0826 | OpenAI Jalapeño beats Nvidia Blackwell; New Mac mini M6, Nitter C&D, Dolly Parton Dies

Mia: Welcome to HackerNews Daily on Bri Radio. I'm Mia.

Milo: And I'm Milo. We've got a packed lineup today, starting with news that OpenAI's new chip, codenamed Jalapeño, is being described as better than Nvidia's Blackwell.

Mia: Also on the way, a new Mac mini featuring the M6 and M5 Pro chips, plus reports that Dolly Parton has passed away.

Milo: And we'll dig into just how much of Hacker News is actually AI-generated. Let's get started.

Mia: OpenAI's next accelerator, codenamed Jalapeño, is being positioned as better than Nvidia's Blackwell, in reporting from Semi Analysis picked up on Hacker News. The conversation immediately pivoted to a bigger question, though: one commenter put it this way — the impressive part is what happens outside the LLM accelerator space, when LLM techniques get used to design the next generation of chips. So the Blackwell comparison becomes almost the setup, and the open question is whether that capability spills into general silicon design, giving general-purpose chips a leap forward rather than just the dedicated LLM accelerators.

Mia: Qwen 3.8-Flash-Next is releasing tomorrow, described in the Hacker News post title as a redesigned multimodal mixture-of-experts model: 125 billion main parameters, another 51 billion in n-gram embeddings, and 6 billion activated per token. That count comes from an early English version of the ModelScope page a commenter copied, which also claimed that at about one-ninth the training cost it reaches capability comparable to Qwen 3.7 Plus while being stronger in coding and cowork — though the page had another paragraph about new attention that wasn't captured.

Milo: The parameter count itself proved hard to pin down — one commenter couldn't find it on the page and asked for a source, while another said it had briefly appeared in the description under the countdown before being removed. And expectations had actually run lower: one commenter hoped for a 35 billion parameter model, and a 5090 owner wanted a 35B-A3B variant to run locally. Instead, the 27 billion model from last week apparently runs great on a 5090, and the new one is basically aimed at Macs, Strix Halo, and DGX Spark — another commenter admitted they're nowhere close to running something this size locally, and hoped for a free endpoint.

Mia: One developer got a 35 billion parameter Qwen model running on a Raspberry Pi and reports impressive intelligence and stability. They built connectors for the car's OBD port to read the car's internals, plus connected the manufacturer's cloud service so it can handle things like changing the air conditioning or locking and unlocking doors.

Milo: It gets fed the full car manual and is hooked up with their other agents in a shared discussion room, so they can ask the car questions directly — when to add oil and what kind. Effectively, it's a local AI assistant living inside the car, running on a Pi.

Mia: Apple's new Mac mini lineup now reaches the M6 chip alongside an M5 Pro, and some longtime mini owners aren't entirely sentimental about the change. One commenter called it a bit of an emotional goodbye to the super cheap Mac minis they'd been able to pick up for family members and even use as headless iOS and macOS build servers sitting in a closet at a previous job.

Milo: On the pricier side, the Mac Studio gets an M5 Max and an M5 Ultra. The Studio with M5 Max starts at 2,499 dollars in the US, or 2,299 with the education price, while the M5 Ultra model starts at 5,499 — and a European commenter was already working through what an M5 Ultra with heavy GPU memory would cost them once configured. So the emotional part is about the budget tier changing, but the high end clearly keeps climbing.

Mia: Python's six pre-declared constants are, in the words of the original article, a bit weird. True, False, and None are keywords resolved as lexical tokens during parsing, which is why something like x.True is a SyntaxError — yet they also exist as normal builtins and can be overwritten through setattr on the builtins module without changing what the lexical token returns. The constant debug is normally True but flips to False under the -O option, and it's the only identifier in the language that can't be assigned to, with its own dedicated SyntaxError message.

Milo: Ellipsis and NotImplemented are documented as constants but are really ordinary builtins, so globals can shadow them — while the three-dot token itself keeps its value even after Ellipsis is reassigned. One commenter asked whether that three-dot token behaves like True, False, and None as a lexical token that resolves to a hardwired value, and the answer from another commenter was that the token does, even though Ellipsis is just an ordinary pre-defined constant with the same value.

Mia: The thread then widened into a debate about Python's design direction. One commenter said Python felt like a breath of fresh air in the 2010s compared to PHP and Perl, but now feels like a weird PHP itself, pointing at loose typing, a potluck standard library, and a horrible package manager — and another commenter pushed back from there.

Mia: A Show HN post called LatticeDB pitches it as like SQLite but for graph databases. The author says graph databases keep coming up at work and they found them painful to work with locally, so they built this sqlite-esque local file approach instead — and one commenter thinks that design makes sense for a lot of use cases.

Milo: That local-file angle is genuinely why SQLite took off, since having a plain file on disk means no server process to run. If LatticeDB delivers that same feel for graph data, it could lower the barrier for people who want graph-style queries without standing up something heavier — though the commenter did ask what scales of data the author has actually tested the design on, and that question is hanging unanswered.

Mia: Firefox 157 will include JPEG XL by default on all platforms. On Hacker News, the author of that announcement asked how many people in 2026 still haven't heard of JPEG XL, and that sparked a more practical thread where someone is curious why a PDF containing a jxl image takes ages to load.

Milo: That PDF lag traces to the file being embedded rather than the format being slow — the browser has to decode the jxl before it can render the page, and that decode path is where the loading time shows up. What's unresolved is whether the bottleneck sits in the decoder, the container integration in Firefox, or something about how the particular PDF was produced, since the comments never settled it.

Mia: A blog post argues tooltips need a delay, and then they need to skip it — meaning they should pause before appearing, but once one appears, that delay should be skipped for the subsequent ones. A commenter about twenty years on recalls hitting the same problem implementing hover-based dropdown menus: roughly ten lines of JavaScript and a couple of global variables, yet they were the only person who understood it, and today they'd have added more comments.

Milo: That only-person-who-understood-it detail is the real lesson, because it shows the cost is maintainability rather than writing the code. A delay mechanism is trivial to hack in, but keeping the timing logic predictable for whoever inherits it is what makes or breaks it — and the commenter basically admits they didn't comment the code well enough back then.

Mia: A Hacker News report titled Nitter and XCancel receive cease and desist notices centers on the open source Twitter front-end Nitter going dark. A GitHub issue in the zedeus/nitter repository reports that all public Nitter instances return the same Instance has been rate limited error, and a commenter confirmed many instances were showing rate limiting too.

Milo: A commenter named raphman then said the project had received cease and desist letters, was awaiting legal advice, and that all Nitter instances should stay down for the foreseeable future. Another quoted the Nitter Matrix group's account of the letter, which is sweeping: it demands permanently taking down nitter.net and the GitHub repository, deleting all so-called X Data in both, stopping all use of the Twitter and X marks, ceasing all access to X data including copies, deleting all X account credentials and session tokens, and confirming compliance in writing within three business days.

Mia: The letter asserts breach of X's terms of service, developer agreement and developer policy, plus the Texas Harmful Access by Computer Act and the Lanham Act, and threatens litigation against the recipient personally. It also says X has sent notice to GitHub and to the recipient's domain registrar. Commenters split on the response — one said anyone can send such a request, so checking with a lawyer before complying is reasonable, while another noted legal advice typically runs twenty thousand to one hundred thousand dollars and suggested Nitter seek donations to cover legal fees.

Mia: A new paper titled Black hole singularity is a surface not a point, by Andrew J. S. Hamilton and Tyler McMaken, takes aim at the widely repeated claim that a black hole's central singularity is a point. In general relativity, two observers who free-fall into a spherical black hole along different angular trajectories at the same time never actually meet at the central singularity — they lose causal contact well before reaching it. Counterintuitively, two points can be spatially close yet causally distant, and the singularity is a surface.

Milo: For rotating black holes, the authors argue the singular surface almost certainly sits at the inner horizon, where even the tiniest perturbations ignite an exponential mass inflation instability that collapses into a spacelike singular surface. They also propose black hole quantum states probably live on that effectively two-dimensional surface, coevolving unitarily with, and in thermodynamic equilibrium with, the hot atmosphere of trapped Hawking radiation inside the event horizon. Changing course, on Hacker News one commenter framed it as a critique of common popular-science journalism tropes rather than a new result, since a graduate-level general relativity class would teach the same — and another recommended Susskind's The Theoretical Minimum: General Relativity, noting that for a non-spinning black hole the singularity is a surface in time, not in space, because space and time exchange roles at the horizon.

Mia: SpaceX has been quietly building out a new site called Starbase, and a Hacker News discussion pointed out that southern Louisiana is where that buildout would land — and what that could mean for the region's workforce. One commenter there noted that coastal Louisiana is among the poorest areas of the US, poorer even than the Rio Grande valley in Texas where SpaceX's existing Starbase sits, and they suggested that a site buildout could mean a decade or two of steady work for welders, concrete workers, and tradespeople in the area.

Mia: The FDA has authorized the first wearable device that continuously monitors both ketone levels and blood sugar. The device is meant for people managing their metabolism, and the Hacker News discussion around the announcement flagged an immediate crowd it would likely appeal to — marathoners and ultra runners looking to tune their performance. One commenter was quick to note, though, that it is not actually watch-sized, and wondered whether a wrist form factor might be on the way later, pointing at Garmin as one brand that could move in that direction.

Mia: A BBC report describes a small number of US restaurants that have banned tipping outright and raised menu prices instead, so they can pay staff higher salaries. At La Cigale in San Francisco, wine waiter Caroline Kraetzer earns forty dollars an hour, about double what most service staff in the city make. The restaurant charges a flat one hundred forty dollars per person and tells guests that no tips are accepted and their return visits are enough thanks in themselves.

Milo: Other places made the switch in different circumstances and some of them reversed course. Rachel Miller, who owns Nightshade Noodle Bar in Lynn, Massachusetts, moved to a tip-free model five years ago after reopening following the pandemic, arguing that tipping let guests compensate people differently along lines of gender, race, or sexuality, and that kitchen staff were taking home only a fraction of what front-of-house staff got in tips. Her tasting menus start at one hundred two dollars for seven courses before 6 pm, and one hundred twenty-six dollars for nine.

Mia: But Talulla in Cambridge, Massachusetts dropped tips in 2020 and switched back last September. Co-owner Danielle Ayer says the 23 percent menu-price increase they leaned on was only sustainable through the winter, and that running without tips actually costs more overall because tips do not count as revenue while higher menu prices do, which raises both revenue and sales tax. Cornell professor William Michael Lynn adds that higher menu prices make a meal look pricier and tend to pull demand down. Amanda Cohen's Dirt Candy in New York dropped tipping back in 2015 and now pays staff around thirty dollars an hour.

Milo: Bomb fishing is wrecking Indonesia's coral reefs, according to a Yale Environment 360 report, and the Hacker News comment thread zeroed in on the roots of the practice. A couple of commenters pushed back on blaming the fishers directly, arguing instead that poverty and subsistence living are the forces driving adoption of the technique, and one of them suggested that subsidies from large economies could remove the need for it entirely.

Mia: A Hacker News post called "My Friend Aaron," dated late August 2026 and written by Rory McMeekin, is a short story about a narrator who meets Aaron at school at age eleven. The narrator asks to be his best friend, but Aaron already has one. By seventeen Aaron is ranked in the top six worldwide at World of Warcraft.

Milo: Later Aaron loses a few thousand dollars — basically his whole inheritance from his dad — on unregulated prediction markets, calls the system rigged, and proposes building his own prediction-market exchange, which the narrator agrees to help build. In the story those prediction markets are unregulated sites where anyone could bet on nearly anything, from oil prices to game release dates to a pop star's like counts.

Mia: The discussion around the story ran with the darker implications of prediction markets. One commenter predicted a prediction-market murder within five years, calling murder systematically a fairly dramatic act. Another replied that every prediction market is effectively an assassination market, because they trade influence more than information and sabotage is the simplest way to steer an event. Someone else called that an old story, pointing to fixing a fight or a game.

Milo: One commenter recalled seeing the notion of assassination markets back in the early 2010s as bounties rather than markets on when someone would die, and argued that a large enough prize pot would eventually make it worthwhile for someone — with incentives beyond money, including adoration, protection, and power — and that they were surprised it had not been seriously built with crypto. Another commenter countered that powerful targets have legal or government muscle on their side, and that millions of people usually do not agree on someone dying unless that person is genuinely powerful.

Milo: Dolly Parton has died, and the news is drawing a lot of reaction on Hacker News. One commenter who said they were never much of a country music fan still always considered themselves a fan of her.

Mia: That speaks to how broad her reach was. The tributes show up alongside the announcement itself across several major outlets, so this is being treated as a genuine loss far beyond the country and western audience she came from.

Milo: Beyond the Dolly Parton news, there's a post on the coredump.cx blog asking a pointed question: how much of Hacker News is AI-generated? The piece runs under the assumption that Pangram's AI detection is highly effective.

Mia: And that assumption is exactly what's being challenged in the comments. One reader is asking whether Pangram's detection capability is actually considered strong, or even on the higher side, among English-speaking circles. In other words, if the detection isn't as reliable as the post assumes, the whole estimate of AI share starts to look shaky.

Mia: From OpenAI's Jalapeño going up against Nvidia's Blackwell, to Apple's M6 and M5 Ultra day with a new Mac mini and Mac Studio — that's what we covered today.

Milo: Thanks for listening along with us. We'll catch you on the next one.