0930 | Agents, Gigawatts, and a Vibe-Coded Website

Mia: Welcome back to Hacker News... B: ...A: Mia, B: Milo. Welcome back to the show — I'm Mia, and as always I'm here with Milo. B: Hey everyone. A: So we've got a full day on Hacker News, and honestly, the thread that runs through almost everything today is artificial intelligence — not just the models themselves, but what happens when they show up in pricing, in publishing, in surveillance, in gambling, even in how countries build software.

Mia: B: Yeah, and then the flip side of that — all the human-scale stuff: power grids, drive failure reports, a solar system in your browser, and a pretty sobering stat about how little we host people at home anymore. A: Right, and that's actually where we'll end things. But let's start where the news started — with OpenAI's big day. B: Let's do it. A: Okay, so OpenAI DevDay happened, and by all accounts it was a lot. We're talking more than twenty announcements from the company in one event.

Mia: B: Twenty-plus. And the headline items, as people were describing them: there's a product called Dots — these are always-on agents powered by a model called GPT-6 Astra, and each one gets its own cloud computer to run on. They've got integrations with over four thousand apps, and they're available on the Pro and Business Premium tiers. A: And then there's the model lineup news — a new one called GPT-6.1 Sol, which is positioned as being near-Astra in intelligence but at a fifth of the price.

Mia: We're talking two dollars per million input tokens and ten dollars per million output tokens, versus whatever Astra is charging. Apparently it matches Astra on a benchmark called DeepSWE and beats Claude Opus 5.5 on something called GDP.pdf. B: Which, if those numbers hold up, is a pretty aggressive price cut for something claimed to be roughly equivalent in capability. A: And then on the consumer side, there's a restructuring of the subscription tiers.

Mia: A new top tier called Pro 500 at five hundred dollars a month that includes the Ultrafast option. And Pro 200 is coming back for new users, but with lower usage allowances than before. B: So the shape of all this, to me, reads pretty clearly: OpenAI is pushing upmarket with the expensive agent product and the five-hundred-dollar tier, while simultaneously commoditizing the base model layer by cutting Sol's price dramatically.

Mia: A: Exactly — you sell the cheap model to everyone, and you sell the expensive orchestration and speed to the people who'll pay for it. B: Now, the part that nobody could verify from the announcements themselves is whether these always-on agents actually work reliably in the real world. That's the open question hanging over the whole thing. A: Which is a perfect segue, honestly, because there's a project that's trying to answer a version of that question empirically. B: Oh, livenerf.

Mia: A: Right — it's a daily benchmark that's tracking whether Claude Opus 5.5 gets worse after its launch. The idea is you establish a baseline in the first ten days after release, and then you keep running the same benchmark day after day and watch for drift. The first call on it is apparently expected around October twenty-fourth, 2026.

Mia: B: And the reason this exists at all is that "the model got worse" has become this recurring complaint that nobody can settle — because you've got no baseline, you've got prompt changes on the provider's side, you've got sampling randomness. Setting the first ten days as the reference point is at least a principled attempt to control for that. A: It's community auditing, essentially.

Mia: And it pairs nicely with another story that had people doing a very different kind of audit — the statistician Nicholas Polson. B: Right, so the claim is that he has two hundred and fifty-eight academic papers published in 2026 alone. A: Two hundred fifty-eight. In one year. B: And the suspicion on the thread — and it is just suspicion, nobody's proven anything — is that heavy AI use is behind that output volume. The detail that really took off, though, is that his name reversed spells "noslop.

Mia: " A: Which is either a remarkable coincidence or the internet reading tea leaves. I'd caution listeners that reversing a name and finding a joke in it proves nothing about the papers themselves. B: No, but it captures the mood — people are increasingly doing this kind of informal quality control on AI-assisted output, whether it's benchmarks like livenerf or eyeballing publication counts that seem physically implausible.

Mia: A: And that trust question cuts both ways, because the next story is about what happens with the data you hand over when you use these tools. B: Yeah — this one's uncomfortable. There's a study finding that AI chat providers are leaking conversation titles, actual prompts, screenshots, and unprotected permalinks to third-party trackers, and those trackers are attaching persistent user IDs to all of it. A: So not just metadata — the substance of what you typed.

Mia: Your prompts going to ad trackers with an identifier that follows you across sessions. B: And the "unprotected permalinks" part is its own problem — if a share link isn't protected, anyone who obtains it can read the whole conversation. A: And then there's the DraftKings story, which is the uglier end of the same data pipeline. They're using AI on betting records to identify gamblers who are likely to lose, and then targeting those people specifically with promotions.

Mia: B: Which is targeting the most vulnerable users with the product most likely to harm them. The EFF is calling for a ban on behavioral advertising in response. A: It's the logical endpoint of behavioral ad tech — the same machinery that decides which shoe ad to show you can decide which losing gambler to hook. B: Whether that leads to regulation is the open question, but the EFF's position is at least now on the record.

Mia: A: And speaking of technology that doesn't deliver what it promises — there were two public-sector stories that fit that pattern. B: The London one first: a six-month facial recognition trial at a station scanned over five hundred thousand faces, cost three hundred twenty thousand pounds, consumed a hundred hours of police time — and produced exactly one false positive and zero arrests. A: One false positive, zero arrests. Half a million faces.

Mia: B: And the other is the new White House site, america.gov — a twenty-five megabyte page that fails the WAVE accessibility test, while usa.gov, the old site, still exists and weighs in at eight hundred kilobytes. A: So you've got expensive surveillance tech that accomplishes nothing measurable, and a public interface that's bloated and excludes users with disabilities — while the cheaper, older version is still sitting right there.

Mia: B: It raises the question of who these projects are actually for. A: From public failures to private ones that worked suspiciously well — the vibe coding story. B: Oh, the Railcode one. The founder vibe-coded his company's site — meaning he built it largely by iterating with AI agents rather than hand-writing it — and the result was so good that people were asking which designer he'd hired. A: And the argument that took off in the discussion is that iterating with agents is design.

Mia: B: Right — the pushback to the old "vibe coding isn't real engineering" take is that the iteration loop itself, the taste and the judgment applied over many rounds, is design work. It just doesn't look like the traditional craft. A: And on the systems side of the same world, there's NSL — which spins up WSL-style Linux machines using systemd-nspawn containers inside a single VM, via systemd-vmspawn.

Mia: B: And the key contrast people drew is with distrobox — because distrobox mounts your home directory into the container, the environment isn't really isolated from your actual system. NSL keeps the host clean instead. A: So it's the same instinct — lightweight, disposable environments — but with real isolation. B: Which matters when you're running half-vetted AI-generated code in them, frankly. A: Ha — connecting the two stories whether we planned to or not.

Mia: B: Staying in tools: Godot's GDExtension, combined with godot-cpp 10 and Conan, now lets developers pull in any C or C++ library. The demo people pointed to simulates a hundred thousand particles using flecs. A: So the engine stops being a walled garden and becomes a hub for the whole native ecosystem. B: And then Tcl/Tk 9.1.0, released September twenty-ninth, 2026 — Tk gains screen-reader accessibility and bidirectional text, and Tcl gets unicode support, a timer, and lfilter.

Mia: A: A toolkit old enough to remember when that meant something different, still adding accessibility features in 2026. There's something quietly admirable about that. B: From old-but-alive to broken-on-purpose and broken-by-accident. A: The Relapse exploit first — it's a PS5 kernel exploit covering firmware seven point zero zero through thirteen point six zero.

Mia: Technically it uses JavaScriptCore typed array corruption to get execution, then chains a use-after-free in aio_multi_wait to get kernel read and write. B: So a full kernel-level compromise across a wide firmware range, which is genuinely rare for a current console. And then macOS 27 "Golden Gate," which was supposed to be the bug-fix release, ships with misaligned icons, menu bar crashes, broken search, and leftover Tahoe content still in there.

Mia: A: One team broke through the PS5's walls; another couldn't fix its own. And per the ChromeOS news, Google will end ChromeOS support in 2034, two years early, steering Chromebooks toward "Googlebook OS" with new paid licenses — so platform lifecycles are getting shorter from both directions. B: Which brings us neatly to a story where geopolitics forced a platform decision — the Netherlands and the ICC. A: Right.

Mia: US sanctions on the International Criminal Court created real risk for the Netherlands in depending on Microsoft, so they're building a NixOS-based ecosystem to replace it. Trials are running now, with a first release due late 2027. B: And the notable thing is the driver — not cost, not ideology in the abstract, but sanctions. That's a much harder motivation to ignore than a procurement debate. A: It's sovereign open-source as a geopolitical necessity.

Mia: And that leads naturally into infrastructure more broadly — grids and storage. B: Vermont first. Green Mountain Power runs a virtual power plant of over five thousand five hundred home batteries, paying customers fifty-five dollars a month for ten years — and it's now the state's largest power source. A: Distributed storage beating centralized generation, in a whole state.

Mia: B: And Delhi cut its electricity losses from fifty percent to five percent since 2002 — not with fancy tech, but by fixing obsolete equipment and cracking down on theft on a failing grid. A: Half the power was being lost. Basic repair took it to five percent.

Mia: B: And if listeners want the framework for thinking about this, there's David MacKay's Sustainable Energy Without the Hot Air — free online since 2008 — though critics note it has the primary-energy fallacy issue and its 2008 solar and wind prices are badly outdated. A: Still a good way to learn to do the arithmetic, as long as you update the numbers.

Mia: B: And while we're doing arithmetic: Backblaze's Q2 2026 drive report covered three hundred fifty-four thousand four hundred fifteen drives with an overall annualized failure rate of one point seven three percent. And they announced a multi-exabyte storage agreement with CoreWeave. A: The AI demand side showing up in storage economics — someone's signing multi-exabyte deals, which tells you where the data is going. B: Which connects to the browser Solar System — space.bl2.net.

Mia: Real scale, five hundred twenty-six thousand asteroids from the JPL Small Body Database, satellite positions from CelesTrak TLEs, and SGP4 propagation running in web workers, updated daily. A: That's serious data infrastructure, and it's just... open in a browser tab for anyone. B: And a sobering counterweight: the Lancet Oncology study finding infections caused two point three million cancer cases in 2024 — about twelve percent of new cases — led by H. pylori and HPV.

Mia: A: Which means a meaningful share of cancer is preventable with things we already have. B: And that brings us to our last pair — how we live and how we read. A: So, the hosting statistic: in 1975, forty-two percent of Americans hosted friends or family monthly. By 2026, that's twelve percent. B: A seventy percent collapse in hosting over fifty years. A: Which is a hard number to read any way other than downward — less in-person community, sustained across generations.

Mia: B: And the counterpoint, or maybe the companion, is the Ask HN reading thread — Dungeon Crawler Carl is hugely popular there, sitting right alongside recommendations like Seeing Like a State and Normal Accidents. A: LitRPG next to books about how states fail and systems go wrong. B: Less hosting, more solitary reading. Which, honestly, is a fitting place to end — and also a good reminder to invite someone over. A: Ha. On that note, thanks for listening, everyone — we'll see you next time.

Mia: B: Goodnight.