Episode notes
OpenAI posted a computer-checked proof that three-dimensional Navier-Stokes can blow up when a smooth extra force is added, and says it will not claim the Clay Millennium prize. That is not the unforced prize problem. NYU's Tristan Buckmaster says the lab moved after hearing about his line of work. OpenAI says it did not look up live user data and cannot rule out de-identified past use. Also: a watchdog found 332 paid ads with child sexual abuse imagery that passed Meta's review; Meta launched Muse, a US agent that can open email, calendar, and payments, while Reuters says internal tests still hit data exposure; Britain's NATS blamed a flight-processing fault for disrupted UK departures and officials say it was not a cyber attack. The side of tech news nobody talks about.
Hosts: Alex & Jordan
Show: Chief Skeptic Officer — The side of tech news nobody talks about.
Drop: Daily at 7:00 A.M. America/New_York
Episode date: 2026-09-09
In this episode
Forced proof, unclaimed prize — OpenAI claims a Lean-checked blowup with a smooth force, Clay C and D, and will not seek the million dollars. Clay has awarded nothing. A rumor on 1 September preceded a large agent run. Buckmaster's written account is that credit was offered only if his Anthropic-employed collaborator was dropped. OpenAI cannot rule out de-identified training data.
Paid ads, approved anyway — Tech Transparency Project counted 332 paid CSAM ads that passed Meta review in 2026. Some used real photos of minors. After a 2 September briefing, remaining visible ads came down. Meta says most had few impressions and little spend. The report button often said no violation.
Muse wants the inbox — Meta launched Muse in the US. Email, calendar, payments, a card even on the free tier. Confidential encryption of the whole VM is later this year. Reuters reported internal tests still showing stalls and data exposure. No confirmed public customer leak.
Processing fault, passengers wait — NATS said a flight-processing issue disrupted UK departures on 8 September. Airspace stayed open. A fix landed the same afternoon. Recovery outlasted the fix. BBC reported officials were expected to say it was not cyber, and not the 2023 system.
Links
AI disclosure
This episode was created with artificial intelligence. Alex & Jordan are AI hosts; their voices and conversation are generated with AI. Research and editorial judgment shape the skeptic angles; we do not invent quotes, scores, or viral claims about the news.
Transcript
Alex and Jordan, turn by turn. Tap a line to jump in the player.
0:00
Alex
OpenAI posted a computer-checked proof that a famous fluids equation can blow up when you add a smooth force. It will not claim the million-dollar prize.
0:10
Jordan
A watchdog found 332 paid ads with child sexual abuse imagery that passed Meta's review and ran on Facebook and Instagram this year.
0:20
Alex
Meta launched Muse in the United States, an AI helper that can open your email, calendar, and payments.
0:28
Jordan
A fault in Britain's flight-processing system disrupted UK departures on September eighth.
0:35
Jordan
That's the board. Stay for the audit. We open those up. Today's Chief Skeptic Officer.
0:52
Jordan
What if there was an AI that researched the tech news, checked the sources, and asked what the tech news is not telling you?
1:01
Alex
That's us. I'm Alex.
1:03
Jordan
And I'm Jordan.
1:05
Alex
You're listening to Chief Skeptic Officer. The side of tech news nobody talks about.
1:10
Jordan
Every day at seven A.M. New York time. Wherever you get your podcasts.
1:14
Alex
OpenAI says it has a proof about the three-dimensional Navier-Stokes equations, the math used to describe how fluids move. Its claim is that the equations can break down in finite time when a smooth outside force keeps pushing the fluid.
1:29
Jordan
And that force is the key. The famous million-dollar version asks what happens with no extra force. OpenAI claimed the version with one. Clay calls that the C and D case.
1:40
Alex
Right. They did not prove the unforced prize problem. They say this different proof was checked in Lean, a computer system that checks each math step. OpenAI will not claim the Clay prize, and Clay has awarded nothing.
1:55
Jordan
I'm on Hacker News, a tech comments site. The first comment traces this back to a September first rumor that two other famous math problems had been solved. OpenAI says it then ran about ten thousand AI programs at once across the problems that remained.
2:10
Alex
That is our one scale number. It tells you how fast the lab could pile onto a lead. It does not tell you the proof is right.
2:19
Jordan
The rumor also pointed toward Tristan Buckmaster at NYU and Levent Alpöge, an Anthropic employee working with him on his own time?
2:28
Alex
Yes. They were studying related fluid math, not the prize claim. Buckmaster says OpenAI heard about their direction, moved on it, then offered him credit only if Alpöge was left off.
2:41
Jordan
So the dispute is not, "Buckmaster solved Navier-Stokes first."
2:45
Alex
No. It is about who opened the path and how OpenAI treated them after it rushed in.
2:51
Jordan
Buckmaster also says Sébastien Bubeck asked why he would "ruin your career" while Buckmaster was talking publicly and refusing to drop Alpöge. That is Buckmaster's written account. There is no recording.
3:03
Alex
OpenAI denies any live lookup of user data, but says it cannot rule out past, de-identified product use helping train its models.
3:13
Jordan
Wait. That does not show the lab used these mathematicians' chats. It means OpenAI cannot close that question.
3:21
Alex
Exactly. I'm still in the middle of a logging migration at work. I would not close our highest-severity incident because a lab said a computer checked the steps. I want mathematicians to review the proof, and I want the trail.
3:32
Jordan
So if the proof holds, the conduct question...
3:36
Alex
Still holds.
3:38
Jordan
Then sharing a promising idea has a new cost. A lab can race down the path before your own paper is ready.
3:47
Alex
Let the proof review run. Keep Buckmaster's account under review too.
3:52
Jordan
The Tech Transparency Project found 332 paid ads containing child sexual abuse imagery that ran on Facebook and Instagram this year. Meta reviewed and approved them before they ran.
4:06
Alex
Some used real photos of minors. The watchdog says 274 of the 332 ran in August.
4:15
Jordan
Meta had already told WIRED that new AI would block violating ads when they were uploaded. More still got through.
4:23
Alex
The page above us is text only. The finding is simple: these were paid placements that cleared Meta's review, not ordinary posts waiting for someone to report them.
4:33
Jordan
Someone bought the space. Meta approved the material.
4:37
Alex
Yes. Then the public report route failed too.
4:41
Jordan
TTP reported 129 ads that were still live. Meta replied that 73 did not violate its Advertising Standards.
4:51
Alex
Meta's own policy says ads must not contain content that sexually exploits or endangers children. It calls that zero tolerance.
5:01
Jordan
Hours after TTP briefed Meta on September second, the remaining visible ads came down. Meta later changed 113 takedown notices, the labels it puts on removed material, from adult-policy language to child-exploitation language.
5:16
Alex
So the classification changed after the watchdog arrived.
5:21
Jordan
Yes. The report button often could not reach that answer before the briefing.
5:26
Alex
Meta told the BBC that most ads had fewer than 200 impressions. That means each was shown fewer than 200 times. It also said total spending was under five thousand dollars.
5:38
Jordan
Small reach and small spending do not make approval a small failure. Other major ad networks manage zero paid ads in this category.
5:49
Alex
The Hacker News thread is thin. One defense is that 332 is tiny beside Meta's scale. Another says generated images should count differently.
5:59
Jordan
But the watchdog found real photos in some of them. And scale is a bad defense for a system that accepted payment and cleared the material.
6:08
Alex
We do not need to claim Meta wanted this. The evidence says its controls allowed it.
6:14
Jordan
The upload scanner missed it. Review passed it. The report route often defended it.
6:21
Alex
Meta should be judged by the ads it approved, not only by the policy it published.
6:27
Alex
Meta also launched Muse in the United States. It is an AI helper that can open your email and calendar, browse the web, send messages, and make payments.
6:37
Jordan
After that last story, Meta is asking for a lot more trust.
6:41
Alex
The free use has a cap. Paid plans cost twenty or one hundred dollars a month, and you need a payment card even to start.
6:49
Jordan
Wait, the free start needs a card?
6:53
Alex
Yes.
6:54
Jordan
Then the first step is already identity, billing, and permission to act.
7:00
Alex
Meta's page calls Muse the "world's first personal AI agent built for everyone." That is marketing. The product did launch, and it is US only.
7:10
Jordan
Training is opt out, so users must turn it off. Meta says Muse conversations are not shared with its ads system.
7:18
Alex
It also promises a Confidential VM, an encrypted computer it says will arrive later this year. That is not part of today's launch.
7:27
Jordan
The safety page above us says, "How We Built Safety Into Muse."
7:32
Alex
Reuters says Meta delayed the launch from April over security. In internal employee tests this week, Muse still stalled, failed without warning, and exposed data inside the test flow.
7:45
Jordan
Testers also got around the safety limits in at least one test, and personal photos were exposed?
7:52
Alex
Yes. Those were company tests, not a public customer leak. Meta did not comment to Reuters on those incidents.
8:01
Jordan
Every week, I get a classroom-AI pitch that wants the same inbox. Let it read the teacher's mail. Let it book the meeting. Let it fix the calendar. The permission list arrives dressed as convenience.
8:14
Alex
And you hear inventory.
8:17
Jordan
I hear the product. Access to your life is what makes Muse useful. It is also what makes one silent mistake expensive.
8:26
Alex
So if it can read, click, send, and pay...
8:31
Jordan
Then "mostly works" is not a soft launch condition.
8:35
Alex
Hacker News keeps circling trust. Some commenters like the browser inside the flow. Others already refuse to connect agents like this to personal accounts.
8:45
Jordan
I do not think polish is fake. Good design can make risky power easier to control.
8:52
Alex
Or easier to grant.
8:55
Jordan
Right. That caught me. A calm screen may get more permission than a rough tool would.
9:02
Alex
Muse is asking for inboxes and payment cards now. Meta still owes users proof that its controls survive outside employee tests.
9:12
Jordan
On September eighth, Britain's air-traffic provider NATS had a fault in its flight-processing system. That is the computer that files and sequences departures.
9:21
Alex
UK airspace stayed open, so planes already in the air could keep flying. Takeoffs were the squeeze. Departures were restricted, and flights were cancelled.
9:31
Jordan
NATS began investigating at 1:46 P.M., found the issue by 3:25, applied a fix at 4:40, and declared it resolved at 7:30.
9:41
Alex
The NATS page above us says only that a technical issue disrupted flight departures. We still do not know the kind of bug.
9:50
Jordan
FlightRadar24 estimated up to one thousand cancellations. Ryanair said 65,000 passengers were delayed. Those are separate company estimates.
10:01
Alex
So the computer was fixed at 4:40, but the day's schedule...
10:06
Jordan
Was not. Aircraft and crews were already out of place. Airlines had lost slots, their assigned takeoff times, and the backlog kept moving through the network.
10:17
Alex
NATS repaired the system. Passengers carried the recovery.
10:22
Jordan
The BBC says the transport secretary was expected to tell Parliament that a cyber attack was categorically not the cause.
10:31
Alex
The Hacker News thread keeps reaching for a different NATS failure in 2023. This was a different system. We should not borrow the old failure's cause.
10:41
Jordan
Agreed. Then why could one broken processor squeeze departures for hours without a public spare taking over?
10:49
Alex
By public spare, you mean a backup system that should have taken over.
10:54
Jordan
Exactly. NATS has not told the public what backup was available or why it did not prevent the disruption.
11:01
Alex
We do not know the exact bug. No vendor has been named, and there is no disclosed software flaw to pin on a version.
11:10
Jordan
So "technical issue" is all the public gets.
11:14
Alex
Careful. That is all NATS has confirmed. Missing detail is not proof of a cover-up.
11:21
Jordan
You're right.
11:23
Jordan
The doubt is simpler. A system important enough to shape a country's departures failed, and the public cannot see what broke or what backup should catch it next time.
11:34
Alex
Some comments say the whole airspace went down. It did not. Planes could keep flying. The pressure was on takeoffs.
11:41
Jordan
Passengers still lost the day.
11:45
Alex
Reliability includes recovery. If the backlog lasts for hours, the incident does too.
11:52
Jordan
NATS should publish the failure and backup plan.
11:56
Alex
That's our audit for today. Find us wherever you get your podcasts. Chief Skeptic Officer, every day at seven A.M. New York time.
12:04
Jordan
Tell us what you're skeptical about. Drop it in the comments, the angle you can't stop chewing on.
12:11
Alex
Stay curious. Stay skeptical.
12:14
Jordan
Doubt both.
