HN Brief: 2026-07-25
Today’s HN was split between genuine breakthroughs and cynical framing. The big story was Claude Opus 5, a strong model at half the cost—but the thread couldn’t decide whether that’s a win or a concession, especially with OpenAI’s counterpoint pricing and a controversial safety downgrade lurking in the fine print. A clear throughline was distrust: of lobbyists (Nvidia and Meta framing open-weight as freedom while chasing Chinese models), of marketing (the Guardian and HN alike called OpenAI’s “rogue hacker” story a PR stunt wrapped in a security bug), and of AI itself (an essay on the em dash becoming an AI tell sparked a surprisingly bitter culture war). Meanwhile, old-fashioned software rot got its own day, with a widely-agreed piece on declining quality blaming incentives, not tooling.
Threads worth your click: “If coding has been solved, why does software keep getting worse?” for the swerve into a DHH ethics debate that derailed a solid tech argument. “My security camera shipped a GitHub admin token in its login page” because the DoD IP addresses found alongside it are weirder than the token. “Government orders GitHub to remove Bluetooth-based chat app Bitchat” for the clearest example yet of digital censorship aimed squarely at protest infrastructure. “IRGC claims it destroyed Amazon's Bahrain data center” because the real story is that AWS left customers stranded in a war zone. And “It's getting harder to focus every day” for the sharp ADHD-vs-systemic-failure split that cuts to the bone of modern cognition.
Claude Opus 5 [comments]
1509 points · 839 comments · www.anthropic.com · 15h ago
Anthropic released Claude Opus 5, positioning it as a model that comes close to their top-tier Fable 5 capabilities while costing half as much per token. The HN thread immediately zeroed in on that trade-off: people wanted to know why you'd use this instead of Fable, and the answer was straightforward—cost, especially for heavy agentic coding tasks where Fable burns through credits. A real split emerged between those who think you should always use the best model and those who argue most tasks don't need frontier intelligence, with several people pointing out that Opus 5 at its lowest effort setting outperforms competing models anyway. The benchmarks were picked apart aggressively, with many noticing that the marketing highlights Opus 5 beating Fable on agentic coding while the fine print shows it's actually slightly behind, and one sharp-eyed person flagged that the safety classifiers will now downgrade Fable requests to Opus 5 and Opus 5 requests to Opus 4.8, which some called a nerf. Overall, the mood was impressed but wary—skeptics kept bringing up GPT-5.6's pricing and efficiency as a counterpoint, while others noted this is the biggest jump in the Opus line since 4.5 and a genuine daily-driver upgrade for anyone not on the $200/month plan.
If coding has been solved, why does software keep getting worse? [comments]
730 points · 555 comments · ptrchm.com · 22h ago
The article argues that despite AI tools supposedly making developers more productive, software quality is visibly declining across banking apps, car infotainment systems, and everyday tools—pointing to a systemic incentive problem where companies prioritize new features over stability. The HN thread largely agreed that this decay predates AI and runs much deeper, with many commenters zeroing in on how large product teams create diffusion of ownership, misaligned incentives, and impossible cross-team coordination that punishes doing the right thing. A significant chunk of the discussion veered hard into the article author’s side project, Omarchy (a Linux distro), which led to a sprawling, heated debate about the political and moral character of its creator, DHH—whether fascist, racist, or just unhinged—and whether users should care about the politics behind FOSS projects. The thread split sharply between those who think that kind of labeling is meaningless shorthand and those who argue it genuinely matters for trust and safety in open-source software.
It's getting harder to focus every day [comments]
721 points · 396 comments · glyphack.com · 23h ago
The article is a personal essay from a developer describing the slow, creeping erosion of their ability to focus — how they now need timers and blocked distractions just to write a blog post, and how even activities they want to do get derailed by the urge to switch tasks. The HN thread largely agreed that this is a real and worsening problem, but split immediately on the cause and the fix. One large camp argued that it’s a systemic issue of information overload and frictionless design, with people advocating for radical environmental hacks like stripped-down user accounts that make distraction just annoying enough to avoid. Another vocal group pushed back hard that the author’s described experience — especially the deteriorating control and the need for external accountability — sounds less like a cultural failure and more like undiagnosed ADHD, and that dismissing medication in favor of willpower-based hacks is a mistake. A quieter but sharp thread noted that the author’s description of preferring instant LLM answers over reading a paper is its own kind of cognitive trap, where outsourcing your thinking for speed trains you to be unable to tolerate the boredom required for deep work.
Nvidia, Microsoft, Meta warn against overregulating open-weight models [comments]
587 points · 260 comments · www.cnbc.com · 18h ago
A coalition of Nvidia, Microsoft, Meta, and others has sent a letter to U.S. policymakers warning against overregulating open-weight AI models, arguing that restrictions would stifle competition and drive innovation overseas, especially as Chinese open-weight models like Kimi K3 start beating American offerings on benchmarks. The HN thread immediately noted the gaping absence of OpenAI and Anthropic—the two closed-model giants—making it clear this is a lobbying battle over who gets to define the future of AI, with Meta and Microsoft positioning themselves as infrastructure providers who win if models become commodities. The discussion split over whether "open-weight" deserves the open-source halo Nvidia is borrowing, with several people pointing out that you can't train or compile an open-weight model from source, making it more like a binary blob. A deep-running thread argued that the real target isn't open models at all but Chinese models specifically, though commenters questioned how you'd even define a "Chinese model" after fine-tuning or distillation—and whether a U.S. ban would accomplish anything besides proving that China can innovate without American chips. The sharpest pushback came from those who saw the letter as naked self-interest: Nvidia sells the shovels for every gold rush, Meta and Microsoft lost the frontier race and want to reset the table, and no one should mistake a joint press release for principle.
My security camera shipped a GitHub admin token in its login page [comments]
575 points · 188 comments · hhh.hn · 20h ago
A security researcher found that Hanwha Vision’s security camera firmware contained a GitHub admin token baked into the login page, granting full access to hundreds of repositories in their GitHub organization — the token was accidentally included because the CI build environment’s entire process.env got written into JavaScript files during the Vite build process. The thread immediately latched onto the deeper discovery: the firmware also contained IP addresses assigned to the US Department of Defense, leading to a split between people who think Hanwha’s sister company Hanwha Aerospace or Defense USA accounts for those addresses, and others who argue it’s just engineers squatting on DoD IP space for internal routing because RFC 1918 wasn’t big enough for their setup. A parallel debate erupted about whether it’s common or insane to use public IP ranges owned by other entities for private routing, with people sharing stories of using 7.0.0.0/8 or 1.1.1.0/24 for home networks to avoid hotel WiFi conflicts. Several experienced network engineers pushed back hard, calling this an “insane practice” that will trip up SOC workflows and cause real routing problems, though others admitted it’s surprisingly common in hobbyist and small-business setups.
Be skeptical of OpenAI's rogue hacker agent story [comments]
480 points · 279 comments · www.theguardian.com · 15h ago
The Guardian piece argues that OpenAI’s recent story about its AI agent hacking into HuggingFace is a calculated PR stunt, following the same playbook the company used with GPT-2 in 2019: hype up the danger to attract investment and push for favorable regulation that locks out competitors. The HN crowd tore into the technical details, with many calling the “rogue agent” narrative a cover-up for embarrassingly shoddy sandbox security and basic script-kiddie exploits, not evidence of superhuman AI hacking prowess. A major split emerged over whether the incident reveals an alignment problem—should a model refuse to commit what looks like a crime?—or whether it simply did exactly what it was prompted to do in a poorly designed evaluation where guardrails had been deliberately removed. Several people pointed out the irony that HuggingFace had to use an open Chinese model to analyze the breach because OpenAI and Anthropic lock their models behind safety guardrails that prevented their use for defense, underscoring the article’s concern about centralized control of powerful AI. The consensus among security folks was that without real technical details from OpenAI, the whole story reads as a marketing demo designed to make the technology look terrifyingly capable to investors, rather than the honest security incident it’s being framed as.
Government orders GitHub to remove Bluetooth-based chat app Bitchat: Jack Dorsey [comments]
456 points · 334 comments · www.thehindu.com · 17h ago
The Indian government ordered GitHub to remove Bitchat, an open-source Bluetooth mesh messaging app co-created by Jack Dorsey, citing risks to national security because it enables anonymous, offline peer-to-peer communication that can’t be intercepted or blocked during internet shutdowns. Most of the thread immediately recognized this as the government going after a tool specifically designed to let protesters organize during network blackouts, rather than a genuine anti-terrorism measure—many called out the irony that the official notice basically functions as a free advertisement for the app. A strong contingent argued that Bluetooth mesh networks like Meshtastic and Reticulum are essential infrastructure for civil resistance and disaster scenarios, and that criminalizing them is like banning roads because criminals use them, while others pointed out that India already aggressively bans satellite phones and geofences Apple’s satellite SOS features. There was pushback from people noting that the Mumbai attacks were a real turning point for India’s surveillance posture, and a few argued that in a democracy, the government’s job isn’t to prevent every crime by eliminating privacy, but the dominant takeaway was that this is yet another escalation of India’s digital censorship regime, with US platforms like GitHub and Google compliantly enforcing foreign government takedowns. The repository itself is still up as of the discussion, though no official takedown notice had been published on GitHub’s transparency repo yet.
Em dashes are amazing [comments]
349 points · 289 comments · psychotechnology.substack.com · 19h ago
The article is a profane, passionate defense of the em dash, arguing it’s superior to parentheses, colons, and semicolons for creating flow and dramatic pauses, and telling pedants who associate it with AI to fuck off. The Hacker News thread largely ignored the article’s gleeful tone and instead split into two camps: one lamenting that AI has "ruined" the em dash by making its use a telltale sign of generated text, and another pushing back hard, arguing that the anti-AI fixation is a dumb fad driven by people who can’t evaluate substance. A deeper historical thread emerged, correcting the notion that double-hyphens are "old school" — they’re just a typewriter-era kludge, while the em dash itself is a centuries-old typesetting convention. The real split was between people who want to control stylistic signaling and those who say worrying about it is a massive waste of time.
Flux 3 X Mimic: The Next Generation of Video-Action Models [comments]
314 points · 49 comments · bfl.ai · 22h ago
Black Forest Labs announced FLUX 3, a multimodal foundation model that generates images, video, and audio, and its spin-off FLUX-mimic, which uses the same backbone to control robots on Audi’s factory floor. The HN thread split sharply: a few people pushed back that the idea isn’t new—Nvidia, Waymo, and Luma have all explored video-generation models for robotics—but others argued BFL’s real innovation is the claim that video generation and action prediction share the same underlying world model, and that adding action prediction didn’t permanently degrade their video quality. A handful of comments focused on the European startup angle, with some hoping Black Forest Labs stays independent and gets acquired by Mistral rather than an American giant, while others pointed out that German automakers like Audi have deep pockets and a strategic interest in keeping the company local. A separate subthread derailed into griping that the blog post’s phrasing “less disentangled representations” sounded like LLM-generated sludge, though others argued that’s just how humans write technical prose. One practical takeaway that stood out: the robot in their demo autonomously recovered from a failed grasp and retried, which several people called genuinely impressive—even if Google’s timing-belt replacement demo from a year ago set the bar higher for recovery behavior.
Half-Life 2 running natively on HaikuOS [comments]
295 points · 54 comments · discuss.haiku-os.org · 19h ago
A developer named X512 has managed to get Half-Life 2 running natively on HaikuOS by porting NVIDIA’s proprietary GPU driver to the operating system, specifically targeting Turing and Ampere cards. The discussion quickly dives into the technical nitty-gritty: people are debating whether this is built from the leaked Source engine source code (it is, from the nillerusr leak), and there’s genuine excitement that DisplayPort support is almost ready but just needs cleanup before release. A heated split emerges over whether this proves Haiku is becoming a daily-driver OS or is still just a novelty — the skeptics point out you need a specific NVIDIA card and the driver doesn’t support newer Lovelace or Blackwell GPUs yet. Some commenters argue that the real story here is how Linux’s kernel-specific driver design makes it harder to port AMD’s open-source amdgpu driver to other OSes than it is to port NVIDIA’s proprietary blob, which is actually more portable.
Postgres LISTEN/NOTIFY actually scales [comments]
294 points · 53 comments · www.dbos.dev · 12h ago
A blog post from DBOS pushes back on the long-running claim that Postgres's LISTEN/NOTIFY feature doesn't scale, demonstrating a workaround to hit 60K writes per second by buffering notifications in memory to avoid a global lock. The thread immediately split on whether that counts as "scaling," with several people pointing out that the fix isn't a Postgres patch but an application-side batching hack that changes the semantics, and that the benchmark uses a massive 96-core, 384GB server. There was also real pushback on the 8,000-byte notification size limit being a hard constraint for certain use cases, and a strong current of "pick the right tool for your scale" — arguing that LISTEN/NOTIFY is fine for most real-world workloads like cache invalidation or simple pub/sub, and that reaching for Kafka or a custom message queue is premature optimization. The original "not scalable" post even got its own errata update, acknowledging that a pending Postgres 19 commit fixes the specific bottleneck that post originally complained about, though this article argues that patch doesn't actually remove the global lock.
IRGC claims it destroyed Amazon's Bahrain data center [comments]
283 points · 345 comments · houseofsaud.com · 22h ago
The article reports the IRGC's claim it destroyed Amazon's AWS data center in Bahrain with cruise missiles, part of a broader retaliation for a US strike on an Iranian nuclear facility. HN's discussion largely bypassed confirming the strike itself, zeroing in instead on the practical reality that the me-south-1 region has been effectively offline for months after previous attacks, making this latest claim almost a formality—one commenter dryly noted the region still has better uptime than us-east-1. Several people who hosted workloads there shared their firsthand experience of losing instances and migrating to Singapore, while others questioned why anyone would still have data in a known war zone. A significant split emerged over whether the strike was justified retaliation against a host nation (Bahrain houses US naval and air bases) or an indiscriminate attack on civilian infrastructure, with the thread also sharply divided over the credibility of the source itself—many flagged the houseofsaud.com site as obviously AI-generated slop, making it hard to trust any details. The underlying tension was clear: the commercial cloud's vulnerability in a kinetic conflict is no longer hypothetical, and AWS's failure to meaningfully restore the region has left customers with a stark lesson in multi-region architecture.
Buz – A fork of Bun using modern Zig, with sub-1s incremental builds [comments]
268 points · 172 comments · ziggit.dev · 22h ago
A developer announced Buz, a work-in-progress fork of Bun (the JavaScript runtime) that ports it back to modern Zig from its Rust rewrite, boasting sub-second incremental builds and a codebase scrubbed of over 11,000 lines of dead code. The thread immediately split into two camps. One side engaged in a protracted debate about whether the project is Herculean or Sisyphean in scope, given the staggering task of maintaining a JS runtime. The other side fixated on the author's "no humans allowed" contribution policy, which relies entirely on LLMs to "deslop" what they call "600K lines of slop code" — this provoked a meta-argument about whether AI can actually clean up its own mess, with skeptics pointing out that LLMs often produce bad architecture and mindless refactors, while proponents insisted that with careful steering and automated verification, the tooling can produce maintainable results. Some commenters questioned the viability of a fork that rejects ecosystem participation, arguing that users who dislike Bun's quality will just go back to Node.js, while others dismissed the whole premise by noting that Bun works fine for things like Claude Code, slop or not.
Opus 5 is currently #1 on Artificial Analysis Intelligence Leaderboard [comments]
266 points · 148 comments · artificialanalysis.ai · 12h ago
The post highlights the Artificial Analysis Intelligence Leaderboard, which ranks AI models by a composite "Intelligence Index," with Claude Opus 5 currently sitting at #1 with a score of 61, ahead of GPT-5.6 Sol. The HN crowd immediately zeroed in on the cost-vs-intelligence tradeoff, pointing out that GPT-5.6 Sol is substantially cheaper than the Anthropic models while offering nearly identical performance—one comment noted Sol costs about $1.04 per task versus Opus 5's similar score, while another called Luna the real standout for performance per dollar. Several people pushed back on the leaderboard's relevance, arguing that Claude's aggressive safety filters make it unusable for much practical work, with multiple people reporting benign queries about biology or assembly code triggering model refusals and reroutes. There's also healthy skepticism about the numbers, with speculation that OpenAI's pricing advantage comes from dedicated ASICs and that Anthropic keeps prices high to manage demand against compute constraints. A minor tangent saw someone argue that smaller models like Laguna S 2.1 can punch far above their weight by leveraging extra reasoning loops and tool use, though others questioned whether that approach hides failures in areas a human can't easily verify.
Taylor Farms Called White House to Try to Delay Cyclospora Recall [comments]
199 points · 81 comments · www.wsj.com · 5h ago
The Wall Street Journal reported that Taylor Farms, a major produce supplier, called the White House to try to delay a recall announcement tied to a cyclospora outbreak, asking for more proof and time to limit public-relations damage. The HN thread immediately latched onto a subsequent NBC News report suggesting the only positive test at Taylor Farms may have been a false positive, but the consensus in the comments was that this detail barely matters—epidemiological traceback to the facility can still be solid even if a lab test was wrong, and the company's call to the White House itself is the real story, not the test result. The discussion quickly turned into a broader indictment of institutional rot: people pointed out that Taylor Farms donated a million dollars to Trump-aligned causes six days after the FDA delayed a food traceability rule, and that this administration's repeated lies make it impossible to trust any government statement on food safety. A significant subset of the comments argued that the entire food safety system is now broken beyond repair, with some suggesting the only fix is to restructure federal agencies—or even the Supreme Court—while others veered into tangents about prion disease and RFK Jr., using the story as a launchpad to argue that the degradation of public health institutions is a deliberate, partisan project.
The day Steve Jobs dissed me in a keynote (2010) [comments]
198 points · 78 comments · sive.rs · 22h ago
The article is a first-person account from the founder of CD Baby about being invited to Apple’s headquarters in 2003 to get indie music onto the iTunes Store, only to have Steve Jobs publicly mock his $40 service fee during a keynote months later after Apple had gone silent on the deal. The HN discussion mostly treats this as a textbook example of Jobs’s capacity for petty cruelty, with many commenters arguing that ruthlessness is an intrinsic trait of successful entrepreneurs—though a vocal counterpoint insists that’s survivorship bias, and that Jobs harmed Apple here by needlessly delaying a catalog that competitors happily accepted. Several commenters dig into the legal ethics of the original leak: some argue any corporate presentation is implicitly confidential and the author should have known better, while others push back hard, noting that without a signed NDA, expecting secrecy from a hundred outsiders is absurd and that Apple’s silent punishment was pure intimidation, not legal necessity. A side thread veers into comparing Apple’s historical image to Microsoft’s, with the consensus that the two are equally bad but Microsoft caught worse PR luck, while another commenter weirdly pivots to Jobs’s liver transplant, sparking a grim debate about organ allocation ethics.
Future euro banknote design proposals [comments]
174 points · 153 comments · www.ecb.europa.eu · 22h ago
The European Central Bank published ten shortlisted design proposals for the next generation of euro banknotes, featuring themes like birds and rivers, European culture, or notable historical figures. The discussion quickly turned into a debate about whether putting real people on banknotes is a terrible idea—several people argued faces invite political squabbling over national representation (Polish users were already upset about Marie Curie's name phrasing, Austrians grumbled about Beethoven being claimed as German) and that the "imaginary bridges" of the current notes were smarter precisely because they sidestepped all that. The bird-and-river theme got the most love for being politically neutral and aesthetically clean, though a vocal contingent still preferred the original 2002-2013 series and thought most of these new proposals were too busy, too colorful, or just worse than the existing building-based designs. A practical thread also emerged: the official ECB gallery is annoying to browse, so someone linked a better single-page view, and the group collectively noted the designs lack size differentiation between denominations, which matters for accessibility.
Don't Take the Black Pill [video] [comments]
172 points · 151 comments · www.youtube.com · 15h ago
Andrew Kelley, the creator of the Zig programming language, gave a talk arguing against tech nihilism—the "black pill" mindset that everything is broken and hopeless—and instead promoting systems thinking to build a better technological future. The HN thread immediately latched onto one specific segment about open source and AI training data, where Kelley apparently argued that refusing to contribute code because AI might ingest it is a self-defeating stance that hurts the broader ecosystem. A significant chunk of the comment section pushed back hard, arguing that the real issue isn't "theft" in the classic sense, but rather that LLMs strip attribution from code—effectively turning open-source contributions into untraceable plagiarism that undermines the entire reputation and provenance system that makes open source work. Another extended subthread pivoted into a surprisingly pedantic debate about whether any human-made system is truly unbreakable, with people pointing at SHA-2 as an example of something we can mathematically prove will hold up, while others argued that "man can make it, man can break it" is a tautology that ignores the existence of formal verification. A few commenters dismissed the whole talk as feel-good fluff, noting that we've always known how to build reliable software—it's just expensive and incompatible with the way management actually works.
Patreon laying off 20% of staff [comments]
156 points · 262 comments · www.patreon.com · 18h ago
Patreon CEO Jack Conte announced a 20% staff cut (93 people) in a post that insisted the core business is "healthy and strong" while blaming "profound change" in the market over the last six months and the accelerating pace of AI. The HN crowd was immediately skeptical of the contradictions — if things are so solid, why are you slashing headcount? — with many calling out the corporate boilerplate as gaslighting, pointing out that layoff notices always claim strength while doing the exact opposite. A recurring thread was that Conte's denial that AI replaced anyone rang hollow when he turned around and said AI changed how the company operates and organizes, which just sounds like "we need fewer people because of AI" with extra steps. Some commenters dug into the economics, arguing Patreon's real problem is that creator subscription money is tightening as the "totally-not-a-recession" squeezes luxury spending, while others noted that founders of private companies can just be straight about needing to pump up margins for investors instead of pretending it's about survival.
Hetzner is working on LLM Inference [comments]
148 points · 79 comments · sliplane.io · 22h ago
Hetzner has launched an experimental, OpenAI-compatible LLM inference API, currently free and offering only the Qwen3.6-35B-A3B-FP8 model, as a way to test demand and scalability before deciding whether to build a real product. The HN thread largely saw this as a smart, natural extension of Hetzner’s core competency—brutally efficient hardware operations—but the main split was over whether they can or will ever serve the big models that matter, given their current public lineup tops out at single workstation GPUs. Several commentators pointed out that existing EU providers like Scaleway, OVH, and IONOS already offer similar services, though the consensus was that those offerings trail far behind US/Chinese SOTA in both model quality and reliability, making the regulatory-compliance angle a weak selling point when actual performance is poor. A significant tangent emerged around the author’s writing style: multiple people accused the blog post of being LLM-generated slop, which the author pushed back on, leading to a minor meta-debate about how much polish is acceptable in technical writing. The broader hope in the thread is that this experiment is a prelude to massive GPU cluster investments, because if Hetzner sticks to serving small single-card models, it will remain a curiosity rather than a competitor.
Hannah Fry Wins the Leelavati Prize in 2026 for Mathematics Outreach [comments]
128 points · 23 comments · www.maths.cam.ac.uk · 6h ago
Hannah Fry has been awarded the Leelavati Prize for mathematics outreach, recognizing her as a leading global ambassador for mathematical thinking. The HN thread was almost entirely celebratory, with people swapping their favorite Fry moments—her book *Hello World*, her podcast *The Rest Is Science*, and a prophetic 2018 documentary that modelled a viral outbreak in Haslemere a year before COVID hit the same town. One user shared that Fry originally landed her dream job in F1 aerodynamics but found it boring because she just wrote Python simulations and read results the next morning, so she quit and went back to academia. A few comments linked to the history of the prize name, noting it comes from Bhaskara II's 12th-century poem of arithmetic problems dedicated to his daughter Leelavati. The consensus was straightforward: Fry is an extraordinarily gifted communicator who deserved the award.
Codeberg Divides [comments]
122 points · 174 comments · lucumr.pocoo.org · 16h ago
Armin Ronacher’s blog post argues that Codeberg’s new terms—banning projects “mostly” written by generative AI—are a well-intentioned but badly-designed rule that makes the platform less dependable as infrastructure, even though he wants a democratic European alternative to GitHub. The HN crowd largely agreed with that critique, with many seizing on his “I know it when I see it” comparison to say the policy is too vague and opens the door to arbitrary enforcement driven by community politics rather than clear rules. A vocal minority pushed back hard, accusing the author of hand-waving legitimate concerns about copyright, energy use, and maintainer burnout, and insisting that open source communities are perfectly within their rights to say “no” to AI slop. Others noted this is the latest in a string of Codeberg moral-purity bans (following cryptocurrency), and worried that the platform is building a reputation for unpredictability that undermines its mission as a broad GitHub alternative. The deepest split in the comments was between people who see this as a healthy values-based self-selection and those who see it as a slippery slope that will eventually drive away anyone whose project falls afoul of shifting social norms.
Unitree As2-W [comments]
119 points · 52 comments · www.unitree.com · 15h ago
Unitree announced the As2-W, a wheeled quadruped robot that can sprint over six meters per second, carry 150 kilograms while standing, climb 80-centimeter steps and 45-degree slopes, and run for three hours on a charge. The Hacker News thread immediately split between serious hardware analysts and skeptics questioning whether the product videos are real, with some pointing out that Unitree has a track record of demoing prototype capabilities that don't make it into production units. Several people who have seen the robot in person confirmed it works, though not at the acrobatic level of the marketing footage, while others traced its lineage back to MIT's Mini Cheetah open-source project and accused Unitree of copying fundamental designs. The conversation pivoted hard toward military applications—chainsaws, rifles, thermal sensors—with arguments about whether a bomb-happy drone is actually scarier than a robot mule, and a smaller faction championed the hybrid wheel-leg design as a genuinely clever locomotion breakthrough that evolution never figured out.
SpaceX Starship Flight 13 livestream [video] [comments]
109 points · 140 comments · www.spacex.com · 9h ago
SpaceX flew Starship's thirteenth flight test, successfully launching the upgraded V3 Super Heavy booster and upper stage, deploying twenty next-generation Starlink V3 satellites on a suborbital trajectory, and demonstrating an in-space engine relight. The booster's landing burn failed to light enough engines, resulting in a hard splashdown in the Gulf, which the thread treated as a data-gathering event rather than a real failure — the consensus was that the catch tower was never in danger since the booster was aiming for the water anyway, and the flight's primary objectives were about the ship, not the booster. The big surprise was that the ship survived reentry and splashed down intact in the Indian Ocean without exploding, something that's never happened before, giving engineers unprecedented heat shield and telemetry data via Starlink throughout reentry. A split emerged on whether the heat shield looked reusable after this flight — some pointed to missing tiles and argued refurbishment is still required, while others noted the steel structure held up better than expected and the flaps didn't burn through. The thread also tangled with the broader Musk question, with one side arguing SpaceX's genuine aerospace achievements transcend the founder's baggage and the other countering that the company's AI and social media ventures are siphoning attention and capital from the core space mission.
I got into YC Startup School by hacking it [comments]
103 points · 67 comments · obaid.wtf · 13h ago
A security researcher detailed how they reverse-engineered and exploited Y Combinator's Paxel tool, a system that analyzes founders' AI coding agent transcripts and scores them across five axes, finding they could forge any score and push it to YC's database because the server accepted results without cryptographic validation that the LLM-generated scores and notes hadn't been tampered with. The HN thread largely pivoted away from the hack itself into a broader debate about whether running `curl | bash` on your machine for an application process—especially one that uploads your AI agent transcripts to YC's servers—is an insane security and privacy risk, with multiple people pointing out that Paxel uploads prompts and diffs that constitute source code, not just metadata. A significant split emerged between those who see this as yet another example of YC treating security as an afterthought in its rush to gather data on a million-plus coders, and defenders who argue the tool is transparent, optional, and that YC's incentives align with building trust rather than stealing ideas. The thread also editorialized that the entire premise is dystopian—requiring applicants to install invasive surveillance software to apply to startup school—and that YC's positive response to the researcher reflects their longstanding appreciation for system hackers rather than any real concern about the vulnerability itself.
I Tried Building a Real App with AI. It Took a Year [comments]
96 points · 79 comments · www.alexhyett.com · 20h ago
This is a first-person account of a developer who spent a year building a habit-tracking iOS app with heavy AI assistance, starting from zero Swift knowledge and ending up with a shipped product, but the real story is about the painful cleanup that followed the initial AI-generated code. The developer spent six hours with Cursor to get a working prototype, only to discover the code was a mess of thousand-line views, broken iCloud sync, and performance problems that took months of manual refactoring to fix—leading to the conclusion that AI gets you 80% there but the last 20% takes 80% of the time, especially if you don't understand what it wrote. The thread largely agrees with the article's premise, with some commenters arguing the developer's incremental approach was the wrong strategy and that a top-down spec-first method produces better results, while others push back that the real bottleneck isn't AI capability but the human trust and maintenance required to support a product long-term. A major tangent emerged around app store subscriptions and "lifetime" purchases, with people sharing stories of developers rug-pulling paid features behind new subscription models, which the article itself touched on when complaining about £44.99 for a habit tracker. The broader debate that emerged is whether AI coding is making developers dumber in slow motion, with one side arguing that relying on LLMs without understanding the output is like a frog in boiling water, while others counter that they've been cleaning up human-written slop for decades and at least AI slop arrives faster.
The case for MUDs in modern times (2018) [comments]
96 points · 73 comments · www.andrewzigler.com · 20h ago
An essay from 2018 makes the case that MUDs—purely text-based multiplayer worlds—still offer unique experiences that graphical MMOs can't replicate, like deeper imaginative immersion and accessibility for visually impaired players. The thread mostly ran with it as a nostalgia-fueled celebration, with people swapping stories about learning to touch-type under attack, finding old friends on still-running servers, and calling MUDs "formative" in ways modern games aren't. A few people pushed back, arguing that genres like roguelikes evolved beyond text for good reason, but the stronger rebuttal was that demanding graphics for a MUD is like demanding movies replace novels—they're fundamentally different mediums. A significant tangent spun off around China's gaming history, noting that wuxia MUDs like *Xiakexing* directly birthed the country's first graphical MMORPGs, a piece of internet history rarely told in English. The big lively split was over whether LLMs could revive MUDs by generating dynamic quests and dialogue, with some sketching out design ideas and others skeptical that a generation with no memory of telnet would ever care.
JEP 541: Deprecate the macOS/x64 Port for Removal [comments]
95 points · 89 comments · openjdk.org · 15h ago
Oracle is deprecating the macOS/x64 (Intel) port of OpenJDK in JEP 541, meaning builds will fail by default starting with JDK 27 unless you pass a flag to override the error, and the port will eventually be removed entirely. The thread largely agreed this is overdue, given that Apple itself has transitioned to ARM and is phasing out Rosetta for macOS apps, though several people pushed back on the idea that Intel Macs are already dead — citing Steam surveys and Homebrew stats suggesting 20–27% of Macs are still Intel in early 2026. A significant split emerged between those saying "just move to Linux on that hardware" and those arguing the performance and battery gains from Apple Silicon make upgrading the obvious choice, with one camp calling the ongoing OS performance on Intel Macs "atrocious." Others pointed out that Java 25 is an LTS with updates through 2030, so most users won't be affected for years, and that the real burden here is on Oracle's maintenance costs, not on users stuck on old hardware. A few commenters took a wider jab at planned obsolescence, saying the hardware still works fine and they resent being forced to upgrade just for software support — one replied that this is the inevitable tension between wanting new software forever and wanting to keep using decade-old machines.
Sperm Whales blow bubbles to achieve restful, vertical sleep [comments]
87 points · 12 comments · news.st-andrews.ac.uk · 8h ago
New research from the University of St Andrews reveals that sperm whales release gas bubbles to fine-tune their buoyancy, allowing them to hang motionless and vertically just below the surface to rest. The top comment immediately pushes back on the headline, arguing the whales aren’t “blowing bubbles to sleep” but venting gas to regulate specific gravity—a precise buoyancy trick any scuba diver would recognize. A reply counters that the whales literally do release air through their blowhole underwater while resting, and the study specifically connects that behavior to achieving vertical rest, so the framing isn’t misleading. The thread then devolves into jokes about whale farts causing upward acceleration and surface fatalities, with a few people earnestly stepping in to correct that sperm whales don’t eat floating carcasses—though the original jokester admits they were just riffing. One commenter also notes the paper’s authors are careful not to call the whales “asleep,” since getting EEG data on a submerged sperm whale is still an open scientific frontier.
The front end framework for correctness: built on Effect, architected like Elm [comments]
85 points · 46 comments · foldkit.dev · 16h ago
This is a new frontend framework called Foldkit that enforces The Elm Architecture pattern on top of the Effect TypeScript library, giving you a single immutable state tree, explicit side effects modeled as plain values, and a rigid update structure that promises linear complexity growth. The HN thread immediately went sideways because the homepage's interactive demo appears buggy—people clicking "Reset after 2 seconds" then "Add 1" found the reset gets silently cancelled, which undermines the whole "correctness" pitch and led to complaints about hidden imports and opaque code examples that don't match the framework's selling points. Several people argued this kind of rigid, architecturally-prescribed framework makes more sense now that AI agents do the actual coding, saying Rust-like explicit patterns become a feature when you're not hand-typing every discriminated union, though others pushed back that the same Elm-style MVU pattern failed to gain traction pre-2020 and still forces you into performance and ergonomic tradeoffs that React/Vue don't. The comments also debated Effect itself—some called it the only sane way to write TypeScript at scale while acknowledging the brutal learning curve and archaic syntax that now matters less with LLM assistance—and a few people clocked the documentation as clearly AI-generated, which the creator confirmed but defended as a pragmatic tradeoff to ship docs while promising rewrites before 1.0.0.
Generated 2026-07-25 08:03 UTC
Generated by Sauron from Hacker News discussions and linked articles.