0817 | Stripe Finalizes OpenRouter Acquisition, Nvidia Reduces OpenAI Funding, and More

Mia: You're listening to HackerNews Daily on Bri Radio. I'm Mia.

Milo: And I'm Milo. Today we've got a packed mix — Stripe is reportedly closing a deal to buy OpenRo, and Nvidia is apparently pulling back on how much OpenAI infrastructure financing it's willing to guarantee.

Mia: We'll also dig into who the token brokers are, why some people think models are getting dumber on purpose, and Anthropic's published system prompts that have the community talking.

Milo: Plus a RISC-V debate from a third-world embedded engineer, an ad blocker landing in Firefox for iOS, and a nuclear plant in Florida that was shut down after control rods dropped into the core. Let's get into it.

Mia: Bloomberg reports, citing people familiar with the matter, that Stripe Inc. has finalized an agreement to acquire OpenRouter Inc., the startup that helps companies switch between artificial intelligence models, for more than seven billion dollars. The deal comes just months after OpenRouter raised money at a reported one point three billion dollar valuation, and it could give Stripe, a payments processor, a stronger footing in the AI sector. Bloomberg says the acquisition underscores business demand for cost-friendly AI solutions, while OpenRouter itself provides access to hundreds of AI models, aiming to match developers with the most efficient and affordable options. In the Hacker News discussion, commenters dug into the strategic fit: one pointed out that OpenRouter already uses Stripe for payments, so the acquisition could cut OpenRouter's costs and increase Stripe's revenue, while another argued that both companies put a single API key in front of a fragmented ecosystem and charge a convenience fee, so Stripe would rather own the largest middleman than let a rival running routing, limits, and policies operate outside its rails. One commenter went further, comparing Stripe's ambitions to competing with the US Federal Reserve as the core dollar payment processor while also handling tokens, adding that the transaction volume must flow.

Milo: So that is a big step up in value in a very short time, and the reporting frames it as business demand for cost-friendly AI. For Stripe, it's a payments processor buying its way into the AI sector. And there's a quote that really captures the logic here — OpenRouter's own CEO once described the company as the AI equivalent of Stripe.

Mia: A Hacker News post picks up Reuters reporting based on the Wall Street Journal: Nvidia has dramatically reduced the amount of OpenAI infrastructure financing it may guarantee for the proposed Ohio data center project. Nvidia is now expected to initially guarantee less than one hundred twenty billion dollars, down from the two hundred fifty billion previously discussed, and it would provide a financial backstop only for the first phase of the project, with a deal that could be signed as early as this weekend. The revision came after investors raised concerns about Nvidia's risk exposure tied to large financing commitments, and it follows Nvidia's Monday announcement that it partnered with six major financial institutions to launch compute financing platforms aimed at raising over five hundred billion dollars in third-party capital for AI infrastructure. OpenAI is still discussing a binding lease for the full ten-gigawatt Ohio project, developed by SoftBank subsidiary SB Energy, which would be the largest data center project announced to date if completed. Nvidia did not immediately respond to a request for comment, while OpenAI declined to comment.

Milo: And there's a sharper argument from a third commenter. Both companies basically put a single API key in front of a fragmented ecosystem and charge a convenience fee, so rather than let a rival that handles routing, limits, and policies stay outside its rails, Stripe would rather own the largest middleman in that space.

Mia: One commenter even compared Stripe's ambitions to competing with the Federal Reserve as the core dollar payment processor while also handling tokens, adding that the transaction volume must flow. We should flag that the Hacker News conversation also quoted a claim that every company on the Forbes AI 50 that monetizes does so on Stripe — though that's presented as a claim in the thread, not something confirmed by our reporting.

Milo: Hacker News readers are discussing Who Are the Token Brokers, a Vectoral article by Matt Lenhard that follows up on his earlier reporting about the token relay market. Lenhard reports that token brokers buy unused AI credits from startups and resell them, and that founders told him they receive inbound email pitches for off-market inference, including one offer of direct relays to OpenAI and Claude at forty to fifty percent below list price with only an API key switch. He catalogs marketplaces like AI Credits and AICreditMart, which list credits from providers including MiniMax, ElevenLabs, Google Gemini, OpenAI, Microsoft Azure, and Anthropic at thirty to eighty percent discounts, plus services like CheapCredits, Tokvana, and Neokens claiming bulk-pricing discounts. CheapCredits advertises a flat forty percent off list on the GPT-5 series and names OpenAI and Anthropic as sub-processors in its GDPR data processing agreement. Lenhard says a forty percent discount would be very unlikely for anyone but a provider's top customer and suspects CheapCredits is acquiring supply another way, while Telegram channels and Reddit posts also trade credits, including an offer of roughly twenty-five hundred dollars in OpenAI credits earned through YC Startup School.

Mia: And it wouldn't cover the whole project either. Nvidia and OpenAI are nearing an agreement where the chipmaker provides a financial backstop only for the first phase, and a deal could be signed as early as this weekend. The reduction came after investors raised concerns about Nvidia's risk exposure from large financing commitments, and it follows Nvidia's Monday announcement that it partnered with six major financial institutions to launch compute financing platforms aimed at raising over five hundred billion in third-party capital for AI infrastructure.

Milo: OpenAI is still discussing a binding lease for the full ten-gigawatt Ohio project, which is being developed by SoftBank subsidiary SB Energy. If completed, it would be the largest data center project announced to date. On Hacker News, commenters dug into the financing structure — one ran the numbers and looked at what happens if Nvidia sells a hundred billion of hardware with a specific margin and provides a backstop on that same hardware. That is a case where the company is unprofitable despite an eight hundred fifty two billion valuation, so its ability to fund large-scale commitments stays under scrutiny.

Mia: A Hacker News post links to a blog argument that AI models are getting dumber on purpose, and in the discussion a commenter makes a sharp technical point: the benchmark the argument leans on, SimpleQA, stopped measuring back in September 2025, so the claim is partly built on older data. That same commenter notes it seems like an AI-generated argument that uses old facts, something that happens fairly often. This undercuts the central evidence behind the post's premise, since it leaves the more recent performance question unanswered. The discussion thus turns on how much weight the argument can carry when its core benchmark is stale, and on whether the reasoning still holds up once that data gap is taken into account.

Milo: The gist: token brokers buy unused AI credits from startups and resell them. Founders told Lenhard they get inbound email pitches for off-market inference — one pitch offered direct relays to OpenAI and Claude at forty to fifty percent below list price with just an API key switch. One broker claimed their account could spend a hundred thousand dollars a day, sent an API endpoint instead of keys, and billed after a usage milestone.

Mia: He catalogs the marketplaces — one bills itself as a credit marketplace, and another lists credits from providers like MiniMax, ElevenLabs, Google Gemini, OpenAI, Microsoft Azure, and Anthropic, at discounts of thirty to eighty percent. Lenhard's own test listings were still pending approval. Other services advertise bulk-pricing discounts, and one advertises a flat forty percent off on the GPT-5 series with a data processing agreement naming OpenAI and Anthropic as sub-processors.

Milo: Lenhard says a forty percent discount would be very unlikely for anyone but a provider's top customer, so he suspects that particular broker is acquiring supply some other way. Telegram channels and Reddit are in the mix too, with offers running from a couple thousand dollars in OpenAI credits earned through YC Startup School up to ten thousand dollars in OpenAI API credits expiring later this year.

Milo: A Hacker News post picks up the news that Anthropic has published the system prompts for claude.ai and its mobile apps, and one commenter noticed the earliest prompts were slightly over three hundred words while the latest run three thousand-plus. Simon Willison keeps these prompts as a git commit history so changes are easy to diff, and he highlighted one notable addition in the Opus 5 prompt: it explains that Anthropic suspended Claude Fable 5 and Claude Mythos 5 on June 12, 2026, to comply with U.S. Department of Commerce export controls, lifted on June 30, with access restored July 1, and instructs Claude to treat those events as outside its training-data cutoff. Several commenters argued the ballooning system prompt hurts performance, with one saying the token budget is being spent on useless nonsense at the most important part of the context window and a third pushing back against prompt-engineering tricks in general. Willison also flagged that Anthropic publishes the prompts but not the tool definitions, which he called more interesting for understanding what Claude can actually do.

Mia: That resonates, because the same pattern that post describes is the thing that just got people looking at Anthropic's published system prompts for Claude and its mobile apps. One commenter noticed the earliest prompts were a little over three hundred words, and now they run past three thousand.

Milo: Right — Simon Willison keeps those prompts as a git commit history so changes are easy to diff, and he flagged a notable addition in the Opus 5 prompt: Anthropic suspended the Claude Fable 5 and Claude Mythos 5 models on June twelfth, twenty twenty-six, to comply with U.S. Commerce Department export controls, which were lifted June thirtieth, with access restored July first. The prompt tells Claude to treat those events as outside its training-data cutoff, confirm them matter-of-factly, and otherwise treat export controls as a normal current political topic.

Mia: Several commenters argued the ballooning prompt hurts performance — one said the token budget is being spent on useless stuff at the most important part of the context window — though a quieter counterpoint is that prompts accrete over time in response to mishaps and loophole exploitation, like building codes and boilerplate contracts do. Willison also flagged what's missing: Anthropic publishes the prompts but not the tool definitions, and they don't publish Claude Code's system prompts at all, which is the content he says is actually the most interesting for understanding what Claude can do.

Mia: A new paper from the Max Planck Institute for Intelligent Systems, the ELLIS Institute Tübingen, and ETH Zürich reports what happens when you train a language model so it only knows what a fifth grader knows. The work, called LittleLearner, built an 88-billion-token training corpus distilled from FineWeb-Edu through a five-stage filtering pipeline aligned with U.S. Common Core standards from kindergarten through fifth grade, deliberately excluding concepts, facts, and vocabulary taught above that level. Researchers trained models from scratch at three sizes — 0.6 billion, 1.3 billion, and 5 billion parameters — each paired with a matched unfiltered control sharing the same architecture, tokens, and recipe. The central finding: scaling, post-training, and in-context learning amplified what the curriculum had taught, but none of it meaningfully improved out-of-scope performance, indicating the pretraining filter set the effective capability ceiling. Even GRPO post-training with out-of-scope data boosted in-scope capabilities without recovering beyond-fifth-grade ones, and in-context learning with the tested prompts did not unlock new reasoning in beyond-grade-five material for the 5-billion-parameter model. One person in the Hacker News discussion claimed the model knows about quantum entanglement, clearly beyond fifth grade, but another commenter pointed out that claimer had not actually read the example being referenced.

Milo: MathCode is a new terminal AI coding assistant on Hacker News that bills itself as a frontier mathematical coding agent, with a built-in math formalization engine: you describe a problem in plain language and it converts the statement into a Lean 4 theorem, then attempts a formal proof. The quick start shows a command to prove that the square of an even number is even, with outputs written to a LeanFormalizations folder, and it requires macOS on Apple silicon or Linux on x86-64, plus the codex CLI for the default backend. Its selling points include a persistent Lean REPL that brings compile checks down to roughly 0.4 seconds after a one-time warmup instead of about thirty, an auto-named reusable theorem library, a persistent axiom library for conversational assumptions, searches of leansearch.net and Loogle for verified Mathlib lemmas, and an Obsidian vault visualizing theorem-to-lemma dependencies. The project says its pipeline is based on the older AUTOLEAN project, which prompted one commenter to ask whether this is really just a wrapper around it. Another thread turned to the business side: with no licensing terms attached, one commenter argued commercial use is blocked, while another countered that AI-generated code isn't copyrightable anyway, so once it's released people can use it as they please. Asked what commercial setting would actually want a Lean theorem-proving agent, one commenter jokingly suggested, quote, Mathematics, Incorporated.

Mia: Buf has announced what it calls the first fully-featured, production-grade language server for Protobuf, meaning the kind of editor integration other major languages have long enjoyed: go to definition, code completion, finding references, and semantics-aware syntax highlighting. The server ships with the Buf command-line tool, so VSCode users install the Buf extension and Neovim users run the command with proto filetypes. Buf says Protobuf previously lacked that LSP support. Under the hood it's built on Buf's own Protobuf compiler frontend, which the company says is fully compliant with the spec, much faster and more flexible than Google's protoc, and which Google itself uses alongside protoc in parts of its codebase. A new query-driven frontend enables incremental compilation and smarter diagnostics, like correctly flagging duplicate repeated modifiers, and a new abstract syntax tree and intermediate representation replace the older format. On the roadmap are automatic fixes for imports, tighter Buf configuration integration, completion and reference lookup for custom options, suggested field and enum numbers, and support for Protovalidate rules with CEL snippet highlighting. The Hacker News thread, though, asked a bigger question: do Protobuf's advantages still hold in the LLM era, with debate turning on how the format's strengths stack up against what modern AI tooling now handles.

Milo: A Cloudflare user is raising privacy concerns about a silent change to their website. A Hacker News poster reported that after switching nameservers to Cloudflare to serve files from their R2 storage through their own subdomain, the company quietly injected a JavaScript analytics snippet into what was supposed to be a plain HTML-only, JavaScript-free site, and they had to go into the Analytics dashboard, add the site, and only then disable the snippet, calling the whole approach entirely invasive and arguing it should be opt-in rather than opt-out. Others chimed in with the same story. One commenter said they'd explicitly turned off all analytics and still found the script on every site using Cloudflare for caching, calling it gross and unacceptable and noting that caching does not mean modifying their site. Another suggested reporting it as a bug, since it seems more like a mistake than deliberate malice. And when the question came up of whether this is well known and how Cloudflare gets traffic and DDoS telemetry otherwise, one commenter pushed back that Cloudflare serves the HTML anyway, and therefore has every ability to track individuals regardless.

Mia: Firefox for iOS has a native ad blocker now, but the Hacker News conversation zeroed in on what it won't touch. A commenter quoted Mozilla's own support page: the blocker doesn't block ads on search engine results pages, whether from Google, Bing, DuckDuckGo, or other search providers, and it doesn't block sponsored content on Firefox's Home or New Tab page. One commenter guessed those limits must be an Apple requirement, since they don't exist on Windows or Android. But another pushed back, noting the ad-blocking list called EasyList doesn't block those things on any platform, and they know of no blocker that kills the ads on the built-in New Tab page, which can be disabled in settings. On desktop, one commenter said, you can set the Home page to blank and use an extension, while another noted sponsored content can be managed in Homepage settings. The revenue angle surfaced too, with one person joking that blocking ads on Mozilla's own homepage raises the question of keeping the lights on. Another said Mozilla knows not to bite the hand that feeds it, since Google feeds Mozilla hundreds of millions of dollars, while a reply countered that it's actually Mozilla funneling billions of dollars worth of traffic to Google. And one separate thread debated avoiding ads by not visiting pages that show them — though, as another commenter pointed out, it's tricky to tell which pages those are before you visit them.

Mia: NIH is winding down a key grant program for early-career clinical researchers, and a Hacker News thread went straight to why. Commenters split between bureaucratic chaos and deliberate destruction: one said NIH has been chaotic and mismanaged for two years, with labs defunded for no good reason and research directions put on ice, "likely forever," and called computational biologist Sean Eddy a clear case of malice. Eddy, who at 60 planned to work another decade with his team, said, "For someone of my career stage, this is probably not recoverable." Another commenter argued there is no reason to give the administration the benefit of the doubt, citing OMB head Russ Vought on putting government workers in trauma. One framed it as a knee-jerk reaction to China's achievements that can't mature in four years without long-term process changes. On why Eddy doesn't move to big pharma, commenters said pharma researchers must do work bosses consider productive, not curiosity-driven public research, and that Eddy's lab builds tooling the whole field depends on with no commercial funding. Others called the policy goal weakening US science or blamed voters who elected leaders on that promise.

Milo: Honestly, they split hard. Some commenters read it as bureaucratic chaos — NIH mismanaged for two years, labs defunded for no good reason, research directions put on ice likely forever. But others framed it as deliberate destruction, and a bunch pointed to Sean Eddy as the clearest case.

Mia: Right, Eddy — the computational biologist a commenter called foundational to their whole field. One person quoted him directly saying that at sixty, with plans to work another decade with his team, this is probably not recoverable at his career stage. That quote kept coming back.

Milo: And it sparked a really pointed exchange about why he wouldn't just go work for a big pharma company. One commenter explained pharma researchers only get to do work their bosses consider productive — not curiosity-driven public research. And another added that Eddy's lab builds the tooling and concepts the whole field depends on, and there's no commercial funding for that work at all.

Mia: So the thread circled back to intent. Some commenters argued there's no reason to give the administration the benefit of the doubt, citing OMB director Russ Vought on putting government workers in trauma. Others floated a kinder reading — that the goal is revamping institutions to accelerate timelines and cut waste. But one commenter pushed back hard on that, saying that can't happen in four years without long-term process changes, and framed it as a knee-jerk reaction to China's gains.

Milo: And then you had the harshest take of all. One commenter said flatly that the policy goal is to weaken US scientific research, and another blamed voters who elected leaders on that promise. The counterargument was pretty thin — mostly that it was worth waiting to see how it plays out. Either way, the thread landed on one grim shared conclusion: a whole generation of clinical researchers is losing its on-ramp, and there's real doubt about whether the pipeline comes back.

Mia: A Hacker News post flagged academic papers using the phrase "kidney disappointment" in place of "kidney failure," linking to a Google Scholar search for the term. Commenters started questioning whether it was a translation issue or AI-written text, and one shared sentences like "gauges an individual's gamble of kidney disappointment." Some suggested AI might swap words for watermarking; another argued it was more likely disguising plagiarism, then updated that it was probably just poor machine translation. The name "tortured phrases" came up for such synonym substitutions, with a Scholar search turning up more examples, while others offered their own names from "tax avoision" to "expressions subjected to enhanced interrogation." One commenter cited a 2021 arXiv paper on the phenomenon and noted SCIgen had been generating nonsense papers that got published way back in 2005. The definition landed on bizarre, nonsensical word combinations like "counterfeit consciousness" for artificial intelligence, and one user suggested text spinners could be an untapped gold mine for creatively naming things. The linked article itself was unreachable, returning instead a Google block page saying the request could not be processed.

Milo: They came at it from every angle. One commenter initially wondered if it was a translation issue or possibly AI-written text, then flagged sentences like "gauges an individual's gamble of kidney disappointment." Someone suggested AI might be swapping words for watermarking purposes, but another argued it was more likely disguise for plagiarism — before later updating that it was probably just poor machine translation.

Mia: And the community had a name for these. One commenter said these synonym substitutions are called "tortured phrases" and shared a search turning up more examples. Someone else called them "distressed phrases," and others piled on with their own names — "tax avoision," "moderately inconvenienced phrases," "phrases experiencing discomfort," and "expressions subjected to enhanced interrogation." The naming contest was clearly the real show.

Milo: Right, and one commenter cited an older arXiv paper and noted that SCIgen had been generating nonsense papers back in 2005 that actually got published. Someone quoted the definition of tortured phrases as bizarre, nonsensical word combinations — like "counterfeit consciousness" standing in for "artificial intelligence" — and suggested text spinners could be an untapped gold mine for creatively naming things. The whole thread basically turned a cautionary tale into a comedy bit.

Mia: Unit 1 at the St. Lucie Nuclear Power Plant was manually shut down after three control rods dropped into the reactor core. Per an NRC filing, the manual reactor trip happened at 9:47 a.m. EDT on Aug. 13, 2026, while the unit sat in Mode 1 at 100% power; the NRC classified it as non-emergency. The licensee reported the trip as uncomplicated, with all systems responding normally, and operations stabilized the plant in Mode 3, or Hot Standby, removing decay heat by discharging steam to the main condenser. Unit 2 was unaffected, and the NRC website shows Unit 1 still shut down. In the discussion, one commenter couldn't find the NRC filing and noted the reactor had its operating license extended earlier in the year, while another pointed to the official event listing. The leading speculation was an interruption of power to the field windings of the affected control rods—maybe a fried control cabinet or an electrical gremlin—with the three inserted rods disrupting the neutron flux profile, making a manual shutdown the right call until the cause is found, repaired, tested, and the reactor restarted.

Milo: According to a notification filed with the Nuclear Regulatory Commission, the manual reactor trip happened mid-morning on August 13th, while the unit was running at full power. The NRC classified it as a non-emergency. The licensee reported the trip was uncomplicated, all systems responded normally afterward, and the plant stabilized in what's called Hot Standby, with decay heat being removed through the main condenser. Unit 2 wasn't affected, and the NRC's site still shows Unit 1 shut down.

Mia: And the Hacker News thread was mostly speculation about the cause. One commenter pondered an interruption of power to the control rod field windings — maybe a fried control cabinet or an electrical gremlin — and said the three inserted rods would have disrupted the neutron flux profile, making a manual shutdown the right call until the cause is found, repaired, tested, and the reactor restarted. He noted this kind of event has happened many times before, even if it's not typical. Another commenter suggested it could actually be an intentional design feature, with control rods held by an actively powered retaining mechanism so all three drop automatically if power's lost.

Mia: A response essay on rvembedded.com rebuts Dmitry Grinberg's "RISC-V: They Should Have Known Better," which had hit the Hacker News front page and started an argument on Lobsters. The author writes from Trinidad and Tobago and about a year ago moved their stack from STM32 and ARM to RISC-V. The core argument is economic: from outside the US and Europe, shipping a one-dollar chip costs US$60 to US$200, and a PCB company reportedly turned down a sponsorship based solely on shipping to the author's location. The difference between a ten-cent and a one-dollar part, the author argues, is the difference between thirty students each holding a chip and thirty students watching one demo board. The author concedes Grinberg's complaints about the compressed store offsets and Zicsr, and notes RISC-V International denied them membership, but argues Grinberg's own requirements for a cheap microcontroller core—low interrupt latency, small die area, good code density, no multiplier, no divider, no privilege separation—land exactly on RV32IC or RV32EC, and that the ten-cent CH32V003 the author shipped two products on matches that core.

Milo: So it's a direct counter to Dmitry Grinberg's essay, which had hit the front page and started an argument on Lobsters. The author writes from Trinidad and Tobago and switched their entire stack from STM32 and ARM to RISC-V about a year ago. Their core argument is economic: from outside the US and Europe, shipping a one-dollar chip costs sixty to two hundred dollars, and a PCB company reportedly turned down a sponsorship purely because of the author's location.

Mia: And they make the cost gap vivid. The difference between a ten-cent part and a one-dollar part, they argue, is the difference between a class of thirty students each having a chip and thirty students watching one demo board. So it's not about bleeding-edge performance — it's about whether the cheap end of the market can even participate.

Milo: The author actually concedes some of Grinberg's complaints — the compressed store offsets, and Zicsr being a separate thing you have to ask for — and notes RISC-V International denied them membership outright. But the key point is that Grinberg's own stated requirements for a cheap microcontroller core — low interrupt latency, small die area, good code density, no multiplier or divider, no privilege separation — land squarely on the RISC-V baseline. And Grinberg himself writes he's "100% sure" RISC-V will own the cheap-as-dirt single-use microcontroller space. The author shipped two products on a ten-cent RISC-V chip with sixteen registers, no multiplier, no divider, machine mode only, and just a couple of kilobytes of memory — including a bin monitor with a time-of-flight sensor.

Mia: A New Yorker piece in the Annals of Inquiry section by Shayla Love, dated August 6, 2026, centers on Katherine Dunn's new book "Little Blue Dot," which raises questions about the unintended consequences of knowing where we are. It opens with an unsettling story: a neuroscientist who studies spatial navigation at Ben-Gurion University, driving north with his children and following Waze, was told to turn left and drive roughly two hundred miles south—because his phone thought he was in Beirut, with Israel often jamming GPS signals. The Hacker News discussion turned to pre-GPS navigation. One commenter called orienteering a good skill, recalling a detour in a new city into what he described as the poorest, gang-controlled neighborhood. Another said he got turned around even in familiar, non-grid cities before consumer GPS, that driving with a map open on his lap at night was a nightmare, and it took years to build a passable sense of Boston's geography. Others added Los Angeles via Thomas Guide, and that European cities and Boston were especially tough, with Boston giving even GPS users "borderline paranoia."

Milo: A neuroscientist who studies spatial navigation at Ben-Gurion University was driving north with his children following Waze when an automated voice told him to turn left and drive roughly two hundred miles south. He realized his phone thought he was in Beirut. The article notes that Israel often jams GPS signals — so the systems we lean on for where we are can quietly be telling us the wrong thing.

Mia: And the Hacker News thread ran with the pre-GPS angle. One commenter called orienteering a good skill to have, noting his wife and a close friend could get lost trying to get out of a cardboard box, and recalled deviating from a planned route in a new city and crossing the wrong street into the poorest, gang-controlled neighborhood — then buying donuts through bulletproof glass.

Milo: Others weighed in on navigating without GPS. One commenter said that before consumer GPS he got turned around even in familiar cities that aren't on a grid, that driving at night with a map open on his lap was a nightmare and not especially safe, and that it took years before he had a passable sense of Boston's geography. Someone else recalled the Thomas Guide for Los Angeles, and others noted European cities and Boston were especially tough — one saying even GPS in Boston gives borderline paranoia. It all feeds back into the piece's core question: what do we give up when we always know exactly where we are?

Mia: The two-day weekend just turned a hundred years old, and a Guardian essay by Jess Cartner-Morley argues that employees might be the ones ruining it. It goes back to 1926, when Henry Ford gave his staff both Saturdays and Sundays off, becoming the first big employer to do so. Before that, a weekend on this scale didn't exist, and the piece recalls how the Soviet Union had no weekends at all from 1929 to 1940, giving citizens a randomly allocated day off in every seven so that about eighty percent of the workforce was on the job any given day and factories never had to power down. The problem was people lost the shared day that families could spend together, and one worker's letter to Pravda complained it was no holiday if you had to have it alone. The workforce became demoralised, the promised productivity spike never materialised, and the policy was modified and eventually abandoned. Portsmouth history professor Brad Beaven argues the weekend represents the triumph of collective time over market time, calling it not just rest but reclaiming autonomy from the industrial clock.

Milo: There's a great line in there from historian Brad Beaven, who calls the weekend the triumph of collective time over market time. Not just rest, he argues, but people reclaiming control of their lives from the industrial clock. And the piece brings that all forward to today, asking whether hybrid working and so-called Skiveaday Fridays have actually spoiled what the weekend was meant to be. On Hacker News, people were debating just how recent this whole idea of invented leisure really is. One commenter pointed out that a musician's grandfather was already alive in Napoleon's day. Someone else mentioned a grandmother who said music festivals hadn't even been invented yet, because without electronics there was no electronic music and no big raves. Another person recalled their mother, who lived through segregated water fountains and a time when women couldn't open bank accounts, while her grandkids have never known a world without Fortnite on iPads and watch-like gadgets straight out of Dick Tracy. The whole thread just lands on how strange and very recent our idea of free time actually is.

Mia: A blog post making the rounds asks whether its author is building the world's first true telnet bulletin board system running on a Casio calculator. The author admits that about eight weeks back they had virtually no interest in calculators at all, until a SOTA Reflector post from Mark M5TEA about a ported calculator emulator drew them in. It was BASIC and RS232 serial communication that changed their mind. Coming from limited programming knowledge, they used an AI model to vibe code an amateur-radio logging program and an unfinished CW keying program that needs an external bridge circuit to the radio. They now own two Casio pocket computers imported from Japan, this one a rarely-seen Z-1GR model that arrived dead and sold as junk, with planned revival attempts planned, plus the on-board eight kilobytes of RAM they'd host the BBS from. So far that's four programs written, one published on GitHub, with the rest promised as open source down the line.

Milo: And here's the hardware setup. The main Casio arrived boxed, near mint, with its manuals and a serial cable, plus optional memory. It comes with eight kilobytes of RAM on board, expandable to thirty-two with an add-on card, and even further if you swap out the chip. That's an incredibly tiny amount of memory to run any kind of server, which is exactly why the question is worth asking. The author also picked up a rare second unit, a Z-1GR, which arrived dead and sold as junk, with plans to try reviving it. On Hacker News, the nostalgia ran strong. One person remembered programming a Casio through secondary school, others called it a childhood fever dream, and one commenter was reminded of pretending to be Ethan Hunt with early handheld gear. A few people just admitted they'd gone out and ordered their own retro hardware after reading the post.

Mia: If you ever fiddled with interactive physics graphs in school, you might already know Paul Falstad's work. His collection of math, physics, and engineering applets is getting fresh attention on Hacker News. These are small simulations he wrote to make abstract ideas visible, spanning waves and acoustics, electrostatics, analog circuits, quantum mechanics, thermodynamics, mechanics, and more. Originally built in Java, most have now been converted to JavaScript so they run without a Java-capable browser, which is a big deal for keeping them alive since Java applets stopped working in modern browsers years ago. In the discussion, one commenter noted Falstad also made circuitjs, useful for prototyping because you can modify and probe elements in real time, unlike SPICE simulators, though they said it can get unstable in complex circuits with nonideal op amps so they check final designs in SPICE. Another recalled that Falstad wrote zsh as a sophomore in 1990, the shell that became macOS's default in 2019.

Milo: The discussion pulled in a lot of personal history. One commenter shared their own wave simulation written back in two thousand five, which they'd managed to revive using the Ruffle emulator. Others recalled using the applets back when they actually were Java applets. Someone highlighted Falstad's separate project, a circuit simulator that's handy for prototyping because you can tweak and probe components live, though real-world parts can sometimes make it unstable, so a lot of people still check their final designs against a proper simulator. One commenter also pointed out that Falstad wrote the zsh shell back in nineteen ninety, which later went on to become the default shell on modern Macs. And a researcher even cited the collection as a key step in the history of educational simulations, with a couple of favorites being the ripple tank and the quantum mechanics sims. All in all, a fairly remarkable body of work from those little browser toys.

Mia: That wraps things up for today. Thanks so much for listening and for sticking with us through every story.

Milo: If you enjoyed the episode, share it with someone you think would like it — and we'll be back next time with more.

Mia: Until then, take care of yourselves, and we'll talk again soon.