0810 | Amazon's AI Data Center, Kids' Online Safety, CSS Bombs, and Code Provenance

Mia: Welcome back to HackerNews Daily. I'm Mia, and I'm joined by Milo.

Milo: Hey everyone, good to have you. Today we've got a packed lineup, starting with a surprise in California — an Amazon AI data center that residents say just appeared on them.

Mia: We'll also dig into a new Illinois law targeting kids' social media, a researcher finding dangerous tricks hiding in your inbox, and a tool that tracks words back to their original source.

Mia: Tom's Hardware reports that residents of Gilroy, California were caught by surprise when an Amazon AI data center started construction in their city. The headline frames it as Amazon secretly circumventing a community vote for a massive AI data center using 45-year-old rules, and the article's tagline is simply, surprise, it's a data center. Negotiations started back in 2020, public comments stayed open until 2024, and the project reportedly flew under the radar. One resident, Rosa Rodriguez, wanted to raise concerns about the damage a data center can do to a region already suffering from drought, but she was told the comment period had already lapsed back in 2024.

Milo: The big question on Hacker News is whether any actual circumvention happened. One commenter, who discloses that Amazon pays his salary though his job has nothing to do with data center construction, says the article never actually identifies a circumvention or a loophole. He reads it as Amazon following the correct city permitting and legal processes years ago, and argues approvals have to become final at some point. He also suggests residents are partly complaining because data centers are this year's culture war hot topic.

Mia: Others agree. One notes the site went through a long approval process with required public notices and comment periods, and asks whether the supposed circumvention is simply that comment periods are time-boxed. He concedes residents may not have known the negative implications back in 2024. And another commenter goes further, saying there's no loophole at all, because Gilroy has a law on the books that M2-level industrial projects, including data processing, are approved unconditionally.

Mia: It's FOSS reports that Illinois Governor JB Pritzker has signed HB5511, the Children's Online Social Media Safety Act, making Illinois the latest state to pull operating systems into the age verification picture. The framework applies to covered manufacturers — device makers, operating system providers, and app stores — for every device sold or used in the state.

Milo: So how does it work in practice? By January 1st, 2028, these companies have to build an interface at account setup that asks a parent or account holder for the child's birth date, and that gets converted into one of four age brackets: under 13, 13 to 15, 16 to 17, or 18 and up. Apps then pull that bracket through an API on download or launch. A minor signal gives apps actual knowledge, which triggers required safety settings — feeds get restricted, profiles hidden from adult strangers, adult messages blocked, precise location masked, and notifications go dark between 10 PM and 7 AM.

Mia: Parents can override those defaults, and minors over 16 can override them themselves. Apps have to start requesting signals by July 1st, 2028, and violations can cost up to 50,000 dollars each, enforced by the state Attorney General. Good-faith device manufacturers and apps aren't liable if a signal turns out wrong. The bill passed 82 to 27 in the House and 57 to 0 in the Senate. Notably, the article points out that Colorado and California have similar OS-level age signal laws, but both were amended to exempt open-source distribution — and System76 founder Carl Richell worked with Colorado lawmakers on that.

Mia: PortSwigger researcher Gareth Heyes has published a paper titled CSS: the bomb inside your inbox. The core argument is that webmail clients routinely render untrusted CSS inside a trusted interface and rely on CSS sanitization, but attackers can create discrepancies between what the sanitizer thinks is safe and what the browser actually renders. He says he'll show how to break out of trust boundaries, exfiltrate tokens, compromise third-party websites, and steal passwords.

Milo: The client list reads like the state of webmail itself — Yahoo Mail, AOL Mail, Fastmail, ProtonMail, Gmail, Outlook, and in the targeted-list section, OpenAI's Atlas and Firefox. One commenter's reaction was simply, oh that's all, is it? But there's real substance here. Heyes found that HTML label elements can target any form element with an id attribute through the label's for attribute, inheriting that element's click action. He says sanitizers missed this technique in at least three webmail clients.

Mia: He also found a real bug in Outlook that would let an email control Outlook's UI, including pinning messages. The paper's table of contents reads like a laundry list of attacks: CSS-sanitization bypasses, image-proxy bypasses, tracking whether an email is viewed in Fastmail and Gmail, displaying a recipient's IP address in ProtonMail, combining an image proxy bypass with indirect prompt injection, CSS mutation, defacing Outlook, CSS hotwiring in Fastmail, and future attacks including an HTML only keylogger and a Chrome real time keylogger.

Mia: Finally, there's a GitHub project called us-vs-them, which tackles what's becoming a central question in the age of agentic coding: who actually wrote this line of text, us or them? It's line-level provenance derived from version history, not from markup in the file, and it works as either a library or a CLI tool.

Milo: The README makes a strong argument for why this matters. With agentic editing, text a human wrote or edited should be treated as close to sacred — an agent should hesitate and have a very good reason to touch it. But slop another agent has produced is completely up for grabs. So the tool runs a diff against version history and outputs what it calls islands of human-authored lines inside a sea of machine-generated text. The key constraint is that text shouldn't need to be marked up specifically for this — the tool just leverages that each new version of a text is created under identifiable authorship.

Mia: The authors want to track meaningfully coherent pieces of text rather than individual lines, factoring in joining, splitting, and dilution of authorship without converging to all sea or all island. In the discussion, one commenter says his team uses annotations for that, but even if agents ignore them one time out of a hundred, they still see it happen. Another says his team does the same thing, but also flags when a line was AI generated but modified by a human. The real takeaway here is that as agents write more of our code and docs, tools that can tell us which parts deserve careful human eyes are going to become a bigger and bigger part of the workflow.

Mia: Let's start with the FCC's proposal to ban LiDAR-equipped foreign drones from US airspace, which it frames by classifying LiDAR as military-grade technology. Beyond the high-end sensing packages, the proposal could also hit thermal imaging models and the swarms used in drone light shows. DJI, the big Chinese drone maker, pushes back on that label directly, calling it a category error.

Milo: Hacker News readers split most sharply on what's actually driving this. One commenter asked whether it's simply another move to remove competition, and another agreed, saying that seems to be the whole point of this country these days. But a counterargument is that the real aim might be avoiding handing China yet another future-relevant technology, given that China already manufactures the bulk of the world's solar panels and wind turbines and is heading for electric vehicles.

Mia: Someone else added that China has been leading in drones for a very long time, and it's not something they took from the US. On the practical side, one reader argued the effective approach would be subsidizing or encouraging local manufacturing at competitive, non-luxury prices. And another questioned whether US companies can even compete, noting US firms don't sell consumer drones at all, and asking whether people would be forced to buy four-thousand-dollar enterprise models for their kids to play with.

Milo: Then there's the question of whether any of this changes what China does. One reader pointed out the ban won't stop China from developing LiDAR, and China already has two market leaders in that technology. Another made the point that the rest of the world is generally happy to buy reasonably priced Chinese tech, citing the suggestion that the US could destroy a smaller economy overnight as evidence of a US-centric myopia. On the risk itself, one commenter asked how LiDAR is a national security risk at all, arguing it's easy to build.

Mia: From drones to a different kind of tool, there's a project on Hacker News called OpenChamber, described as an agentic development environment for AI coding. The idea is you run AI agents across your desktop, a browser, your phone, and inside VS Code, so you can watch the agents work, review the diffs they make, branch sessions, and keep the board visible.

Milo: Some of the features that stand out: session goals, where an agent keeps working toward a finish line even when the app is closed; running one task across up to five models at once and keeping or fusing the best results; a walkthrough of the changes, preview, a GitHub issue-to-pull-request flow, and scheduled prompts running on a cron timer. It ships desktop apps for Mac, Windows, and Linux, a browser version, and a native mobile app in beta.

Mia: Privacy is a big part of the pitch. The claim is nothing about a user's work is collected or sent anywhere, code and session content stay on the machine, and the code is open source so the privacy model can be inspected. It runs on the OpenCode SDK today and is free to use. The footer does note it's an independent project, not affiliated with the OpenCode team, and one commenter pointed to its MIT license.

Milo: In the discussion, one reader noted it looks somewhat like another tool called Paseo, but specialized for OpenCode under the hood. Someone asked whether sessions can run on a remote server so tasks keep going after you close your laptop, and a commenter replied yes. Another reader said they'd prefer Paseo for flexibility across different harness and model combinations.

Mia: There's an article from Laurentiu Raducu about how he uses LLMs to learn complex topics, and the twist is he doesn't just ask a model to explain something. Instead, the method is built around LLM-generated simulations. The flow goes something like this: ask a model to build foundational knowledge on the topic, have it review that knowledge for accuracy, then ask it to build a simulation of the topic in a low-poly, RollerCoaster Tycoon-style animation, add user experience controls, push it to a new repository, and publish it as a web page.

Milo: His claim is the result is a beautiful animation that's fully accurate and free of hallucinations, and that it works better for him than reading endless material or digesting a bulleted list from a language model. His worked example is ChipTycoon, which follows a chip from the moment sand is collected to when it's finalized and delivered to a data center. He's also built similar pages on how rocket engines are made, how LLMs work, how F1 engines are built, and how an EUV machine is built.

Mia: Now, the comments pushed back on what the post really is. One reader said the title doesn't represent the content, arguing the actual method is having an agent describe the problem area and then implement a simple web-based simulation game, and the author learns the topic and its constraints by playing it. Another commenter countered that the whole game could be generated by the model too, which circles back to how much of the learning is actually happening.

Mia: Finally, there's a video shared on Hacker News from the myNoise site, titled I Made Tinnitus My Friend, Then It Disappeared, from the site's founder. Tinnitus is that persistent ringing in the ears many people live with, and the framing of making it your friend drew some of the strongest reactions in the thread.

Milo: One commenter pushed back hardest, saying people with tinnitus tend to get outright angry when told to just change their attitude toward it, because it's not that easy. They were also surprised there's still no effective drug, noting that lidocaine works, but only very temporarily. Their own tinnitus went from a five out of ten to about a one out of ten over ten years, so they only hear it in an extremely quiet room while trying to sleep, but they insist it's not their friend and never will be, reading it as an annoying side effect of nerve damage.

Mia: Another commenter made the opposing case directly: they have tinnitus, changed their attitude, and made it their friend, arguing the way is open or you can continue to see it as aversive. But a commenter at thirty or so years with tinnitus still at a five out of ten said it was the first time they'd ever heard the change your attitude advice, comparing it to telling someone with depression to just be happy.

Milo: Experiences varied widely. One person has dealt with tinnitus for at least thirty-five to forty years, describing it as high pitched and varying in intensity. The real takeaway here is that there's no one-size-fits-all experience, and the advice to simply befriend the ringing lands very differently depending on how severe and how long-standing the condition is.

Mia: There's a study doing the rounds on Hacker News that tries to put a name to something a lot of people in the industry have suspected for years. Tim Weiss, from Imperial College London, and Nevena Radoynovska, from Emlyon Business School, published in the journal Organization Science what they're calling "Criminal Deception in Silicon Valley." They looked at court data on Silicon Valley ventures and founders prosecuted for fraud between 2000 and 2023.

Milo: So what did they actually find? I imagine it's more than just "founders lied."

Mia: Exactly. They argue the entrepreneurs engaged in what they call "façading" — constructing, performing, and protecting illusory appearances that project high-growth performance outward while masking how the venture was actually doing underneath. And they break it into three forms depending on how wide the gap was between what audiences were led to expect and what was really going on internally: surface, reinforced, and deep façading.

Milo: And I'm guessing this has real consequences for how regulators and investors should actually operate, not just an academic taxonomy for its own sake.

Mia: Right, they propose expanding SEC surveillance and its whistleblower program, reworking how investors do due diligence, and adding entrepreneurship education that clearly marks the line where founders cross into criminal territory. The discussion in the comments was split though. One commenter argued a lot of the numbers startups cite are basically soft — a user count given to TechCrunch doesn't really tell you how it was calculated — but drew a hard line at building a database of millions of synthetic users. They cited the startup Frank, which sold to JPMorgan before its founder was eventually prosecuted. But another commenter rejected the "fuzzy math" framing entirely, saying the metric exaggerations are a different thing altogether.

Milo: Separately, there's a discussion around a TechCrunch interview with historian Jill Lepore, whose headline is that Silicon Valley misreads science fiction and undermines democracy. Lepore is a Harvard historian and New Yorker staff writer who recently won a Pulitzer for her history of the U.S. Constitution. Her next book is "The Rise and Fall of the Artificial State."

Mia: And she's arguing that the artificial state is replacing the liberal democratic nation-state — a return, in her words, to tyranny and mystification, but in the form of rule by algorithms, corporations, and machines.

Milo: She's careful to say she isn't anti-technology though, right?

Mia: Right, she says her beef is with private corporations taking on the functions of the state, something she argues has only in the last twenty to twenty-five years become a deliberate effort at usurpation. She also notes many nineties futurists were libertarians who wanted the internet to do away with the nation-state entirely, and she called Elon Musk a bad science fiction reader — saying the stuff he likes actually contradicts his own political beliefs. The conversation also touched on Apple's famous "1984" Macintosh ad, why it's "bananas" to call Twitter a digital town hall, and the current backlash against data centers.

Milo: And on Hacker News, some people rejected the whole premise. One commenter said it doesn't misread anything because it doesn't read at all. Another argued most of it isn't deeply philosophical — that tech leaders got extraordinarily rich and now worry more about the concerns of the extraordinarily rich.

Mia: Switching to something a bit more introspective for the community itself. Hacker News is discussing "The Hacker's Renaissance," a manifesto published in Phrack issue seventy-two, released in August 2025. It looks back forty years to the original 1986 Hacker's Manifesto and argues the hacker spirit has been commodified.

Milo: Right, the claim is that bug bounty platforms monetize curiosity into CVEs and t-shirts, CTFs have become resume-building exercises, exploits get embargoed rather than shared. And the manifesto's framing is that the underground never died — it just moved to quieter places: encrypted chat rooms, dead-drop Git repositories, and onion forums.

Mia: And the piece calls itself not a eulogy, but a reckoning. Except a lot of the discussion isn't about the argument at all — it's about whether a machine wrote it.

Milo: Right, I saw that. One commenter said it "sounds like ChatGPT." Another quoted the article's own line — "this is not a eulogy, it's a reckoning" — in full agreement. And then opinions split hard on whether the method even matters. One commenter said they should have used Claude, that GPT is embarrassing, but that content counts and not the method used to create it. Someone else reversed it: the content sucks precisely because it was written with an LLM.

Mia: And another commented that they can barely read it anymore because they've consumed so much machine-generated text that their brain just skims it now. There's even a side thread hunting for style giveaways — like a single character the original 1986 piece never used appearing all over this one, maybe an em-dash. Although someone pointed out the original Phrack was pure seven-bit ASCII, so any non-ASCII character would be a giveaway.

Milo: Finally, a sobering story out of New Zealand about music journalism — and what happens to a scene when the coverage collapses before the venues do. It's a feature from the outlet Propel titled "The sound went quiet: how New Zealand lost its music media, and what we're building to replace it." It opens in late June with two announcements on Karangahape Road: Flying Out, the record store that anchored the Auckland strip for eleven years, said it was done, and the basement venue Neck of the Woods, which had launched a decade of electronic music careers, posted that after eleven years of fighting mounting debt it had reached the end.

Mia: But that one at least had some hope in it. Nearly three thousand people, including the band The Beths, put a hundred fifty thousand dollars into a crowdfunding page in under a week, and the venue reopened in July.

Milo: Still, the piece lists a string of other closures — a venue going into liquidation in April after thirty-four years, another closing in February, one the year before, and a festival cancelled in May for the foreseeable future. And it asks pointedly why a thirty-four-year-old venue went under owing close to seven hundred thousand dollars.

Mia: And the core argument is that music journalism collapsed before the venues did. A local music journalism award winner laments that the number of music journalists left in the country can be counted on one hand, with no full-time music critics at any major publication. Research put arts and culture at around thirteen percent of total media coverage against roughly twenty-five percent for sport, and barely three percent for art forms outside film, music, and television. Internationally, Pitchfork was folded into GQ back in 2024 with its editor-in-chief laid off. The piece's answer to all of this is what they're building to fill the gap.

Mia: Let's start with a blog post from Shayon Mukherjee, dated August 8th, with the argument that every fast write in a storage engine just moves work somewhere else. His core point: before a database tells a client that a write succeeded, it has to decide what it's actually guaranteed. Returning after copying bytes into memory is fast, but a crash can lose the write entirely. Waiting for a local SSD to sync is slower but survives a process or kernel crash. Waiting on a network volume or on several database servers to each save their own copy survives even more failures, but you pay for it with extra network time on every single write.

Milo: And the tricky part he flags is that the same system call, fdatasync, can be used whether the database is waiting on an SSD sitting right in the database host or on a remote network volume exposed through a similar interface. So an NVMe latency figure tells you very little unless you also know what acknowledgment point it's attached to.

Mia: Right. For scale, Amazon describes its S3 Express One Zone storage as having single-digit millisecond latency and claims it's up to ten times faster than S3 Standard. Turso, meanwhile, measured a four-kilobyte write at about 6.4 milliseconds on average and 7 milliseconds at the 99th percentile. Compare that to a local disk sync at one millisecond, and the local path is roughly six times faster. If your local sync is near one-tenth of a millisecond, that gap widens to somewhere around sixty-four times.

Milo: Though he's careful to note that gap isn't a property of every device or every object store. And the newer architecture he highlights uses immutable sorted files in object storage, with a local SSD handling the write-ahead log or cache, all organized in an LSM layout. The key takeaway being, once you've sent a successful write to the storage service, the visibility and durability of that write now belongs to the service, not to your database. Every layer just shifts the responsibility one step further out.

Mia: Sticking with the Hacker News feed, there's a developer post worth unpacking. Terry Godier, who builds a calm RSS reader called Current and makes visual essays about technology, posted a mea culpa on August 9th. Last week he launched Dark Hours, a utility that gives you a sense of what you could see in the sky that night.

Milo: And a developer who built another web app, DarkHours.app, replied on Bluesky to point out how similar the projects were — including the name.

Mia: Godier says he initially promised to significantly differentiate the features, change the name, and write a blog post. But about an hour later, he realized the app he'd launched using Claude was strikingly similar to that open source project — down to reproducing a bug the other developer had already fixed. So he redirected the domain to the developer and killed his plans to launch an iOS app.

Milo: He apologized for what he called irresponsible use of AI, said he'd genuinely never seen the other site before that day, and promised not to use AI that way to create web stuff again. The reaction from commenters is split. One praised him for taking responsibility and doing the ethical thing. Another replied with some skepticism, and a third called the account implausible, suggesting the AI supposedly made him plagiarize a whole project down to the name and lie about the review process.

Mia: And there's the lighter side too — one commenter joked that the dog plagiarized their homework. The thread captures the real tension here: when tools make copying that easy, the line between accidental similarity and willful reuse gets very hard to draw, and the community is genuinely unsure which side this story lands on.

Mia: And that brings us to the end of today's roundup. We covered a lot of ground, from an Amazon data center catching a California town off guard, to Illinois signing new online safety rules for kids, to a fresh look at how hackers are framing their craft in a decades-old underground magazine.

Milo: We also dug into a haunting piece about how New Zealand lost its music media, and a quieter story about one developer who turned tinnitus into a companion—then watched it fade. If any of these threads grabbed you, the show notes have the original articles linked.

Mia: Thanks for listening. We'll be back next time with more from across the web's deepest corners. Until then, stay curious.