HN Brief: 2026-06-25
Today’s Hacker News was dominated by three currents: AI hardware escalation, free-tier skepticism, and founder reckonings. OpenAI unveiled its first custom inference chip and Qualcomm bought Modular, both moves that signal the datacenter inference arms race is just beginning—but the real friction was around the Bunny.net “free DNS” announcement, which the thread quickly panned as misleading since every account still carries a $1 monthly minimum. In the founder corner, John Carmack owned his early mistakes at id Software, sparking a sharp debate about whether startup intensity permanently burns people out, while the German incorporation nightmare story split the room between those who saw bad structural choices and those who saw a fundamentally hostile bureaucracy. Threads on open-source vs. proprietary AI also ran deep, from Anthropic accusing Alibaba of distillation to Yann LeCun arguing that only open models can serve most of the world.
The thread on Bunny’s DNS offer is worth clicking for the furious argument over what “free” actually means in 2026. Carmack’s retrospective is a rare unfiltered look at the human cost of startup intensity and whether the industry ever learns from it. The Germany founding saga delivers a visceral blow-by-blow of bureaucratic friction, but the real draw is the clean split between structural critics and self-responsibility advocates. The Thomann vs. Fender legal battle is the day’s outlier—a deeply flimsy IP claim over the Stratocaster body shape, with private equity and deep-pocketed retailers lining up for a fight you won’t see in any AI thread. And the Anthropic/Alibaba accusation is worth reading not for the charges, but for the technical takedown of “distillation” as a manufactured panic.
We’re making Bunny DNS free [comments]
877 points · 260 comments · bunny.net · 23h ago
Bunny.net announced they're making their authoritative DNS service free, dropping query fees and offering hosting for up to 500 domains per account. The HN thread immediately zeroed in on the catch: every account still carries a $1/month minimum spend, so unless you already use Bunny for CDN or other services, the "free DNS" isn't actually free—several people felt burned by the marketing language. A lot of commenters clarified that this is authoritative DNS hosting (like Route53 or Cloudflare's DNS), not a public recursive resolver, and debated whether the $1 floor is reasonable or a dealbreaker for small projects. The discussion also split into comparisons with Cloudflare's free tier, complaints about Hetzner's recent price hikes, and a side tangent about how hard it is to auto-import DNS records from other providers.
OpenAI unveils its first custom chip, built by Broadcom [comments]
672 points · 378 comments · techcrunch.com · 14h ago
OpenAI announced its first custom inference chip, the Jalapeño, co-designed with Broadcom and manufactured by TSMC, aiming to cut dependence on Nvidia for running trained models. The thread quickly split between skeptics who see this as a smart long-term bet on inference costs overtaking training, and hardware realists who pointed out that the chip’s performance-per-watt claims are hand-wavy and that Broadcom is just repeating its TPU playbook for Google. A surprisingly deep tangent erupted over Apple Silicon’s GPU architecture — people argued that Apple squandered its GPGPU lead by chasing raster performance and that memory bandwidth alone won’t fix its prefill bottleneck, making local high-end inference a pipe dream until Apple re-architects from scratch. Others grumbled that naming it “Jalapeño” feels like cringey California tech branding that actively hurts keyboard portability, though most dismissed that as noise next to the real story: OpenAI is now in the custom silicon game alongside Google and Amazon, and the datacenter-vs-desktop gap is only widening.
Founding a company in Germany: €9600, 152 days and I still can't send an invoice [comments]
577 points · 705 comments · paolino.me · 19h ago
The article is a founder’s blow-by-blow of spending €9,600 and 152 days trying to incorporate in Germany, only to be blocked from sending an invoice because the VAT ID never arrived. The HN thread mostly disagrees that this is a fair indictment of German bureaucracy: the loudest argument is that the author chose an unnecessarily complicated structure—a UG & Co. KG, typically used for tax optimization and hiding investor identities—when a plain GmbH or UG would have given him limited liability in weeks for under €1,000. A vocal camp counters that even that simpler path is still slower and more expensive than what you’d get in Estonia or the UK, and that Germany’s notary requirement and cascading fees are a real drag on founders. Others note the author could have started with a sole proprietorship or GbR to invoice immediately, but that he prioritized liability protection and tax structuring from day one—a choice he owned, but that critics say he then blamed on the state rather than on his own over-engineering. The thread splits cleanly: you either see a person who made a bad structural bet, or a system that makes even a well-informed founder jump through expensive hoops for the basic right to bill a customer.
There are a few things that I look back on as my mistakes in the early days [comments]
524 points · 259 comments · xcancel.com · 16h ago
John Carmack posted a candid thread owning his early mistakes at id Software—over-ambition with Quake, pushing everyone too hard, a bad stock arrangement that misaligned incentives, and expecting level designers to also be visual artists. The HN thread seized on his admission that running people at startup intensity permanently wears them out, sparking a debate about whether that wisdom only comes with age and wealth, and whether founders who extract that level of effort then get to humanize themselves later. A strong split emerged: some argued that stress and panic produce worse quality and a death spiral, while others cynically noted the industry just cycles through young hires who haven't learned the lesson yet. The discussion also swerved into a broader critique of tech's CS/math-dominated culture, with several people arguing that the real blind spot is ignoring humanities and human factors—and that the same mistakes get repeated because leaders have no incentive to change. There was also a pragmatic nugget about needing to pair artists with designers earlier, referencing the friction between Romero and the other level designers.
RubyLLM: A Ruby framework for all major AI providers [comments]
381 points · 64 comments · rubyllm.com · 17h ago
The article showcases RubyLLM, a Ruby gem that provides a single, elegant interface for working with dozens of AI providers—OpenAI, Anthropic, Gemini, local Ollama, and others—handling chat, vision, agents, and structured output with minimal dependencies. The thread was overwhelmingly positive, with many developers reporting happy production use, but one early comment claimed you still had to set provider-specific parameters for things like temperature and effort; the author promptly replied with documentation links showing those are unified, the commenter quickly walked it back, and several people jumped in to defend the maintainer’s free work. A substantial side debate erupted around whether building in a dynamically typed language like Ruby in 2026 is wise—some argued static typing provides clearer signals for LLM tooling, while others countered that Ruby’s expressiveness and the massive amount of Rails context in LLM weights are real advantages, with Stripe and Airbnb cited as evidence that dynamic typing works at scale. Several users highlighted practical pain points—cache mismatches with xAI’s API, difficulty with observability before the recent instrumentation release, and the lack of native support for OpenAI’s newer Responses API—but the author confirmed that a major 2.0 refactor is underway to split protocols from providers, which will transparently route models to the correct endpoints and address that gap. Overall, the community treated it as a serious, well-designed tool for the Ruby ecosystem, with the usual static-typing heat but genuine appreciation for the maintainer’s responsiveness and vision.
Anthropic says Alibaba illicitly extracted Claude AI model capabilities [comments]
356 points · 601 comments · www.reuters.com · 12h ago
Anthropic publicly accused Alibaba of illicitly extracting capabilities from its Claude AI models, a claim Reuters reported as fact. The HN discussion largely dismissed this as performative geopolitics, arguing Anthropic is trying to convince the U.S. government to ban Chinese AI access while selling its own “Fable” system to the rest of the world. Several technically minded commenters zeroed in on the term “distillation,” explaining that it’s a normal, nearly undetectable byproduct of using any API—if you can see the output, you can train on it—and that Anthropic’s framing of it as an “attack” is a PR move, not a technical one. A strong current of criticism points out the hypocrisy: Anthropic itself trained on copyrighted data and settled for $1.5B, and many suspect the real audience for this accusation is defense hawks and IPO timing, not any genuine security concern.
45°C cooling design cuts data center water use to near zero [comments]
310 points · 202 comments · blogs.nvidia.com · 17h ago
NVIDIA announced that its next-generation Rubin servers will run entirely on liquid cooling with coolant temperatures up to 45°C, claiming near-zero water consumption and dramatic cuts to cooling energy by eliminating evaporative cooling towers and fans entirely. The Hacker News crowd immediately zeroed in on two practical concerns: where does all that heat actually go, and what about the humans who need to work in a building where everything is operating at 45°C? Several people with domain experience pushed back on the idea that this solves the local heat island problem—waste heat still dumps into the environment, just at lower energy cost, and one commenter linked a CNN piece about data centers cooking their surrounding neighborhoods. The district heating angle got serious traction, with multiple people arguing that 45-55°C water is actually viable for heating buildings or feeding heat pumps, though others pointed out it's borderline for Legionella prevention and requires a different distribution infrastructure. A side conversation erupted about whether the blog post itself was AI-written, with some defending it as eating their own dogfood and others calling it a depressing sign that even NVIDIA won't pay a human to write their announcements.
Slate EV truck starts at $24,950 [comments]
270 points · 409 comments · www.slate.auto · 18h ago
The Slate EV truck, priced at $24,950, pitches itself as a bare-bones, modular electric pickup with a DIY ethos and NACS charging. HN quickly dug into the company’s size comparison page and found the truck is genuinely small—like a 1985 Toyota—which triggered a heated debate over pedestrian safety: one camp argued the tall, square hood is a death trap for kids, while others countered that the leading edge is barely 40 inches and the short hood actually improves forward visibility over the F-150s and Silverados that dominate the road. Most people agreed this isn’t a Rivian competitor—it’s a feature-stripped utility vehicle for people who just want an EV truck without the luxury tax, and some grumbled about the lack of comma.ai compatibility. A long, detailed thread on EV ownership costs got into the weeds: 120V charging is enough for a 30-mile commute, real Seattle power rates make it absurdly cheap vs. gas, and the battery-life worry turned into a debate about whether Leaf-specific degradation data generalizes. One commenter went on a sprawling anti-EV rant about oligarchs and battery waste, but got shut down with recycling data and notes that modern batteries seem to outlast the cars they’re in.
NSA lost access to Mythos amid Anthropic dispute [comments]
245 points · 259 comments · www.nytimes.com · 20h ago
The NYT piece reports that the NSA lost access to Anthropic’s Mythos AI tool because a government directive forced Anthropic to cut off all users, including its own employees, after the agency demanded they block non-Americans and Anthropic couldn’t reliably verify identities. The thread immediately split into a legal argument over whether the government could just seize Mythos via the Defense Production Act—only Congress and the President have that power, so the NSA can’t simply force access. Then the conversation completely derailed into a fierce technical debate about whether the NSA can MITM or decrypt TLS traffic at scale, with one side insisting Snowden proved they already have hardware taps inside every major US cloud provider, and the other side countering that modern TLS1.3 and server-side encryption make that infeasible, calling the claim technically illiterate. A handful of commenters dismissed the whole article as Anthropic IPO propaganda, noting the NSA’s past surveillance capabilities are less relevant when companies voluntarily hand over data anyway.
Show HN: Nub – A Bun-like all-in-one toolkit for Node.js [comments]
234 points · 66 comments · github.com · 17h ago
Nub is a Bun-inspired toolchain that doesn't replace Node.js but wraps it with preload hooks for TypeScript transpilation (via oxc), module resolution, and polyfills for modern APIs like Temporal and Worker, while also shipping a faster script runner, package manager, and Node version switcher all in one Rust binary. The HN crowd was mostly impressed with the pragmatic design choice to augment rather than fork, though a vocal contingent argued that fundamental problems like the inability to document package.json config are unfixable without breaking from Node entirely — the author pushed back that you can monkeypatch those too without throwing out the ecosystem. A long thread spun off into jokes about "vibe coding" and the Bun-to-Rust-to-Zig rewrite memes, especially after someone noticed Claude is the second-largest contributor to the repo, which some found hilarious and others found unsettling for a production DX tool. Several people questioned whether this is production-ready given the polyfill attack surface and lack of a build/compile step, and the author acknowledged a `nub build` is coming but argued the overhead is negligible. There was also solid technical depth around why they use `--require` instead of `--import` (0.5ms vs 4.6ms overhead) and how synchronous `module.registerHooks()` was a big enabler for this approach.
The Xteink X4 E-Ink Reader [comments]
229 points · 131 comments · blog.omgmog.net · 15h ago
The article is a detailed review of the Xteink X4, a £40 e-ink reader the size of a stack of credit cards that you can magnet to the back of your phone. The HN thread zeroed in on the device's surprising openness—it runs on an ESP32 microcontroller and has a thriving custom firmware ecosystem (CrossPoint, Inx, Papyrix, even a PalmOS-inspired OS) that makes it genuinely good for the price, with one faction arguing that its lack of a backlight and touchscreen is actually a feature, not a bug. Others pushed back hard, saying the missing front light is a dealbreaker and pointing to Android-based readers like Boox as more versatile, though those were criticized for terrible battery life and GPL compliance issues. The marketing gimmick of sticking it to a phone got mostly eye-rolls—everyone agreed the real win is just how pocketable it is, with multiple people saying they forget it's there and it's the best reader they have on them, not the best they own. The biggest split was between the "keep it dumb and minimal" crowd and those who want more features like a light, with a consensus that if they release a version with a front light, it'll be an instant classic.
Stealing Is a Skill [comments]
228 points · 133 comments · ben-mini.com · 18h ago
This post argues that literally copying someone else’s work—down to the pixel—is a legitimate skill, using Virgil Abloh’s “3% approach” as justification for rebuilding Mintlify’s marketing site practically verbatim. HN pushed back hard on the ethics and taste of that: several people called it outright distasteful and foolish to publicly admit to, with a few pointing out the obvious copyright and trademark risks. Others defended the practice as a time-honored learning exercise—copywork in writing, jazz musicians transcribing solos, or the Kalashnikov principle of knowing your field before innovating—but the thread split sharply on whether that justifies using it for a corporate landing page rather than for personal study. A side debate emerged around the sorites paradox of copying from many sources vs. one, and whether the real value is in combining influences rather than cloning a single site, while the author’s own commenters got into a heated exchange about how hard pixel-perfect reproduction actually is, especially with LLMs.
Reid Hoffman says SpaceX 'not an AI company', xAI 'complete train wreck' [comments]
227 points · 261 comments · fortune.com · 19h ago
Reid Hoffman went on a podcast to call SpaceX’s AI strategy a “buy your way into relevance” move and dismiss xAI as a “complete train wreck” that’s on its third restart after all cofounders left. The HN thread largely ignored the substance of his Anthropic and Gen Z commentary and instead focused on whether Hoffman himself is credible—many argued he’s just a bitter rival with a LinkedIn fortune and a stake in OpenAI/Anthropic, so his shots at Musk are self-serving noise. The real meat of the discussion was on SpaceX’s IPO financials: multiple people dug into the S-1 filing’s claim of a $28.5 trillion TAM and pointed out that the rocket and Starlink businesses don’t remotely support the valuation, so the AI narrative is pure hype to justify a price that assumes SpaceX will capture a quarter of global economic output. A few commenters went further, arguing SpaceX is essentially operating as a “SPAC X” that will use its market cap to roll up AI companies like Cursor, and that the whole thing is a pump-and-dump that will contract spectacularly once buyers realize the numbers are fantasy.
PR spam today looks like email spam in the early 2000s [comments]
213 points · 123 comments · www.greptile.com · 17h ago
The article argues that AI-generated pull request spam is now plaguing open source projects (specifically OpenClaw) the same way email spam exploded in the early 2000s, and proposes reputation systems like Vouch as the fix. HN pushed back hard on the idea that most of these PRs come from people trying to help—the dominant take was that contributors are gaming GitHub for resume fodder, not altruism, and that "I contribute to open source" will soon be as worthless a signal as "I donate to charity." Several maintainers said they'd rather ban AI contributions entirely than filter them, comparing spam filters to a losing battle that still has false positives decades later. The thread also split over whether donating token credits to projects (instead of submitting slop) would work, with many noting that maintainers want cash, not tokens, and that the real problem is people who genuinely believe their AI-generated code is valuable despite lacking the skills to review it.
For most of the world, open-source AI is the only way forward [comments]
211 points · 137 comments · techstrong.ai · 17h ago
Yann LeCun made a politically charged case at the UN that open-source AI is the only path to global sovereignty and cultural diversity, warning that proprietary systems from US and Chinese big tech will concentrate power dangerously. The HN thread mostly sidestepped the geopolitical angle and dove straight into the affordability question—can “most of the world” actually run a useful open model locally? The consensus was that you need something like 32GB of VRAM, which right now costs $3k–$4k, but a lot of commenters argued that Moore’s law and cheaper hardware will bring it down to $1k–$2k within a couple years, and that models like Qwen 3.6 27B already run well on consumer gear if you’re willing to quantize and split across cards. Others pushed back hard, saying the real barrier isn’t local hardware at all—open-weight models can be hosted by third-party providers at much lower cost, so the future is probably cloud-based open inference, not everyone owning a monster GPU. The RAM shortage debate got heated, with one side pointing to historical cycles where new fabs ended gluts and the other saying this time demand is exponential and won’t let up, especially since hyperscalers are hoarding HBM for training.
Computer use in Gemini 3.5 Flash [comments]
210 points · 134 comments · blog.google · 14h ago
Google's integrating "computer use" — basically letting an AI agent see a screen and click around — directly into Gemini 3.5 Flash, making it a built-in tool rather than a separate model. The HN thread immediately split: a vocal camp argued computer use is slow, insecure, and a fundamentally wrong approach compared to building proper APIs or using bash tools, while others countered that it's pragmatically useful for automating legacy enterprise portals, UI testing, and bypassing gatekeepers. A huge chunk of the discussion wasn't even about the article's announcement — it was a coordinated complaint that the Gemini app still lacks MCP support, with people saying they've abandoned the official interface for CLI work or custom frontends. Some commenters also called out Google's benchmark chart, noting that while it makes 3.5 Flash look great, it actually loses to Opus 4.8 and GPT-5.5, and they mocked the selective axis scaling. Safety got a real airing too: people pointed out that until these agents can securely handle your SSN or bank login, the "iPhone moment" for AI agents is still far off, and that Google's own warnings about not sharing secrets undercut the vision.
Blogging can just be stating the obvious [comments]
201 points · 68 comments · blog.jim-nielsen.com · 8h ago
Jim Nielsen’s post argues that blogging doesn’t need to be groundbreaking—it can just be pointing out something obvious that nobody else is saying, like John Gruber’s rant about websites shoving cookie popups and newsletter signups in your face before you can even read the content. The Hacker News thread overwhelmingly agreed with that premise, diving deep into the “curse of knowledge” and the fact that there’s always a new cohort of people who haven’t heard a basic idea before, so re-stating it in a fresh voice can be genuinely valuable. A mathematician pushed back hard, though, saying that once you’re deep in a field, the fear of reinventing the wheel and adding noise kills any motivation to write, and he questioned whether the world actually needs yet another take on the Serenity Prayer. Others countered that repetition isn’t redundancy—it’s how ideas actually sink in, and every writer’s lens is unique, so even a “basic” post helps someone. A side argument flared up about Substack’s own user-hostile design and whether LLMs make blogging pointless or actually help you find underexplored topics, with several people groaning that AI gets shoehorned into every discussion these days.
Qualcomm to Acquire Modular [comments]
191 points · 60 comments · www.modular.com · 18h ago
Qualcomm is buying Modular, the AI infrastructure startup behind the Mojo programming language, for about $4 billion. The thread goes heavy on Mojo angst: plenty of people think Chris Lattner wasted his talents bending Mojo to be Python-compatible instead of building something clean from first principles, with one camp arguing that the Python bait-and-switch was inevitable for VC-funded developer mindshare, while others are just sad Mojo's future under Qualcomm looks murky. A strong undercurrent points out the irony—Lattner spent years writing about how hardware companies fail at AI software stacks, and now he's selling to one. Some commenters see the bigger picture as Qualcomm bulking up to challenge Nvidia in data-center inference, assembling a portfolio of bets from RISC-V cores to server-class CPUs, but the dominant take is that Mojo's initial promise is effectively dead as a standalone language.
Thomann takes legal action against Fender [comments]
190 points · 122 comments · www.thomann.de · 12h ago
Thomann, Europe’s biggest music retailer, is suing Fender after Fender used a default judgment from a German court—won against a no-show Chinese Aliexpress seller—to claim copyright over the Stratocaster body shape and then sent cease-and-desist letters to builders like PRS, Suhr, and even Thomann itself. The HN crowd quickly zeroed in on how flimsy the legal basis is: German courts don’t operate on precedent, the Strat shape has been in the U.S. public domain since 2009, and the law firm Bird & Bird is basically bullying everyone into destroying inventory they can legally sell in America. A deeper thread points out that Fender has been owned by private equity (Servco Pacific) since 1985, and the new CEO comes from luxury goods—so this looks less like protecting IP and more like injured pride that John Mayer’s PRS Silver Sky outsells equivalent Stratocasters, especially now that Fender owns Reverb’s sales data. Several commenters note that Thomann’s own cheap brand, Harley Benton, sells S-types, but the real target might be PRS at the high end, and that luthier LSL has already hired the lawyer who won the 2009 public-domain ruling against Fender. The consensus is that Fender badly miscalculated by picking a fight with a deep-pocketed, vertically integrated retailer that’s willing to take this to a real trial in Germany.
A deadly fungus that can infect cats and people is spreading [comments]
187 points · 137 comments · www.sciencenews.org · 20h ago
The article reports that *Sporothrix brasiliensis*, a fungus causing oozing skin ulcers in cats and potentially fatal infections in humans, is spreading across South America and expected to reach the U.S., carried by infected cats. The HN discussion quickly pivoted to arguments over whether humans are the real invasive species and if drastic population reduction is the only answer, with some pushing back that species migration and extinction are natural, not a moral catastrophe. Others invoked *The Last of Us*, but the thread corrected that this fungus doesn't affect brains—though its airborne yeast, 10-week surface survival, and three-year symptom delay make it uniquely unsettling. A split emerged between those downplaying the threat as mainly to immunocompromised people and those pointing out that any respiratory pathogen with these transmission traits is a serious concern for everyone. Practical pushback included calls for better testing at borders, skepticism about pet ownership in cities, and a tangent arguing that bleach and hypochlorous acid kill it easily, so panic is overblown.
Elastic lays off 7% of employees [comments]
184 points · 166 comments · www.elastic.co · 10h ago
Elastic's CEO announced a 7% workforce reduction, framing it as a restructuring to lean into AI and automation while promising to keep hiring in sales. The HN crowd was deeply skeptical, pointing out that the memo reads like an AI-generated nothingburger designed to placate investors rather than explain real business logic. Ex-employees chimed in to say the company culture has deteriorated since the IPO, and a side debate erupted over whether the Japanese model of hiring for life might be saner than the Western reflex to lay off employees whenever efficiency improves. Others argued the real story is that Elastic's move to abandon open source licensing and the subsequent chill with AWS and GCP has left the company in a weaker competitive position—so the AI excuse is just cover for bad strategic calls.
GitHub shouldn't be a dependency for publishing Rust on crates.io [comments]
170 points · 67 comments · github.com · 12h ago
The linked article wasn't available to this summarizer; from the discussion, the core complaint is that crates.io currently requires a GitHub account to log in and publish Rust crates, creating a single point of failure and a dependency on a for-profit platform. A ten-year-old GitHub issue exists to fix this, and an RFC was recently merged to unblock the work, but progress is slow because it's boring volunteer labor that no one is paid to do—several commenters push back that the Rust Foundation could fund it but doesn't, while others counter that the foundation isn't structured to hire developers. A split emerges: some argue the real problem is that crates.io shouldn't depend on *any* single internet property (including itself), while others point out you can already use alternative registries or vendored dependencies, making the GitHub dependency only an auth issue, not a distribution one. There’s also an unexpected tangent where someone gets called out for using the phrase "load-bearing," triggering a meta-argument about LLM-isms, and a few people bring up Go’s and Java’s package ecosystems as supposedly better models, only to be told they have their own flaws.
Boffin claims Microsoft’s “quantum leap” is invalid due to “basic Python errors” [comments]
166 points · 56 comments · www.theregister.com · 16h ago
The Register reports that Dr. Henry Legg published a peer-reviewed critique in *Nature* showing that Microsoft's touted quantum computing breakthrough relied on basic Python errors—specifically, their analysis code reversed an array by index instead of by physical value, and hardcoded a filter to only display the most favorable data region. HN spent a surprising amount of time debating the word "boffin" (it's a classic British Register-ism, and opinions are split on whether it's endearing or lazy), but the technical crowd zeroed in on the actual bug. Several commenters dug into the GitHub source to confirm the `x[::-1]` reversal issue and pointed out that this isn't the first time a high-profile Nature paper has unraveled over a simple coding mistake—a 2016 paper on gradient descent had a similar sign-error retraction. The broader takeaway in the thread was that Microsoft's topological quantum computing story has smelled fishy for years, and this critique (plus the convenient timing of a "Majorana 2" announcement right before the paper dropped) only reinforces the skepticism; as one person put it, the company seems to be chasing a qubit that doesn't exist yet with code that can't even index a list correctly.
Ending respiratory infections [comments]
147 points · 70 comments · blog.interceptfund.com · 6h ago
The linked article is a detailed proposal from a new philanthropic fund called Intercept, backed by $500M from Stripe, OpenAI, Bill Gates, and others, to essentially eliminate respiratory infections by developing broad-spectrum preventatives and air-cleaning technologies, drawing a direct analogy to how clean water ended waterborne diseases a century ago. The HN thread was sharply split: a top comment described a girlfriend’s death from human metapneumovirus, grounding the stakes in visceral tragedy, while several people pushed back hard on the waterborne analogy, arguing you can filter all the water you drink but you can’t filter all the air you breathe unless you walk around like Bane. The claim that healthy people spend 15-25 days a year sick sparked a major debate—parents and office workers said that sounds right or even low, especially with kids in school, while others insisted it’s wildly overblown unless you count every sniffle as “sick.” A recurring frustration was that society collectively decided to ignore air quality improvements after COVID, with some commenters noting they now mask everywhere or run HEPA filters at home, while others questioned whether the real barrier is funding or the technical strategy itself (e.g., universal vaccines that still only cover one virus family).
Half-Life 2 in a Browser [comments]
139 points · 44 comments · hl2.slqnt.dev · 2h ago
The linked article wasn't available to this summarizer; from the discussion, someone built a browser-based port of Half-Life 2 running via WebAssembly, and it actually loads and plays on phones and low-end laptops. The thread immediately split into two camps: one side marveling at the technical feat and the nostalgia of playing HL2 anywhere, the other side arguing the site is almost certainly illegal since it redistributes copyrighted assets without a license from Valve. The copyright pushback got detailed—people pointed out that Valve has never made the HL1 engine open source (contrary to a common claim), and that just because Valve gave away the game for free before doesn't make unauthorized redistribution legal. A running theme was that the port is visually glitchy (missing character eyes, broken shaders, no lip sync), which prompted a few people to list other impressive browser ports like Quake 3 and Unreal Tournament that pull off similar tricks more faithfully.
LuaJIT 3.0 proposed syntax extensions [comments]
138 points · 84 comments · github.com · 7h ago
LuaJIT 3.0 is proposing a batch of new syntax: bitwise operators, a ternary `?:`, safe navigation `?.`, compound assignments like `+=`, and a handful of other QoL extensions. The HN crowd is split—some see this as a welcome revival of a project many thought was dead, but others argue these changes push LuaJIT into being its own language, not just a JITted Lua. The ternary operator gets the most heat: a bunch of people want `if` as an expression instead, and the perennial “just use `and/or`” crowd gets reminded that that idiom breaks on `false` values. There’s also a notable thread pointing out that several of the additions (like `?.` and `??=`) come from Luau, the Roblox fork, making it easier for younger coders and AI agents to pick up. Overall, the tension is between modern ergonomics and the “just one way to do it” philosophy Lua was built on.
Big AI labs are hiring philosophers [comments]
136 points · 121 comments · archive.is · 15h ago
The Economist reports that philosophy graduates are increasingly in demand at AI labs, who hire them to tackle ethical reasoning, model alignment, and the bigger questions around consciousness and safety—the kind of stuff that makes AI less sycophantic and more willing to say "I don't know." HN immediately split: some commenters saw the article as thinly veiled PR (one even flagged the closing quote from an "AI theoretician" as a giveaway), while others took the hiring trend seriously, noting that philosophers like Amanda Askell at Anthropic actually shaped the Claude constitution. A recurring pushback was that "philosopher" here mostly means a few high-profile hires, not a flood of humanities grads replacing engineers—and that the real job might be more about justifying the labs' existing choices than challenging them. People also jumped on the SBF connection, arguing that a consequentialist upbringing by his mother (a law professor, not a trained philosopher) shows how badly that framework can go when fed bad data, which led to a wider debate about whether the labs are really after insight or just a respectable-sounding stamp of approval.
I rewrote PostHog's SQL parser, 70x faster, while barely looking at the code [comments]
129 points · 61 comments · posthog.com · 13h ago
The article describes how PostHog used multiple parallel Claude Code sessions to rewrite their SQL parser from an ANTLR-generated C++ parser to a hand-rolled Rust recursive-descent parser, achieving a 70-454x speedup by replacing generic graph-walking with explicit code. The HN thread quickly veered into a meta debate about the site's controversial design — half the comments are people complaining it broke their browser or scrolling, while others defended it as boldly different from standard vibe-coded startup pages. The technical conversation was surprisingly nuanced: some pushed back that this only works because of pre-AI accumulated parser knowledge and worried about long-term skill erosion, while others (including the author) argued it's a proper powertool approach with property-based testing and an oracle, not "vibecoding." A few people asked why they didn't just use an existing fast SQL parser like sqlglot or DuckDB's, and the author clarified they needed a custom dialect tied to ClickHouse, while others pointed out that the real magic was the testing loop — using fuzzing and production query logs as an oracle to catch every divergence.
Cloudflare launched self-managed OAuth for all [comments]
129 points · 52 comments · blog.cloudflare.com · 5h ago
Cloudflare opened up self-managed OAuth so any customer can create OAuth clients for delegated API access, rather than only a handful of manually onboarded partners. The HN crowd split hard: some welcomed it as a better, more secure alternative to shared API tokens, while others groaned that OAuth is an over-engineered mess that makes simple access management a nightmare of confusing specs and centralized identity tracking. A vocal camp worried this is another step toward Cloudflare turning into a bloated cloud platform that will eventually kill its free tier and core CDN/DDoS protections—others shrugged and said that’s just the logical expansion of their business and the free tier is a cheap funnel anyway. Several commenters dismissed the blog post as a self-congratulatory engineering diary for something that should’ve shipped years ago, and a few got into the weeds arguing whether OAuth is fundamentally a privacy risk because the provider sees every site you log into.
Minimus container images are now free [comments]
123 points · 75 comments · minimus.io · 20h ago
Minimus, a company selling hardened container images with minimal CVEs, just made their entire catalog of over 1,200 images free to pull without login. The HN thread immediately turned into a comparison with Chainguard and Docker Hardened Images, with people pushing back on why they'd switch—the CTO jumped in to argue their distroless base, FIPS images, and no-auth-as-in-beer approach are genuinely differentiators, though some commenters weren't buying it. A big split emerged around trust: several people cynically predicted a future rug pull or bait-and-switch, while others pointed out you can just pull-through cache the OCI-compliant images to insulate yourself. The lack of Dockerfiles and mobile-blocking also got called out as transparency and UX failures, even as the founders promised to consider fixes.
Generated 2026-06-25 08:33 UTC
Generated by Sauron from Hacker News discussions and linked articles.