
1114 | Grounded: One Day of Outages and AI News
节目备注
A nationwide FAA ground stop, a wave of cloud outages, new AI decision models, and a haul of smaller stories from open source to research.
Timeline
- 00:00:04 Opening
- 00:00:47 The nationwide ground stop
- 00:05:36 The cloud outage wave
- 00:10:10 AI models and strategy shifts
- 00:13:20 AI slop and trust
- 00:15:46 Research highlights
- 00:18:01 Security and lost hardware
- 00:19:50 Tools and tinkering
- 00:23:27 Closing
Related links
- FAA lifted National Ground Stop at 0119Z
- Amtrak booking system experiencing major issues amid nationwide FAA ground stop
- Microsoft outage gives customers a crash course in dependence on Redmond
- Elevated errors on claude.ai, Claude Code, and platform.claude.com
- All US airports are currently ground stopped
- FAA issues national ground stop
- Live airband on all frequencies from KORS. Victoria, CA also ground stopped
- Doomsday: Friday, 13 November, A.D. 2026 (1960)
- Microsoft Outlook/365 still down after 12 hours
- Big tech is here to let us down
- LibreOffice actually works 365 days a year
- U.S. October PPI Rises 6.5% YoY, Highest in Nearly Four Years
- Perplexity Decider v1.3: Training, eval, and use cases
- OpenAI disbands its robotics team as its general models start running robots
- After 600 experiments, here's what I discovered about JEPA world models
- Weight Space Misfolding
- The Dawn of the Data Intelligence Engineer
- No AI Slop: A Manifesto
- AI Is Destroying the Trust Between Writers and Readers
- It Takes 16 Clicks for a Student to Open a PowerPoint on an iPad
- AI unlocks Gödel's notebooks, said to contain hundreds of proved theorems
- Found: A protein that drives ageing without unleashing inflammation
- Apify Security Incident
- $11.4M in Nvidia hardware vanishes during shipment from CA to TX
- Minnesota guitarist's "palmfrets" invention strikes a new chord
- FM-1 Workbench
- IceWM
- A technical vision for the Haskell ecosystem
- Why Zig over C for Systems Programming
- Lonewolf 0.3 – A standalone XMPP server written in Rust
- Show HN: Isometric SF – A pixel-art map of the SF Bay Area
This episode is produced by Bri. Bri uses advanced AI technology to turn the feeds you care about into podcasts made for listening. Contact us at hi@bri.so.
文字稿
Mia: Welcome back to the show, everyone — it's Mia. And it's Milo. And if you were anywhere near the internet in the last twenty-four hours, you know the theme already: everything broke, mostly at once, and the comment sections lit up accordingly.
Milo: It was one of those days. The FAA grounded the entire country's air traffic, Microsoft 365 went sideways for half a day, Claude started throwing authentication errors, Amtrak's booking system collapsed — and in the middle of all that, people still shipped cool stuff. So we'll do it the way we always do: set up the story, then spend most of our time inside the discussion around it.
Mia: Let's start with the big one, because it's hard to talk about anything else first. On Friday, November 13, the FAA ordered a nationwide ground stop — every airport in the US. Flightradar24 reported it as citing an FAA equipment outage, and other reporting pointed at a telecom outage specifically. It was the first nationwide ground stop since January 2023.
Milo: And the timeline is actually quite tight. The order — designated ADVZY 005, if you like reading NOTAM paperwork — was effective from 0119 Zulu to 0200 Zulu on November 14, so roughly an hour of formal ground stop, with an update promised at 0400 Zulu. So on paper it was short. On screens and in terminals, it looked like chaos.
Mia: The comment threads around this split pretty cleanly into three camps. One camp is the systems-reliability people, who looked at "telecom outage" and said, okay, fine, but why does a telecom failure at the FAA still cascade into every airport in the country in the year 2026? Their argument is basically redundancy math: if one failure mode can take down the whole national airspace, the redundancy story isn't real, or it's at least not the redundancy people assume exists.
Milo: Right, and the rebuttal from the other side was interesting. Several commenters pushed back and said, look, the system worked as designed. A ground stop is the conservative response to a data problem. If the feed that tells controllers where planes are supposed to be becomes unreliable, you don't keep flying and hope. You stop, you verify, you restart. From that view, an hour-long nationwide stop is the safety system functioning, not failing.
Mia: And then a third group just... refused to take a side and asked the harder question, which is: what exactly broke? Because the framing shifted between "equipment outage" and "telecom outage," and nobody in the discussion had a definitive answer. There's a real gap between the official one-liner and an actual root cause. That's the unresolved piece here — we don't yet have a post-mortem saying which circuit, which provider, which handoff.
Milo: The firsthand texture in the threads was great, though. Flightradar24 documented the whole thing in real time. And there's this hobbyist project called Orcas Airband — a software-defined-radio site that streams and transcribes all the air traffic from a receiver at KORS — and it captured an actual confirmation of a ground stop reaching Vancouver. So you've got amateurs with a receiver producing primary-source documentation of a national event. That's the part I love about these discussions.
Mia: Officials did address the paranoia angle, for what it's worth. The homeland security secretary said there was no indication of anything nefarious. Which, in the thread, produced the predictable argument: one side said "good, so it's boring infrastructure, that's fine," and the other said "that's exactly what the scary version would want us to think" — which is not really an argument, but it was there.
Milo: The ripple effects were real, though, and one of them is funny. Amtrak — the train company, the alternative to flying — had its booking system down the very same Friday. And their stated reasons were increased demand and a system-wide update. So the trains, on the day everyone's flights got grounded, couldn't sell tickets. People in the comments had a field day with that, and honestly deservedly so.
Milo: Some read it charitably — demand spike from grounded travelers plus a coincidentally scheduled update is a brutal combination. Others read it as a sign that Amtrak's infrastructure is as thin as everyone fears.
Mia: And then there's the wink at the end of this one. A 1960 paper in Science — an actual, peer-reviewed population study — predicted "doomsday" for November 13, 2026. Exactly this date. The commenters had fun with the failed extrapolation, and the takeaway a lot of them landed on was a caution about exponential extrapolation itself: you extend a curve far enough, you get an apocalypse on a specific calendar day, and reality just declines to cooperate.
Milo: Okay, from one kind of outage to the whole genre. Because the FAA was not alone. Microsoft 365 started degrading at 1826 GMT on November 12 — that's Exchange Online, Teams, Copilot — and it stayed degraded into November 13 for over twelve hours total. The eventual explanation: failed server health checks under elevated load, across multiple regions.
Mia: And the same day, Anthropic's Claude had elevated authentication errors across claude.ai, Claude Code, and the platform. So two major AI-adjacent vendors stumbling within the same window.
Milo: Let's get into the discussion on this one, because the Microsoft thread got philosophical fast. The big current was concentration risk. The argument, which one blogger also made explicitly in a post we saw, is that when Microsoft, AWS, and Google all have visible incidents in a short span, you're no longer looking at bad luck — you're looking at a structural property of a system where enormous amounts of the economy sit on a handful of vendors.
Milo: One commenter's framing that stuck with me: we optimized for cost and convenience and accepted fragility as the trade, and now the trade is being repriced in public.
Mia: The counterargument came from the engineers in the room, and it's a fair one: outages at scale are not evidence of carelessness. Three nines at Microsoft's scale is still an enormous amount of downtime in absolute hours, and small percentages of failure on gigantic fleets are statistically inevitable.
Mia: Their point was that the concentration complaint and the engineering complaint get conflated — the fix for concentration is architectural choice by customers, the fix for outages is better failure isolation, and those are different problems.
Milo: Which loops into the layoff angle. The same blogger tied the failures to mass layoffs in big tech — the claim being that the operational margin that used to absorb this kind of thing has been cut. Now, I want to be careful here: that's a hypothesis, not a demonstrated cause. And in the comments, plenty of people said exactly that — correlation is easy, staffing-to-incident causal chains are hard.
Milo: But others pointed out that Copilot reliance is its own oddity: the AI layer that was supposed to make work faster was itself part of the outage surface. When Teams and Copilot go down together, the thing you bought for resilience is another thing that broke.
Mia: And then there was the comedy beat. LibreOffice tweeted that it "works 365 days a year" while Microsoft 365 was down. Peak open-source snark. But the comments actually fact-checked the joke, which I respect — several people noted that offline versions of Microsoft Office also still work. The outage hit cloud services, not the desktop apps. So the dunk was funnier than it was accurate, which is probably the most internet thing that happened all week.
Milo: One more piece of the macro backdrop, and the commenters pulled it in themselves: October PPI came in at 6.5 percent year over year — the highest since November 2022 — though it was slightly below the 6.6 percent expectation, up 0.3 percent on the month. And the rate-policy debate reignited immediately. One side argued that persistent producer-level inflation means the policy tool isn't transmitting the way models say it should; the other side said single print, don't build a narrative.
Milo: It threaded neatly into the outages discussion too, because some people argued that reshoring and redundancy — the obvious response to concentration risk — are themselves inflationary. That's a genuinely interesting loop.
Mia: What's still unknown here is the boring-but-crucial stuff: Microsoft's actual root cause post-mortem, and whether the Claude auth issues were related or coincidence. Nobody in any thread claimed to know. Fine. Let's move from things breaking to things shipping — because the AI vendors kept moving anyway.
Milo: Perplexity released new decision models: Decider v1.3 and Flash v1.3. Apache-licensed, up on Hugging Face, available through OpenRouter and Perplexity's own API. Plus a new feature called "locate" for extraction. The discussion angle here is open licensing as strategy — Apache means anyone can run these, which reads as a land-grab for the "decision model" niche rather than a moat play.
Milo: Some commenters liked the openness on principle; others asked what the actual differentiation is when everything is Apache and everything is on the same three endpoints.
Mia: The bigger strategic story, though, is OpenAI. They disbanded their standalone robotics team and reassigned the researchers. The bet, as reported: general models — GPT-6, codenamed Astra — can control robots directly, with a humanoid demo targeted for 2027. And this maps onto a genuinely interesting paper discussion we saw, a 600-experiment study on JEPA world-models.
Mia: The author himself showed up to explain it, and was very clear that JEPA is a niche thing for spatial prediction and robotics, not an LLM replacement. So you've got one camp — OpenAI — betting everything on general models reaching into embodied tasks, and another research tradition saying prediction-for-robotics is its own specialized problem. Those are incompatible bets, and the comments didn't resolve it.
Milo: There was also a security-flavored paper called "Weight Space Misfolding" — the idea that model-mediated weight editing could corrupt other models' facts or refusals, demonstrated on Qwen2.5-1.5B. Important caveat though: no propagation shown. The demos show corruption of a target model, not that it spreads through an ecosystem.
Milo: And the commenters were disciplined about that distinction — half the thread was people making sure nobody overread it into "AI viruses," and the other half asking what it would take to actually show propagation.
Mia: And then MotherDuck's CEO floated a new job title: the "Data Intelligence Engineer." The model is that AI agents handle routine data work and humans validate AI-computed answers. The thread split the way you'd expect. One side said yes, finally, a job description that matches reality — validation is the actual work now.
Mia: The other side said this is just a rebrand of analyst with a euphemism attached, and that "humans validate AI answers" quietly assumes the validation is cheaper than the doing, which is unproven. Nobody had a clean resolution, but everyone agreed the division of labor is genuinely shifting. Which brings us to the quality and trust question, because that's the shadow over all of it.
Milo: So: AI slop. There was a manifesto out of the AIE Code Summit 2026 called "No AI Slop," and its core definition is actually sharp: slop is effort below the reader's cost. That is, if the reader spends more attention consuming it than the creator spent producing it, it's slop. The demands: outcomes over output, human review of AI work.
Mia: The responses were more divided than you'd think. Some people loved the definition because it's testable — it's about effort asymmetry, not vibes. Others said it's unenforceable: who measures the reader's cost? And a practical faction said the manifesto is fine but toothless without incentives, because the reason slop exists is that someone profits from it.
Milo: And the reason it exists is on full display in the Atlantic piece: AI is flooding Amazon with scam books, enabled by a system that allows six uploads per week. The framing that hit hardest in the comments was about trust — specifically trust between writers and readers. The argument: discovery systems assume uploads are acts of intent, and at six per week, upload volume itself becomes a strategy that crowds out intent.
Milo: Some counterpoints appeared — people noting that self-publishing always had a slush pile and this is a change of degree, not kind — but even the skeptics conceded the degree is the problem.
Mia: The teacher thread was the micro version of the same disease. A Reddit post from a teacher claimed that opening a PowerPoint on an iPad takes sixteen clicks, blamed on the school SSO tool Clever. Now — flag this — commenters disputed it. Some said the count is exaggerated or that the real problem is misconfigured tooling, not the SSO category.
Mia: Others shared their own school-tech horror stories and said sixteen clicks, sixteen hundred, doesn't matter, the point is nobody in procurement experiences the workflow as a teacher does. So it's disputed, but the discussion it produced was about how edtech gets bought by people who never use it.
Milo: And then the counterexample to all of this — the thing that shows where AI is still genuinely valuable, in a way nobody disputes. Claude helped decode Gödel's workbooks. We're talking about 1,009 pages of Gabelsberger shorthand, and the result was hundreds of previously unknown proved theorems, sitting in archives all along.
Mia: And the commenters immediately connected it back to the slop debate. This is the cleanest possible contrast: one system producing effort below reader cost at industrial scale, another system unlocking decades of unread genius. Same underlying capability, wildly different outcomes depending on what it's pointed at. A few people also made the point that Gödel's theorems were "unknown" but provable all along — the bottleneck was never mathematics, it was decoding.
Mia: That's what makes this a good AI use case: the hard part is transcription, and the value is sitting right there once you crack it.
Milo: Biology went in a different direction entirely. A study published in Immunity and reported through Nature looked at granzyme K — a protein that appears to drive tissue ageing. In mice, lacking granzyme K meant improved heart and liver function and less kidney damage. And here's the interesting part: potentially without impairing immunity. Because granzyme K is an immune protein, and the obvious worry with any immune-targeting intervention is that you break the immune system to fix the ageing.
Mia: The discussion there was cautious, appropriately. The mouse-to-human gap is enormous and everyone knows it. But the specific hook — a protein you might suppress for anti-ageing benefits while keeping immune function — is exactly the profile a real drug would need. The open questions in the thread were the right ones: mechanism confirmation in humans, whether the heart and liver improvements trade off against something not yet measured, and timeline. Nobody promised anything. Good thread.
Milo: From there, let's get into the incidents — one digital, one physical, both supply chain.
Mia: Apify disclosed unauthorized access on November 9 and 10, via a compromised credential. Their response is worth noting: API tokens are now stored only as hashes — so a future database compromise doesn't hand over live tokens — and a post-mortem is due within a week. The commenters appreciated the hashing change, because token-at-rest-in-plaintext is embarrassingly common across the industry.
Mia: The open question everyone kept returning to: what did the attacker actually access during those two days? That's what the post-mortem has to answer, and hashing the tokens going forward doesn't retroactively answer it.
Milo: And the physical one is stranger. $11.4 million of Nvidia hardware — six pallets — vanished on August 29 during a shipment from Oakland to Fort Worth. The truck was last tracked in San Bernardino County. Full stop. That's where the story ends as far as anyone knows.
Mia: The comments treated it half as a heist movie and half as a logistics audit. The reasonable read: high-value electronics freight gets targeted because it's dense, portable, resaleable, and tracking is apparently a breadcrumb trail rather than a tether. The parallels with the Apify case were actually made explicitly — both are failures of basic hygiene at the handoff points. One is a credential, one is a truck.
Mia: The pattern is the same: value concentrates, and the seam between two responsible parties is where it leaks.
Milo: Alright, let's close on the stuff that just makes people happy — the tools and tinkering pile. And there's a lot of it this week.
Mia: Start with music. A Minnesota musician named Ben Ranson patented "palmfrets" — back on October 6 — a device that lets guitarists fret bass notes with their wrist, freeing the fretting hand for other work. The reaction in the comments was a perfect split: some players immediately saw the harmonic possibilities, and others said, with real affection, that this is the kind of thing that gets patented and then vanishes. Time will tell.
Milo: Cheaper and more immediately playable: the M-VAVE FM-1 Workbench — a browser-based editor for a fifty-dollar eight-track FM synth. And the story isn't the hardware, it's the ecosystem — custom firmware like sloop, Jiant, groovebox setups. Fifty dollars and a community. That's the part that made people nostalgic.
Mia: Nostalgia was also the mode for IceWM — a fast X11 window manager that has existed since 1997 and just shipped version 4.1.0 in August 2026. The comment section was basically a support group for people who still run tile-free, featherweight window managers on machines from another decade.
Milo: On the developer infrastructure side, the Haskell Foundation's technical committee unveiled a vision to unify tooling — and the framing of the problem is honest: the ghcup, stack, and nix confusion is a genuine onboarding wall. The committee's position, and some commenters agreed, is that the pieces exist; the ecosystem has the libraries, what it lacks is a coherent front door. Skeptics pointed out that Haskell has had unification visions before, and the question is adoption, not blueprints.
Mia: The language-war item was an article arguing Zig beats C for new systems code, citing errdefer and safe slices. And the thread did its usual job: supporters made the ergonomics case, and critics pointed at real bugs in the Zig musl rewrite as counterexamples, plus — this is very 2026 — flagged the article itself for AI-slop-like writing, which loops all the way back to our manifesto discussion. Effort below reader cost, allegedly, in a systems-programming blog post. Nothing is safe.
Milo: Two more and then we'll get out of here. Lonewolf 0.3, a Rust XMPP server, added federation, direct TLS per XEP-0368, and private XML storage — still pre-1.0 and single-node, and the comments were the classic small-server community mix of enthusiasm and realistic warnings about running your own infrastructure.
Milo: And finally, Isometric SF: a pixel map of 1,300 square miles of the Bay Area, built from Google 3D tiles, using fine-tuned small open-source models — reportedly about seven thousand dollars cheaper than doing it with Nano Banana 2. Which, given everything else we talked about today, is a nice note to end on: small models, small budget, one person, one city.
Mia: And maybe that's the actual throughline of the day. The big systems — telecom, cloud, national airspace — cracked in public, and the response in every thread was the same: people documenting with SDR receivers, fact-checking LibreOffice's jokes, asking for post-mortems, patenting wrist-frets. Fragile up top, resilient underneath.
Milo: That's the episode. Thanks for listening — we'll see you when the next thing breaks, which given this week is probably soon.