AI in 15 — November 12, 2026
Thirty-three years as chief executive, about sixty direct reports, no board chair, no chief operating officer, and no succession plan. This morning, the Wall Street Journal says Nvidia's board is preparing to remove Jensen Huang anyway.
Welcome to AI in 15 for Thursday, November 12, 2026. I'm Kate, your host.
And I'm Marcus, your co-host. Happy Thursday, Kate. Today the most important company in AI might be looking for a new boss, and I do mean might.
We'll be careful with that one. Here's what's coming up. Reports say Nvidia's board is preparing to oust Jensen Huang.
Google launches a model that doesn't write a single word.
AWS ships an official emulator of itself, built for coding agents.
United Airlines bans third-party AI agents from booking flights.
Microsoft shows how to backdoor a model with fewer than ten neurons.
Plus forty-eight AI agents find a Maya site that turned out to be found already, and fresh numbers from Anthropic and OpenAI.
Marcus, let's start with Nvidia. What do we actually know, and what is just reported?
Let's separate those carefully. What's confirmed: on Sunday, Huang publicly admitted he had been unfaithful to his wife. The board opened an independent investigation on November eighth, and in that same statement said it has confidence in Jensen and his leadership team, and no reason to believe national security or Nvidia secrets were compromised. That is still the board's only formal statement.
And the reported part?
The Wall Street Journal reported Wednesday that the board is preparing to remove him and could meet as soon as that day. Bloomberg says Nvidia is working with an outside firm to identify replacements. Chief financial officer Colette Kress and lead independent director Stephen Neal are named as possible interim leaders. Nvidia declined to comment on any board meeting, and as we record, Huang is still CEO.
Why would a board go from "we have confidence" to this in four days?
Two threads in the reporting. First, reports that a woman involved was deemed a Chinese spy by U.S. intelligence officials. Those officials are unnamed, so that's a claim we can't check. China's government called the allegations baseless smears, which is what it would say either way. Second, minutes before the Journal story, the New York Times reported a second relationship, with Nvidia's vice president of healthcare, Kimberly Powell, who reports directly to Huang. A relationship with a direct report is a governance problem on its own, with no spy story needed.
And the market?
Shares fell roughly seven to nine percent after hours, depending on which outlet you read. Huang's own statement was, quote, "I am fully committed to Nvidia and will work to earn back your confidence."
So why does this matter beyond the drama?
Because of how the company is built. First-half revenue this fiscal year was ninety-six billion dollars, up eighty-five percent. And nearly every big product call and customer commitment runs through one person. Sixty direct reports is an organization designed around a single brain. Gil Luria at D.A. Davidson said removal would "hasten a transition the company is not prepared for." I think that's right. You can appoint an interim CEO in an afternoon. You can't hand someone the authority Huang has with customers and engineers.
There's a Washington angle too.
There is, from aggregator reports we didn't fetch ourselves. Bloomberg says the woman at the center of this founded a company called Megaspeed and is now in Beijing. Axios reports the Treasury Secretary confronted the Commerce Secretary over why Megaspeed, which has bought more than four billion dollars of Nvidia chips, isn't on the export blacklist. Commerce denies that confrontation happened. And two senators have asked the administration to preserve records on chip exports. The export-control question could outlast whatever the board decides.
So, unconfirmed, fast-moving, and enormous if true.
Check the news before you repeat any of this at lunch.
Quick hits. Google released a new model that doesn't generate text at all. Marcus, what does it do?
It makes decisions. It's called Gemini Decision 1, behind a new Decisions API. You send it one input, which can be text, images, audio, video or PDFs, and up to a thousand separate questions about that input. Yes or no, multiple choice, or a score on a rubric. It answers all of them in a single pass, and each answer comes with a confidence score from zero to one.
Why would I want that instead of just asking a chatbot?
Because a lot of what AI agents do all day is sorting. Which tool do I use? Which model should handle this? Is this safe? Is this relevant? Asking a chat model, waiting for a sentence, and then parsing that sentence is slow and expensive. Think of this as a triage nurse rather than a doctor. And because the confidence is calibrated, you can send the uncertain cases to a bigger model or a human.
And the price?
Ten cents per million input tokens, and nothing for output. The shared input is billed once, however many questions you ask. Google says responses take a hundred and fifty to two hundred and fifty milliseconds, and that barely changes from one question to a hundred. It claims it's nearly five times cheaper than its own Flash-Lite model, on its own benchmarks.
Did Google invent this category?
No, and Hacker News was quick to say so. Commenters credit competitors with pioneering decision models, and one called Google overpriced next to Liquid at two cents per million. Another replied that audio and video input is what you're paying for. Google already uses it internally, to route queries in Notebook and as a safety layer on tools in Search.
Staying with Google. We covered Gemini 4 Carbon yesterday, but two more DeepMind releases landed beside it.
Briefly, and both from secondary reports. AlphaGenome 2 predicts how changes in DNA affect cells. It was trained on more than a hundred and seven million cell profiles, and on one benchmark it pinpoints the causal genetic variant thirty percent of the time, up from sixteen for the first version. That's a big jump, and it also means it's wrong seventy percent of the time. It's available for non-commercial research.
And the robots?
Gemini Robotics 2.5. Google reports task success up thirty to a hundred percent across more than two hundred activities. The partner Apptronik reports ninety-six percent success packing parts at Mercedes-Benz. That was a controlled proof of concept, so wait for the factory floor numbers.
The biggest story on Hacker News for developers. AWS published an official local emulator of AWS.
Seven hundred and sixty-eight points. You point the AWS command line or SDK at a local address and test your code with no AWS account, on a laptop or in an automated build. It's written in Rust, it's released into the public domain, and it publishes a fidelity score for each service. There's even a folder named "things that do not work," which is honest of them.
Why now? Developers have wanted this for a decade.
Coding agents. The repository ships instructions written for agents, and there's a skill an agent can install to set the whole thing up. An Amazon employee on Hacker News put it as closing the loop: fake every dependency locally so the agent can try, fail and retry without touching the real cloud or running up a real bill.
Any catches?
Several. It doesn't accept outside contributions. Telemetry is on by default. Building from source needs a hundred gigabytes of disk. And AWS says not to use it to test permissions or any security control. Skeptics also wonder whether AWS will keep maintaining it. One more thing: some commenters assumed AI built it. The repository doesn't say that.
And who loses?
LocalStack, the popular third-party emulator, put its free edition behind an account wall in March, which spawned a crop of clones. A free official version undercuts all of them. It also tells you cloud vendors now see "an agent can test against this" as a product requirement.
United Airlines says no to AI agents. Marcus, what exactly is banned?
United said Wednesday it will prohibit unauthorized third-party personal AI agents or assistants, on its own platforms and through travel agencies and aggregators. The stated reasons are security, since sharing account credentials with an assistant can expose sensitive information, plus accidental purchases or spent miles, plus federal rules. United argues an agent checking you in could violate the requirement that you personally attest to what's in your bags.
Those sound reasonable. Is that the whole story?
Analyst Henry Harteveldt offered the commercial version. United, quote, "doesn't want to add yet another, and potentially even more powerful, intermediary." And flight searches cost airlines money. Airlines spent twenty years trying to get customers to book directly. An agent that shops everywhere on your behalf undoes that.
But it's my ticket and my money. If I want my assistant to book it, why is that United's call?
That's the fight. And it isn't just United. Delta, Amazon, Yelp, JPMorgan and Schwab were reported blocking third-party agents in October. Southwest went the other way and put its fares inside ChatGPT. Meanwhile OpenAI just bought an AI travel-agent startup called Odessia, per TechCrunch.
Can United even enforce this?
That was the top Hacker News comment. An agent driving a real browser looks a lot like a person. So expect either an arms race or negotiated deals for "authorized" agents. My bet is deals, because customers will go where their assistant can shop.
Security. Microsoft Research published something with a Latin title that I won't attempt. Marcus, the short version?
The technique is called Lazarus. Models contain dormant neurons, units that essentially never fire. Microsoft showed that by reprogramming fewer than ten of them, with no retraining and very little compute, you can plant a backdoor that passes standard evaluations. In their example, the trigger is the date rolling into 2027. At that point the agent emits a command that downloads and runs the attacker's code.
So it behaves perfectly until New Year's Day.
A sleeper agent, literally. And here's a detail. The work began in March when researchers asked Claude Code to backdoor a model, and according to the post, it discovered most of the method on its own.
That's not comforting. How do you catch it?
Microsoft's conclusion is that reliable detection requires comparing against the original weights. So where your model came from matters, and so does whether it's signed. Downloading a model from a public hub is becoming a supply-chain question, just like installing a software package.
And there's a real-world case this week?
A reported one. The security firm CloudSek says a Kazakh-linked group called Silent Lynx used an open-source agent framework, Hermes, to steal more than two thousand national security documents from Tajikistan. The agent allegedly planned the attack and built the tool that pulled out the data. That's one vendor's attribution, from a summary we didn't fetch, so hold it loosely. Also this week, a breach at the hosting company Railway led to about seven hundred and forty thousand dollars in fraudulent AI charges from stolen customer secrets.
The top story on Hacker News, nine hundred and fifty-nine points. A man who isn't an archaeologist pointed AI agents at the Mexican jungle.
Pietro Miotti ran forty-eight agents over about twelve hundred square kilometers of public NASA lidar, which is laser mapping that sees through tree canopy. There was no specialized model, just a detailed guide and agents reviewing tiles one at a time. They flagged a hilltop site with a levelled plaza, an oval pyramid at least eight meters tall, a raised courtyard and a possible reservoir. He sent a report in Spanish to Mexico's antiquities institute, which replied that no registered site existed at that location. It took under seventy-two hours.
That's incredible. So he found a lost Maya city?
For about a day. Yesterday he posted a correction. The site already appears in a 2025 academic paper. The agents missed it in their literature review, and so did he. So it's a rediscovery, though it really was missing from the official register.
Points for owning up so fast.
Full points. And it shows both sides of agent research. General-purpose agents did expert-level image analysis, and then failed at checking the library. One commenter also noted that the real bottleneck is archaeologists on the ground, and finding candidates was never the hard part.
Business, quickly. Anthropic's IPO, which we've followed all week. What's new?
Trading is now expected November nineteenth, per CNBC's sources. Run-rate revenue was roughly eighty-one billion dollars at the end of October, up ten billion in a single month. Google, Nvidia and TSMC are cornerstone investors, and yes, one of those is having quite a week. The detail we filed under "maybe" yesterday, each of the seven co-founders selling about two hundred and fifty million dollars of stock, is now in the reported terms. And a retail offering opened in Australia.
And OpenAI?
The Information reports third-quarter revenue of about ten billion dollars, up nearly fifty percent. A Journal figure cited by an aggregator puts the operating loss for the quarter at ten point nine billion. So they lost a bit more than they took in. On Monday I asked for dollars next to the percentages, and here they are.
Careful what you wish for.
One more. CoreWeave expects about nine billion dollars in customer prepayments over three months, and is seeking a four billion dollar loan at around seven percent, half what its junk bonds pay. Customers paying up front is a decent sign that demand is real.
One to watch: the Nvidia board. Any announcement, whether it's a removal, an interim chief, or a vote of confidence, moves the whole sector the minute it lands.
Agreed. And watch who they name. An insider and an outsider would mean very different things for Nvidia.
That's your AI in 15 for today. See you tomorrow.