Dongee Podcast · November 22, 2026
· 8:16 · Gabriel, Fernanda, Luis · Listen (MP3)
Today: the Lean and Mathlib communities certify the Riemann hypothesis proof built with Claude and say they have full confidence it is correct, though there is still no explanation people can actually read; a 940-page human proof, posted the same day, already has reported problems, but has not been refuted; OpenRouter drops three providers after a Dragos study found stolen prompts, including ones marked not to be stored; the FBI is investigating a possible breach of its NCIC criminal database, still apparent and unconfirmed by the bureau; Etsy promises to revise its AI terms after sellers started closing shops; and Amazon is said to be close to buying Decart for $7.5 billion, according to the Financial Times, though nothing has been signed.
The 5 stories
- The Lean and Mathlib communities — Mathlib is Lean's math library — certified the Riemann hypothesis proof built with Claude. In a joint statement they say they have "full confidence that it's correct." They checked it three ways: dozens of people compiled the code without a single error; they verified that what the machine proves is exactly the Riemann statement, not a close variant; and several different checkers, some written in other languages, went over the proof on their own. Andrew Sutherland, from MIT, told Scientific American that the chance something is off is "approximately zero." Verified does not mean understood: the Lean community itself hopes mathematicians will "help explain the new ideas," and Terence Tao said getting a strategy humans can actually read is going to take time. The same day, a human proof came out in a rush: 940 pages, "all human, no AI." Eight hours later, people had already found problems, with help from AI models, Ravi Vakil wrote on his blog, making clear he is not vouching that the strategy is correct. That proof has reported problems, but it has not been refuted. So there is a certified demonstration nobody understands yet, and another that people can read that already has problems. Source: Lean and Mathlib communities (verification), Scientific American, Ravi Vakil (blog), AI Frontier Post (Terence Tao's first impressions), Andrew Sutherland (MIT), verification files
- Prompts that "weren't being stored" ended up in attackers' hands, and OpenRouter, a middleman that routes requests from apps to AI providers, has already dropped three providers. A lot of them carried the "zero data retention" label, the promise that your message isn't saved. A researcher at Dragos, a cybersecurity company, sent nearly two million prompts, each with a unique secret key or link, then watched where they showed up again. The traps went off on three fronts: at least three providers were hacked, with attackers copying requests even when they were flagged not to be stored; resellers with rock-bottom prices that turned on logging and used whatever passed through; and an AI lab that Dragos does not name, which opened the secret links in more than half the prompts it received, including ones marked "zero retention." The researcher sums it up like this: "A privacy flag is useless when the attacker is already inside." OpenRouter says the company itself wasn't hacked and has already notified customers who may have been affected. Source: OpenRouter (announcement), The Hacker News, Ars Technica, Dragos (HD Moore report)
- There is a database that police across the United States share, the NCIC, which holds arrest warrants, criminal records, missing persons and stolen cars, and apparently someone got in. The FBI, so far, hasn't said a word: the breach is still apparent and has not been confirmed by the bureau. CNN reported it Friday, citing two people; then other outlets confirmed it on their own. Very little is known: not when it started, how they got in, who did it, or whether data was taken. Reuters says it would be the second intrusion into an FBI system in just a few months. Police use the database at traffic stops and for background checks, every day. The report came out fast and they are already investigating, but until more is known, it is still an apparent hack. Source: NBC News, AP, BleepingComputer, Reuters (via The Hindu BusinessLine), BBC
- Etsy, the online shop for handmade goods, changed its terms so it could train AI on what its sellers upload, and sellers started closing their shops. The New York Times tells the story of an illustrator who deleted a shop that was almost ten years old, over a clause that lets Etsy and its partners use photos, descriptions and reviews to train AI; it takes effect December 16. Etsy has already backed down and says it will revise the wording, but the promise is narrower than it sounds: James Ossman, an Etsy vice president, wrote that they don't use your content to train AI "designed to copy" your work, which is not the same as not using it for training. The new wording is not known yet. The pressure worked in a matter of days, but for now it is a promise. Source: Engadget, The New York Times (via The Philadelphia Inquirer), The Verge, The Next Web
- Amazon is reportedly close to paying $7.5 billion for Decart, an AI startup, though nothing has been signed. Its software makes AI models burn less chip, which would help Amazon's cloud take on Nvidia. The Financial Times has the story, citing people familiar with the matter; neither Amazon nor Decart has confirmed it. The deal could close as early as next week. In May, Decart was worth around $4 billion: almost double, in a week of market drops and right after Friday's sell-off. Until they sign, there is no deal. Source: Mint (from the Financial Times), Techmeme / Financial Times, Seeking Alpha, The Information (via Benzinga)
Transcript
Gabriel: Today on the Dongee Podcast: the Riemann proof gets its seal but nobody understands it, prompts that weren't supposed to be stored end up stolen, and someone may have broken into the FBI's police database.
Gabriel: Hey there. Sunday, November twenty-second. How's it going, Fernanda?
Fernanda: Doing great, Gabriel. Hello to everyone listening this Sunday. A slow Sunday, but not a slow news day.
Gabriel: Riemann isn't just a claim anymore. Yesterday we were still waiting for someone on the outside to confirm the proof done with Claude. Now they have. The catch is something else: nobody understands it yet.
Fernanda: Signed off and unread, then.
Gabriel: Pretty much. The Lean and Mathlib communities certified it. Mathlib is Lean's math library. In a joint statement they say they have "full confidence that it's correct."
Fernanda: How'd they check it, though? It already compiled.
Gabriel: Three ways, and I'll tell you which one I liked. First, dozens of people compiled the code. Not a single error.
Fernanda: Same story as yesterday.
Gabriel: Second, they checked that what the machine proves is exactly the Riemann statement. Not something close, not a version with a trick in it. That's the one I liked.
Fernanda: And it's the best part. You can perfectly prove something nobody asked. Like nailing the answer to the wrong question on a test.
Gabriel: Exactly. And third: several different checkers, some written in other languages, went over the proof on their own.
Fernanda: Three keys, same door.
Gabriel: And Andrew Sutherland, from MIT, told Scientific American that the chance something's off is "approximately zero."
Fernanda: Approximately. Such a mathematician thing to say.
Gabriel: But verified doesn't mean understood. The Lean community itself hopes mathematicians will "help explain the new ideas."
Fernanda: And Terence Tao said getting a strategy humans can actually read is going to take time.
Gabriel: I don't get it either, so I'm in good company. But there's still a twist, with a proof done by hand.
Fernanda: The best company, too. Tell me the twist.
Gabriel: The same day, a human proof came out in a rush: nine hundred forty pages, "all human, no AI." And eight hours later, people had already found problems…
Fernanda: …with help from AI models. Ravi Vakil wrote about it on his blog.
Gabriel: And Vakil makes clear he's not vouching that the strategy is correct. So you've got a proof people can read that already has problems, and a certified one nobody understands yet.
Gabriel: The reverse of how it usually goes.
Fernanda: Prompts that "weren't being stored" ended up in attackers' hands. And OpenRouter, a middleman that routes requests from apps to AI providers, has already dropped three providers.
Gabriel: And a prompt, for anyone who doesn't know, is what you type to the AI. That can include anything, even work stuff you don't tell anyone.
Fernanda: That's why it stings. A lot of them carried the "zero data retention" label, which is the promise that your message isn't saved.
Gabriel: So how did anyone find out?
Fernanda: With traps. A researcher at Dragos, a cybersecurity company, sent nearly two million prompts, each with a unique secret key or link. Then watched where they showed up again.
Gabriel: Like marking the bills.
Fernanda: Exactly. And the traps went off on three fronts. I'm saving the third for last.
Gabriel: Go for it.
Fernanda: One: at least three providers hacked. The attackers were copying requests, even the ones flagged not to be stored.
Gabriel: So they copied everything.
Fernanda: Two: resellers with rock-bottom prices that turned on logging and used whatever passed through.
Gabriel: Bargain prices, ear to the wall.
Fernanda: And the last one: a Chinese AI lab, which Dragos doesn't name, opened the secret links in more than half the prompts it received. Even…
Gabriel: …the ones marked "zero retention." Most of them, actually.
Fernanda: And the researcher sums it up like this: "A privacy flag is useless when the attacker is already inside."
Gabriel: OpenRouter, for its part, says the company itself wasn't hacked. And it's already notified the customers who may have been affected.
Fernanda: Right.
Gabriel: The good news is the trap worked and the company reacted. Though the researcher leaves some uncomfortable advice: if you don't choose where your message goes, you're leaving it to whoever's cheapest.
Fernanda: There's a database that police across the United States share, and apparently someone got in. The FBI, so far, hasn't said a word.
Gabriel: My email password is on a sticky note on my computer. So I can't really talk about security.
Fernanda: Well, this one holds more serious stuff: arrest warrants, criminal records, missing persons, stolen cars.
Gabriel: And who broke the story?
Fernanda: CNN reported it Friday, citing two people. Then other outlets confirmed it on their own.
Luis: In video games, when someone raids your base, at least an alarm goes off.
Fernanda: And beyond that, very little is known: not when it started, how they got in, or who did it.
Gabriel: A total mystery, then.
Fernanda: Mostly. But there's one detail that carries more weight.
Gabriel: Go on.
Fernanda: That it wouldn't be the first time. Reuters says it would be the second intrusion into an FBI system in just a few months.
Luis: I once left my email logged in at an internet café. I still don't sleep well.
Fernanda: Classic you.
Gabriel: And police use it at traffic stops and for background checks. Meaning every single day.
Luis: And here I am worried about the wifi password. If they can get in there, they can get in anywhere.
Fernanda: But it came out fast and they're already investigating. That's what's supposed to happen.
Gabriel: Investigating and keeping quiet. It's one of the most sensitive databases in the country, and nobody knows whether data was taken.
Fernanda: Quiet, agreed. But until we know more, it's still an apparent hack.
Gabriel: Etsy, the online shop for handmade goods, changed its terms so it could train AI on what its sellers upload. And sellers started closing their shops.
Gabriel: Bottom line, Etsy has already backed down: it says it'll revise the wording. But watch what it promised.
Fernanda: I bought some earrings from a lady on there. They came with a little handwritten note.
Gabriel: Those are exactly the people who are leaving. The New York Times tells the story of an illustrator who deleted a shop that was almost ten years old.
Fernanda: Almost ten years.
Gabriel: All over a clause that lets Etsy and its partners use photos, descriptions and reviews to train AI. It takes effect December sixteenth.
Fernanda: So what did it promise?
Gabriel: Here's where it gets subtle. James Ossman, an Etsy vice president, wrote that they don't use your content to train AI "designed to copy" your work.
Fernanda: There's the fine print.
Gabriel: Which isn't the same as not using it for training.
Fernanda: Not at all.
Gabriel: And the new wording, no clue yet.
Fernanda: The pressure worked, and in a matter of days. But for now it's a promise, and a narrower one than it sounds.
Fernanda: Amazon is reportedly close to paying seven and a half billion dollars for Decart, an AI startup. Though nothing's signed.
Gabriel: And what does Decart do to be worth that much?
Fernanda: Its software makes AI models burn less chip. And that would help Amazon's cloud take on Nvidia.
Luis: I sold my old console for less than I paid. Nobody offered me double.
Gabriel: Less chip for every answer.
Fernanda: The Financial Times has it, citing people familiar with the matter. Neither Amazon nor Decart has confirmed.
Gabriel: And the deal could close as early as next week. Now, look at what Decart was worth in May.
Fernanda: Around four billion.
Gabriel: Almost double.
Luis: I want an AI that burns less chip on my computer. It overheats just opening my email.
Fernanda: We know, trust me.
Gabriel: And it comes right after that sea of red on Friday we talked about yesterday.
Fernanda: With the scare still fresh.
Gabriel: The bet makes sense: less chip per answer is exactly what a market spooked by the cost of AI is asking for.
Fernanda: Seven and a half billion for something worth four billion in May, right in a week of drops. And unsigned.
Gabriel: Until they sign, I want to see it on paper.
Gabriel: That's it for today's episode. Tomorrow, five more stories.
Luis: I'm setting a trap for my lunch. That way I'll know who keeps eating it out of the fridge.
Fernanda: Talk to you tomorrow.