Podcast 5-5 by Dongee · November 12, 2026

Transcript

Gabriel: Today on the Dongee podcast: Nvidia's board and the future of Jensen Huang, United Airlines against AI agents, and the AWS emulator written by agents.

Gabriel: Thursday, November twelfth. Hey. Lina, the Nvidia story has a new chapter… and this weighs more.

Lina: More than yesterday's?

Gabriel: Context in one sentence: we've had days of Jensen Huang, Nvidia's boss, and his relationship with a woman believed to be a Chinese spy. Now, according to The New York Times, there was a second relationship.

Lina: With whom?

Gabriel: With Kimberly Powell, an Nvidia executive, the head of the healthcare division. And she reported directly to him.

Lina: Ah, that's no longer just private life.

Gabriel: That's why, according to The Wall Street Journal, the board is preparing to oust him. It met Wednesday night to decide.

Lina: Careful. There's still no announcement.

Gabriel: Nothing, according to Reuters. But according to Bloomberg, Nvidia is already working with an outside firm to look for possible replacements…

Lina: …and even a temporary boss.

Gabriel: That's right. According to Semafor, a news outlet, some members prefer Colette Kress, the chief financial officer, as interim. And the board notified the White House about the meeting.

Lina: What a turn. On Sunday that same board said it had confidence in him.

Gabriel: And the market felt it: the stock dropped more than eight percent after the close. According to The Guardian, if that holds, that's six hundred billion dollars lost since the investigation began.

Lina: The board is still deliberating, but the market already voted.

Gabriel: And on Polymarket, a site for betting on the news, they give an eighty-eight percent chance he is out before the end of the year.

Lina: And in Washington, Democratic senators Elizabeth Warren and Chris Coons asked the government to keep the records of the chip sales to China.

Gabriel: Seen that way, the board has to act: a relationship with someone who reported to him, and earlier with a suspected spy, is already a matter of how the company is run and of national security.

Lina: Seen from the other side, removing the founder who took Nvidia to the top, without the investigation's conclusions and in the middle of the AI race, could hurt it more than it protects it.

Lina: Gabriel, would you give your card and your miles to an AI assistant so it can buy you a flight?

Gabriel: Sure. Let it find me the cheapest one while I sleep.

Lina: Well, with United Airlines, the U.S. airline, not anymore. It announced it is going to block AI assistants and agents from other companies from booking flights or managing accounts.

Gabriel: Agents… the ones that do things for you, right?

Lina: Those. An AI that doesn't just answer you: it gets into your account and pays. And the rule applies on its own channels and, indirectly, through agencies and search sites.

Gabriel: Wait. And why?

Lina: It says that giving your login details to an assistant "can expose sensitive information." And end in purchases, changes to your account, or spending of miles that you didn't want.

Gabriel: Well, that makes sense. I once bought the same ticket twice… and with no help from anyone.

Lina: I remember. And there's another point: check-in. According to United, doing it with an agent could violate federal rules…

Gabriel: …like the one where you confirm that your bag isn't carrying anything prohibited.

Lina: That one, the traveler confirms in person. But there's also a business angle: according to an aviation analyst, Henry Harteveldt, United doesn't want another middleman besides Google and the online agencies, or to pay for massive robot searches.

Gabriel: And the other airlines?

Lina: Southwest, another U.S. airline, did the opposite: it let its flights be seen inside ChatGPT.

Gabriel: Does United disappear from the search sites?

Lina: No. Its flights are still on Kayak, a flight search site, on Hopper, another travel app, and on Google Flights.

Gabriel: Ah, searching yes, buying no.

Lina: If you look at it as protection, an agent with your card and your miles can make a mistake or be tricked, and the airline is on the hook for security rules.

Gabriel: If you look at it as business, it's keeping control of the customer: it holds back a tool people already chose to use, and that other airlines do accept.

Gabriel: Lina, AWS, Amazon's cloud, released a program that, according to them, was written entirely by AI agents.

Lina: Mmm… entirely? I doubt it.

Gabriel: That's what Matt Goldberg, the engineer behind the project, says: "Our team has not written or reviewed any of the code."

Lina: Not even reviewed. How reassuring.

Gabriel: Let me explain. It's the AWS emulator: a test copy of the cloud that runs on your own computer, without internet. It simulates more than two hundred services, and it's free.

Lina: So you try things out without paying for the real cloud.

Gabriel: That's it. And it started as a weekend experiment. Goldberg asked Kiro, Amazon's coding assistant, for one piece of the cloud, the message-queue service…

Lina: …and it took a week.

Gabriel: Fifteen minutes.

Lina: I take longer inventing a password.

Gabriel: And today, according to him, there are hundreds of agents on three teams: some write tests against the real cloud, others look for bugs, and others fix them.

Lina: Agents reviewing agents.

Gabriel: He says he was inspired by an Anthropic demo, where Claude agents built a compiler, the program that translates code into machine language.

Lina: But it will have its limits.

Gabriel: According to an analyst, it doesn't simulate outages or usage limits. And the public list of bugs is in a folder literally called "things that don't work."

Lina: At least they're honest.

Gabriel: The exciting part: agents can already build and maintain large, useful software, if you give them clear instructions and a way to check their work.

Lina: The unsettling part: nobody reviewed the code, it's used to test real programs, it has known bugs, and the trust is that the agents review each other.

Lina: Gabriel, when you open a web page, there's a program on the other side that delivers it to you. One of the most widely used is Apache, the program that serves a huge number of web pages. And they found a critical flaw in it.

Gabriel: How critical?

Lina: Among the worst: it lets an attacker take control of the server from the internet, with no password.

Gabriel: Whoa.

Lina: And the curious part is who found it: separately, Anthropic's scanner, a program that looks for flaws in open-source code, and some researchers at Google Cloud.

Gabriel: So the AI saw it first…

Lina: …before the attackers, as far as we know. Apache already released the new version with the fix. And if you can't update, they recommend turning off the feature where the flaw is.

Gabriel: And are there a lot of servers at risk?

Lina: According to Wiz, a cybersecurity company, sixty-four percent of cloud environments have vulnerable machines, and fifty-one percent have at least one exposed to the internet.

Gabriel: More than half. And are they already using it to attack?

Lina: That's where they don't agree. Wiz says that at the time of publication, no attacks were known, and there was no public way to exploit it.

Gabriel: But someone says otherwise.

Lina: Red Hat, a software company. It says a public investigation already showed an attack that works against a server as it comes from the factory. And it rates it nine point eight out of ten for severity.

Gabriel: Almost a ten.

Lina: Though in its own systems it downgrades it to "important," because there it's harder to exploit.

Gabriel: On the bright side: AI helped find a serious flaw before attackers used it, and the fix came out quickly.

Lina: On the worrying side: attackers can use those same tools too. And until millions of servers are updated, the risk is still there.

Gabriel: Lina, what would you do with forty-eight AI assistants working for you?

Lina: Sleep.

Gabriel: Pietro Miotti, a programmer who makes a point of saying he isn't an archaeologist, put them to work looking for Maya ruins.

Lina: From his computer?

Gabriel: Yes. With public NASA lidar data: a laser that, from a plane, measures the ground through the trees. About one thousand two hundred square kilometers of Maya jungle.

Lina: And the agents reviewed all of that.

Gabriel: Image by image. He used Gemini, ChatGPT, Claude, and Grok. And they pointed to a hill with a leveled plaza, an oval pyramid, and a raised courtyard.

Lina: Where, exactly?

Gabriel: In Mexico. He sent a report to INAH, Mexico's archaeology institute, and they replied that no site was registered there.

Lina: And all of that in…

Gabriel: …three days. From the idea to the official reply.

Lina: Right. But something tells me a "but" is coming.

Gabriel: It is. On Wednesday he himself updated his post: archaeologists had already mapped the site, in a study from last year, by Estrada-Belli and others.

Lina: So he rediscovered it.

Gabriel: Exactly. The agents didn't find that study when they reviewed what had been published. And it still wasn't in the INAH registry.

Lina: You have to give him credit for saying so himself, without hiding it. And the site stays in the hands of INAH and of the archaeologists who mapped it first.

Gabriel: Some people see this as a preview of what's coming: a curious person, with AI agents, got in three days to where experts had gotten. AI is useful for science, not just for coding.

Lina: And some people see the limit: the agents failed at the basic part, checking whether someone had already done it. And the "discovery" went viral before it was verified.

Gabriel: That's it for today's episode. Tomorrow, five more stories.

Lina: And if I ever find a pyramid, I'll check first that nobody has already mapped it.

Gabriel: Talk to you tomorrow.