0803 | Wikimedia Union Snub, Treasury Yen Intervention, DROP, eBay Harassment Case

Mia: Welcome back to HackerNews Daily, part of the Bri podcast family. I'm Mia.

Milo: And I'm Milo. Today we're starting with a big one — the Wikimedia Foundation's decision to pass on voluntarily recognizing a union, plus a historic move from the US Treasury on the yen.

Mia: We've also got California's new opt-out platform going live, a wild look at eBay's harassment case, and a jump in European EV sales.

Milo: And from the world of community projects, there's Kakehashi and Syncular, plus a colorful debate about AI testing that starts with an SVG of a frog.

Mia: All that and more, right after this.

Mia: Let's start with the lead item in this week's Wikipedia Signpost. It reports that the Wikimedia Foundation has declined to voluntarily recognize its staff union. The U.S. branch, Wiki Workers United, formally asked for recognition on the twentieth of July, following a similar request from the union's U.K. branch back in June. The U.S. branch set a response deadline for Friday the twenty-fourth, during the Wikimania conference in Paris, but no response came by that date, even though unionization was discussed at the conference.

Milo: So what did the Foundation eventually say?

Mia: On the twenty-seventh, it published a statement declining voluntary recognition, arguing that a secret-ballot election run by the National Labor Relations Board is the appropriate path forward, saying that process protects individual choice and ensures any outcome reflects the collective will of eligible workers. The union's response came the next day: it announced it would file for an NLRB election, but not before calling the Foundation disingenuous. Wiki Workers United said thousands of workers across tech and nonprofits have won recognition through the card check method, and accused the Foundation of sending out statements wrapped in what it called classic union-busting rhetoric, which it says was certainly supplied by expensive union-avoidance law firms. The Signpost's own headline goes further, declaring the Foundation hires a union-busting law firm.

Milo: And what about Jimmy Wales?

Mia: On the Wikimedia-l mailing list, Wales said he examined the union's organizing process, which produced a supermajority of — well, that's where the draft trail cuts off. But the core picture is clear: the Foundation is pushing for a formal election, and the union sees that as a delay tactic dressed up in corporate language.

Mia: Over on Hacker News, a Financial Times article is drawing heavy discussion under the headline, US Treasury undertakes historic intervention in the yen market. Interesting wrinkle though — the article content itself is largely locked behind a security verification page, so much of the conversation is happening from what commenters can piece together about how such an intervention intersects with U.S. Treasuries and other central banks.

Milo: So the meat of it is in the commentary, not the piece itself.

Mia: Exactly. One commenter made a point the article apparently missed: that Japan would almost certainly have sold U.S. government bonds to raise the cash to prop up the yen. They then asked a sharp follow-up — if the central banks sold euros instead, might that force European banks to do the same to defend their currency?

Milo: And what do others think about that?

Mia: There's disagreement. Another commenter pushed back on the idea that the European Central Bank would intervene at all, arguing a weaker euro isn't a cure-all, but it does help local industry compete against China. A third commenter added that eurozone countries actually hold even more U.S. bonds than Japan, and could sell them if they chose to — but probably won't, for monetary reasons. The thinking there is that the EU doesn't mind the euro drifting down a bit since it helps domestic manufacturing, and in any case the euro appears to have gained against the dollar since the intervention, not weakened.

Milo: So what would a coordinated Treasury sell-off actually be for?

Mia: That same commenter suspects it wouldn't be about defending the euro's value at all, but about forcing a political concession from the U.S. Meanwhile, on Japan specifically, they note the country holds a huge pile of U.S. Treasuries and was possibly weighing a mass sell-off to fund the yen defense. And they flag that Treasury yields are already sitting at dangerously high levels for the U.S. — which is exactly the pressure point a sell-off would push on.

Mia: Switching gears to privacy, California's Delete Request Opt-out Platform — DROP for short — becomes enforceable starting August first, according to an NBC 7 San Diego report. That's the date registered data brokers must start honoring deletion requests submitted through the portal. Since January first, roughly three hundred fifty thousand Californians have already signed up.

Milo: And the pitch is that one request replaces reaching out to each broker individually.

Mia: Right — about six hundred registered brokers. A single request cuts through all of that. Brokers get forty-five days to process each submission, and the clock resets for every new request. Noncompliance carries fines of two hundred dollars per day, per affected Californian. And Cal Privacy tells NBC 7 Responds it has already fined a dozen data brokers tens of thousands of dollars each — those were for failing to register in the system in the first place, before the deletion obligations even kick in.

Milo: So this is real leverage. Any caveats?

Mia: Quite a few. Only California residents qualify, and the tool can't erase everything about you that's online. The agency stresses that people should still watch for scams and take other protective steps. Cal Privacy's Tom Kemp put the problem in plain terms: personal information gets packaged into digital dossiers and sold, historically for targeted advertising. His example was that shopping for red shoes makes the red shoes follow you for the next three weeks. He called the manual alternative — contacting six hundred companies yourself — impractical, and he noted hackers also lean on this data.

Milo: And there was a Hacker News thread floating an idea off the back of this — someone suggested building a service that automatically sends out deletion requests on your behalf. It's a natural next step given how tedious the manual version clearly is.

Mia: Finally, a Hacker News thread digging into the Financial Times story, Crush this lady: how an eBay harassment campaign led to a fifty-six million dollar payout. The focus is on what happened to eBay's security staff versus the executives who dodged criminal charges. Citing the Justice Department, one commenter laid out the sentences: the former senior manager of special operations drew time served, a year of supervised release with no contact with the victims, and a twenty-thousand-dollar fine. The former senior director of safety and security got fifty-seven months. The former director of global resiliency got twenty-four months. A former senior manager of global intelligence got twelve. A former senior manager of security operations got eighteen months plus a year of home confinement. And three more — including two former managers of global intelligence and a contract analyst — each got a year of home confinement.

Milo: So the security team took the fall.

Mia: All seven were criminally charged and sentenced to prison, home confinement, or time served. Meanwhile, another commenter pointed out that the executives largely moved on before charges were even filed. The chief communications officer and the CEO both left eBay in September of 2019. The chief communications officer later took over the Boys and Girls Clubs of Silicon Valley. The CEO has sat on General Motors' board since 2018 and now runs an AI start-up. Another executive stayed at eBay until 2020 and now serves on Prudential Financial's board.

Milo: That contrast — low-level staff behind bars while the people at the top land on blue-chip boards — is exactly what seems to have chafed people in the thread. It's a striking reminder that the leash of accountability can stop well short of the corner office.

Mia: We start with a striking figure for European electric vehicle sales. A report says battery-electric cars jumped 50 percent and reached a 26 percent market share across Europe. Now, some of the discussion around that headline zeroed in on the numbers behind the numbers. One commenter asked what the sources were and, more to the point, how "Europe" was being defined in the first place. The answer given was straightforward: these are vehicle registrations.

Milo: That distinction actually matters a lot. Measuring registrations rather than, say, deliveries or production can make a big difference in how the trend reads. So when someone sees "50 percent jump," the honest follow-up is exactly the one those commenters raised. A 26 percent share of new registrations for battery-electric vehicles is substantial, but the definition of which markets count as Europe will shape how meaningful that headline really is.

Mia: Next up, a Hacker News discussion built around a Financial Times headline that German carmakers are flooding the jobs market with managers after wielding the axe. The actual article was behind a login wall and only showed a security verification page, but one commenter shared an archived copy in the thread. Early reactions were, well, a little playful. One person joked about the carmakers "dual-wielding," and someone answered, "only handaxes." But the conversation quickly got substantive.

Milo: The real debate was about why German carmakers are struggling and where they go from here. One commenter pointed to ground-level observations from the UK, which has no tariffs against Chinese car manufacturers. He says there are loads of electric BYDs and Jaecoos — which he nicknamed the "Temu Range Rover" — driving around lately, and he cautions that this isn't a scientific study. His take: Chinese makers are building cars people actually want to buy at a price they can afford, while he doesn't see Toyota making anything he wants at any price.

Mia: Another commenter argued China is pumping state subsidies into car manufacturing and can therefore outcompete any carmaker anywhere in the world, with leading battery technology in Asia. His forecast for the German industry: it won't die completely, because the EU can protect its market from Chinese manufacturers through regulation, but it will shrink and transform.

Mia: Then there's a hardware hacking story about rooting a TP-Link router, the TL-841N. The original blog post itself was blocked and couldn't be read, so the substance here comes from the comment thread. One commenter notes the 841N does work with OpenWRT, but as far as he remembers, it's pretty much an end-of-life device with not much flash memory or RAM. Another commenter agreed, conditionally, saying "if this thread is about the same model, then yeah."

Milo: But a couple of commenters pushed back on judging the whole project by the hardware. Their point: the goal here isn't really the router itself. The idea is to get into hardware hacking, and for that, the hardware is kind of irrelevant. What matters most is having a sane serial console exposed, decent documentation, and videos walking you through the process. So even if this particular router is old and limited, it can still be a perfectly good gateway into learning how firmware, credentials, and root access actually work.

Mia: And the thread had its share of tangents too. Someone recommended a serial terminal tool called tio as a featureful alternative to the more common one, though another commenter noted it wasn't available in his particular system architecture. And at some point someone claimed the real problem was clearly that the router firmware wasn't written in Rust — which, given this is Hacker News, is probably a joke.

Mia: Finally, a project called Kakehashi — an experimental piece of software that aims to let you run macOS binaries on Linux running on ARM chips. It's presented as a command-line-first translation layer with no just-in-time compilation. Instead, it loads the macOS executable format on Linux, maps in a standalone system library, translates the underlying system calls, and actually gets real programs running — things like 7-Zip, curl, and multi-threaded workloads.

Milo: The performance picture is honest and early-stage. 7-Zip passes multi-threaded compression tests on a tree of 8,000 files, but it's currently about 5.2 times slower than running natively on Linux. Curl has more than 200 commands and options passing automated tests. The author's biggest goal is full support for Apple's Xcode developer tools, including building iOS apps, plus the macOS version of Homebrew — while acknowledging there's a lot of work ahead.

Mia: Commenters drew the obvious comparison to Wine and Proton — the tools that let Windows software run on other systems — and asked whether the author is familiar with an existing project called Darling that does something similar for macOS apps on Linux. The honest takeaway here: this is a prototype, the performance gap is real, and it's years away from being a drop-in replacement. But translation layers like this have a strong track record of starting rough and getting impressively usable. Kakehashi is worth watching.

Mia: A developer named quambo is showing off Syncular, a project he's spent the last year building, and he traces it back to something he tried in 2019. That earlier project, called debe, used those conflict-free data structures and multi-master sync, but it never became something he could confidently ship. The big lesson for him was that convergence is actually a small part of the hard problem. The truly difficult parts are things like making writes durable, handling authorization changes, dealing with lost acknowledgements, upgrading schemas, bounding the initial bootstrap, and explaining what happened after something fails.

Milo: So what's the architecture that's meant to actually stand up to all that?

Mia: Every client runs a real SQLite database, with sqlite-wasm in the browser and native SQLite everywhere else. Writes apply locally and are written into a durable outbox inside the same transaction. Then a server-authoritative commit log validates and orders them using explicit scopes resolved in the application backend. The sync protocol itself is called SSP2, currently a draft, and it ships with a canonical binary encoding and golden byte vectors so different implementations can be checked against the same ground truth.

Milo: And that's the part where the independent client cores matter. There are two, one in TypeScript and one in Rust, both running against the same 95-scenario conformance catalog. The Rust core is exposed through Swift, Kotlin, Flutter, React Native, Tauri, plain Rust, and a small C API. On the feature side you're looking at realtime WebSocket sync, resumable bootstrap segments, explicit conflict evidence, authorization revocation with local purge, windowed replicas, blobs, optional CRDT columns, and per-column encryption, plus typesafe generated queries for five languages.

Milo: It's worth being clear about the boundaries though. This is pre-1.0 and self-hosted. There's no managed service and no peer-to-peer mode, and native packaging maturity varies a lot by ecosystem. In the comments, one person said it reminded them of Lotus Notes and that they'd love offline-first to be that kind of standard experience.

Mia: On Hacker News, a developer going by projscope introduced Meshdiff, and it exists for a very familiar pain point. Clients keep sending files with names like part v3 FINAL dot STL, and there's no quick way to see what actually changed between versions. Meshdiff solves that the obvious way: you drop in two STL, 3MF, or OBJ files, and a voxel diff renders right in the browser. Green marks added material, red marks removed, with an adjustable tolerance between a twentieth of a millimeter and a full millimeter, plus volume delta metrics.

Milo: And the privacy angle is the headline here, right?

Mia: Exactly. There's zero upload to any server. Everything runs client-side in a Web Worker. The stack is React with TypeScript, React Three Fiber, and a custom voxel diff engine. And the choice of engine is deliberate — they avoided Manifold and CGAL because those assume clean CSG topology, while real-world STL files are triangle soup.

Milo: The author asked people to try breaking it with their weirdest files, and the comments were receptive. One person said they needed something like this literally yesterday. Another found it useful for comparing 3D model generators and suggested a CLI or CI integration so diffs could be generated for later inspection. Several comments converged on syncing the viewports so all three rotate together, with one person noting that should be the default behaviour. There were also concrete bug reports — multi-group OBJ files get ignored with only the first group visible, and someone suggested warning when the same file is uploaded twice. And another commenter said the post should be properly labeled Show HN, citing the site's guidance.

Mia: This one's a fun bit of history, courtesy of The Public Domain Review. Anne Lawrence-Mathers has written an essay on the Ars notoria, an anonymous thirteenth-century work that promised instant knowledge — the sort that would travel rapidly through diagrams, incantations, and arcane rituals to transmit everything a scholar might need. There are fifty-six extant manuscripts, and the diagrams apparently worked almost like religious icons, supposedly offering contact with benign supernatural forces.

Milo: And it didn't survive the scrutiny of a certain famous theologian.

Mia: Right. Thomas Aquinas condemned it by name in his Summa theologiae, concluding it was unlawful and futile. Futile because it can't deliver what it promises, and dangerous because its signs — neither understood by humans like ordinary words, nor sent by God like sacraments — could lure people into contact with demons. The Yale University Library copy is accepted as the earliest known manuscript version, likely produced for a university master, probably at Bologna.

Milo: What did the Hacker News crowd make of the AI angle in the title?

Mia: Mixed. One commenter called the title very tenuous or baity but said the piece was a fun read. Someone else countered that the book genuinely does promise intelligence acquired through non-natural means, so the word artificial does fit — possibly as a tongue-in-cheek jab at people who use language models to appear smarter. Another pointed out that people have always wanted to know things without learning them, and one commenter added, there is no Royal Road to Geometry. Though somebody did push back, arguing the motivation is straw-manned, since they use AI to produce things they actually know how to write but want to do more carefully or efficiently.

Mia: Then there's Andrej Karpathy's post from early August, where he argues LLM testing is moving well past prompts like create an SVG of a pelican on a bicycle. His demonstration: he gave Opus 5 the first paragraph of The Lord of the Rings, a one-million-token budget that cost about ten dollars, and asked for a three.js render. Opus went off for roughly two hours and wrote fifty-five hundred lines of code that procedurally rendered the story. Karpathy called it kind of janky but fun.

Milo: And his framing is that the economics have completely shifted, right?

Mia: Exactly. He put it as moving from no one would ever do this to sure, why not, it's basically free, noting these models have all the stamina and patience in the world. He's excited about hyper-custom worlds — an ephemeral GTA of X on demand, where players could drop into the Lord of the Rings story as a spectator NPC or one of the characters.

Milo: But he also flagged a genuine weakness in all of this.

Mia: Yeah — LLMs can't easily audit their own work because they can't efficiently and natively perceive videos or play games. Opus 5 had to slowly take screenshots at different points, messed up a few times, and that's where the jank came from. Karpathy linked Simon Willison's post on the pelican test, uploaded the source so it's playable and forkable in the browser, and joked to look out for GTA Hobbiton dropping before GTA VI. Elon Musk replied, yah. Comments split between the benchmark idea and a tangent about Google Wave — one person called it a perfect benchmark to burn more tokens, conveniently, and another wished for a timeline where they never had to see the pelican SVG test again.

Mia: A Hacker News user known as thebigship ran their own personal AI benchmark built around a single prompt: generate an SVG of a frog with a Habsburg jaw. Every model got three tries a month, and in the August edition fourteen models ran forty-two tries — and all forty-two produced an SVG. Their argument was that this works better than the more familiar pelican riding a bicycle benchmark, because it hinges on an anatomical feature that a lot of models associate with royalty. Habsburg is a royal lineage, after all, and the Habsburg jaw is an anatomical trait.

Milo: And their headline find was a striking one. Seven of the fourteen models silently brought royalty into a prompt that only named an anatomical feature. Worse, two of them actually knew they were extrapolating — they said, essentially, because Habsburg — and went ahead and did it anyway.

Mia: They also flagged determinism as a practical concern for anyone choosing a model. Mistral returned byte-identical output across separate calls, which one commenter said took them off guard, because many other models varied between runs — which is more what they expected. Another commenter wondered whether the setup accidentally hit a cache somewhere along the line, and someone else teased thebigship with a playful question: bite-identical? — which earned the admission that they'd missed the chance to say it themselves.

Milo: And the behavioral differences run deep. One model narrates its work across some sixty-five comments while another stays silent. So thebigship's core point was that how much a model embellishes beyond what you asked, and whether it behaves deterministically, are directly practical questions when you're picking which model to use.

Mia: From a frog benchmark to formal methods. F-asterisk is a general-purpose, proof-oriented programming language. The headline idea is that it marries dependent types with proof automation — automation based on SMT solving plus interactive theorem proving through tactics. So the code and its proofs live together.

Milo: And it's not some academic toy. Programs compile by default to OCaml, and fragments can be extracted to other targets through toolchains. Its production uses include the cryptographic libraries behind Firefox, the Linux kernel, Python's TLS stack, the Tezos blockchain, the ElectionGuard SDK, and the WireGuard VPN — all extracted from F-asterisk to C and assembly. There's even a parser generator whose proven code is used in Windows Hyper-V, where every network packet moving through the Azure cloud is parsed and validated by the generated code first.

Mia: In the discussion, one commenter particularly liked being able to express calling external libraries while incrementally migrating an existing C codebase into F-asterisk, calling it a very solid language. Another commenter asked what that phrase meant — whether it was about calling not-yet-ported C functions while asserting their behavior — and someone else suggested it reads as expressing, you know, the act of calling external libraries.

Milo: But not everyone was sold. Several commenters took aim at the project's website. One said they clicked through about five pages and never found a single code example — they wanted syntax shown front and center. So even a language with serious production credentials can trip over its own front door.

Mia: That's a wrap for this week's Signpost. From the Wikimedia Foundation's standoff over union recognition to that historic Treasury intervention, and from German carmakers losing ground on EVs to a frog-SVG benchmark — there was plenty to dig into.

Milo: And we loved the clever side of Hacker News this week. Syncular, Meshdiff, the experimental Kakehashi project, plus a great essay on the medieval promise of instant knowledge. Smart, surprising, and worth a follow-up read.

Mia: Thanks for listening. We'll be back with more from the world of tech, open knowledge, and the communities shaping them. Until then — stay curious.