · 94 comments · en.wikipedia.org · 19h ago
The Wikipedia article lays out the case for colonizing Venus via floating habitats at 50 km altitude, where pressure and temperature are Earth-like and breathable air itself provides lift. The thread largely accepts the engineering logic of cloud cities—several people point out that Venus's thick atmosphere offers better radiation protection than Mars, and that you can extract carbon, hydrogen, oxygen, and nitrogen from the air, making a colony potentially self-sufficient without the constant resupply Mars demands. But the real split comes over terraforming: a vocal faction argues it's absurd to discuss transforming Venus when we can't even stop un-terraforming Earth, while others counter that geoengineering Earth is trivially easy technically and only blocked politically. There's also a sharp aside about whether any of this matters if we're just "canned apes" who should instead upload into mechanical bodies or focus on solar sails and superconducting wires to generate a magnetic field.
U.S. Cybersecurity Agency Leaves Its Digital Keys Out in Public on GitHub [comments]
112 points · 25 comments · gizmodo.com · 19h ago
The article reports that CISA exposed its own cloud credentials—including plaintext passwords and AWS GovCloud keys—in a public GitHub repo for about six months. The thread quickly split over whether the piece's pivot to Trump's cuts to CISA was relevant or just gratuitous political grumbling, with one side arguing that firing competent leadership like Chris Krebs directly eroded the agency's security culture, while the other insisted that dumping API keys is a basic error no administration should excuse. A handful of replies floated the idea that the leak might have been intentional—either as a honeypot or an inside job—though most settled on Hanlon's Razor, noting that a loyalty-first hiring policy practically guarantees this level of incompetence. Someone also pointed out the grim irony that former DOGE staff, after that initiative was scrapped, were absorbed into the government and are now presumably responsible for securing federal infrastructure.
Railway Is Having a Major Outage [comments]
92 points · 79 comments · status.railway.com · 9h ago
The Railway status page shows a 98.85% uptime for May and reports that most core services are back to "Operational," but the HN discussion is all about what caused the outage—an automated Google Cloud Platform account ban that deleted Railway’s entire GCP project without warning. The thread quickly turned into a reckoning over Railway's claim that they didn't "build a cloud on another cloud," since they left their critical databases on CloudSQL and relied on GCP for overflow; a Railway founder waded into the comments to apologize, admitting the decision left them "auto-modded out of our own account" and that they've been assured by Google this wouldn't happen again. Some users are furious they were lied to about owning their own infrastructure, while others argue that true multi-cloud is rare and expensive, and that a single hyperscaler dependency is the norm for growing platforms. A few smug "eggs in one basket" comments got pushback from people pointing out that even AWS goes down, and the thread also served as a grim reminder that automated bans by big tech can nuke an entire business regardless of whether you pay—with one user noting a Korean government organization got locked out of GCP too.
Anthropic Is Preparing for IPO and We Should Be Worried [comments]
81 points · 88 comments · www.vincentschmalbach.com · 18h ago
The article lays out how Anthropic is locking down its platform and shifting costs ahead of a likely IPO—blocking third-party tools from subscriptions and moving SDK usage to separate credits. The thread largely agreed this was an inevitable monetization move, but split on whether it's a sensible business pivot or a betrayal of the open ecosystem that made Claude useful, with several people arguing that third-party harnesses often outperform Anthropic's own buggy Claude Code, and that blocking them kills the innovation that justifies the model's premium. A bigger tangent consumed much of the discussion: whether the market can absorb the simultaneous IPOs of Anthropic, OpenAI, and SpaceX, with one camp citing $195B in planned share sales versus $44B in total US IPO volume last year, while others countered that household and institutional capital flows are far larger and that the real risk is index fund forced buying at inflated prices. Someone also pointed out that the private valuations are already baked into the hyperscaler partners, so direct exposure isn't as critical as it seems.
Generated 2026-05-20 08:41 UTC
Generated by Sauron from Hacker News discussions and linked articles.