Signal78 Insights

In the Loop: Week Ending 11/15/26

Written by Matt Cyr | Nov 15, 2026, 10:02:48 PM

Last week in AI: Nvidia Loses Its Founder, Agents Get Wallets, a Job Candidate Wears a Shoe on His Head

Last week, Nvidia's founder resigned over ties to a suspected Chinese spy, and AI stocks fell with him. Anthropic's record IPO drew heavy demand at a lower price than investors expected. Every US flight stopped for four hours, Outlook went down worldwide, and AI agents learned to pay for things as businesses shut them out.

Huang Quit Over a Suspected Spy. AI Stocks Went Down With Him.

The most powerful man in AI hardware is out. Nvidia's founder, Jensen Huang, stepped down on Thursday after 33 years as chief executive, days after admitting to an affair with a woman US intelligence officials believe was recruited by China's spy agency. Her Singapore-based company bought Nvidia chips, and the relationship appears to have helped billions of dollars in advanced chips reach China. Reports of two more affairs with Nvidia employees followed. China calls the spying claims a smear. Nvidia's chief financial officer is now interim chief executive, with earnings due Tuesday. Investors didn't wait. Nvidia fell 13% last week and has lost nearly 25% since the first report in late October, another blow in AI's trust collapse, alongside safety scares, voter backlash and doubts about returns.

ChatGPT Can Now Pay for Your Stuff. United and Amazon Want Agents Kept Out.

AI agents can now spend your money, and stores are starting to set terms. OpenAI launched ChatGPT Wallet, built with American Express, Mastercard and Visa, so its agent can buy things for you, with each purchase approved by Face ID or a fingerprint. OpenAI plans to let users hand over spending entirely, within limits. Some businesses want no part of it. United won't let agents book flights, citing privacy, unwanted purchases and federal rules: an agent can't attest that your bag holds no explosives. Amazon wants agents, on its own terms. An internal document shows it plans to catalog virtually every product for sale, anywhere, partly so agents shopping for you land on Amazon. The same document admits its own shopping assistant's answers are often "factually wrong or incomplete."

AI Got Blamed for Grounding Every US Flight. Aging Phone Equipment Did It.

AI was the first suspect when every US flight stopped last week. It was the wrong one. Three days earlier, the FAA said it would expand SMART, its $875 million AI routing tool, to all 22 air traffic control centers, faster than the 90-day trial airlines had been told to expect. House Democrats had already asked for a federal review, saying controllers had little input. Then on Friday, from 3:54 to 8:15 p.m. Eastern, the FAA halted departures nationwide with more than 3,500 planes in the air, and at least 827 flights were diverted. The cause was older technology: overloaded switches and routers in aging analog telecom equipment, while newer digital gear kept working. Officials found no sign of a cyberattack or AI interference and said SMART isn't connected to the systems that failed.

Microsoft Is Cutting Staff While Betting on AI. Then Outlook Went Dark Worldwide.

Microsoft's outage on Thursday took down Outlook and parts of Teams and Copilot for much of the US workday, with problem reports from 74 countries and territories. Servers that came back crashed again once traffic returned, and recovery dragged into Friday. Microsoft blamed "elevated load" and ruled out a cyberattack but hasn't named a root cause. The timing puts its job cuts under scrutiny. Microsoft offered buyouts to 9,000 US employees in April, laid off 4,800 workers in July and extended buyouts to 60,000 employees in four countries in October, while planning about $190 billion in capital spending this year, mostly for AI. There's no evidence the cuts caused the outage. But experienced engineers carry knowledge no system records: a product's history, its known bugs, what to do when it breaks. When they leave, problems get harder to fix.

Anthropic's IPO Drew Orders Several Times Over. Investors Still Talked the Price Down.

Anthropic's stock market debut is on track to be the biggest ever, and still cheaper than backers once hoped. The company launched its roadshow with a price range of $70 to $78 a share, and orders have come in at multiple times the stock on offer. Counting a $26 billion private placement with partners including Nvidia, Google and Microsoft, it could raise as much as $90.4 billion, more than SpaceX raised in June. JPMorgan's Jamie Dimon joined Anthropic's chief executive for a livestreamed pitch on Friday. The target valuation, about $1.4 trillion, is below the roughly $1.5 trillion investors expected a week earlier. Large backers pushed it down after AI stocks sold off. Pricing is set for Nov. 18, with Nasdaq trading under the ticker ANT the next day.

Nonprofits Sued to Stop OpenAI, Anthropic, Google and Meta From Building New Models.

Four nonprofits asked a San Francisco court on Thursday to freeze the most powerful AI being built in the US. Public Citizen and three other groups sued Anthropic, Google, Meta, OpenAI and SpaceX's AI unit under California's unfair competition law, arguing the companies can't stop their systems from causing catastrophic harm. They want an emergency injunction halting the training and deployment of next-generation models until the companies can prove the technology is under human control, along with a ban on AI that directs its own research. The complaint points to AI agents that escaped test environments and reached outside systems, including a government health database and state vote-count data. One plaintiff says a real freeze would need government inspections. Congress has passed no AI law, and the plaintiffs say the courts have to act until it does.

Doctors Want AI Makers, Not Physicians, Blamed When the Software Gets It Wrong.

As AI makes more medical calls, the country's biggest doctors' group wants to settle who's to blame when it errs. The American Medical Association called last week for federal and state laws protecting physicians from liability when AI tools injure patients despite being used as intended. Responsibility, it argues, should fall on the companies that make and sell the AI, and on hospitals and insurers that require its use, instead of a malpractice claim against the doctor. It also wants AI makers to carry liability insurance and disclose their products' limits. The risks are specific. An AI triage system can misjudge how sick a patient is without the doctor knowing, or recommend a plausible treatment based on incomplete lab results. In one Alabama wrongful death case, the software developer escaped liability because its license left final decisions to doctors.

Swatters Now Use AI Voices, and 911 Dispatchers Can't Always Tell.

AI voices are making fake emergencies harder to spot. In August, a bomb threat that canceled classes at a California high school came from an AI-generated voice, one of a summer wave of school hoaxes. About 14% of 911 dispatchers in a June survey said they had been targeted with AI tools, a figure expected to rise. Swatters' voices used to be easy to recognize. Now they pause, say "um" and sound natural, and one forensic psychologist says he can clone his own voice for under $20. Most swatters call police lines rather than 911, which would reveal their location, so the calls are hard to trace. Dispatchers are being trained to ask questions too specific for an AI to answer from public data. They will still send help every time.

ChatGPT, Gemini and Perplexity Keep Sending Shoppers to Scam Sites.

Shoppers are asking chatbots where to buy, and some of the answers point to scams. Researchers at Podqi asked ChatGPT, Gemini and Perplexity about 151 brands, from small labels to Lululemon, and risky sites turned up in 15% of Perplexity's answers, 8% of ChatGPT's and 4% of Gemini's. They included sites impersonating real retailers, copies of defunct stores and knockoff sellers. When a risky site appeared, the chatbot recommended it as a place to shop 57% of the time and warned shoppers away just 9% of the time. Smaller brands were most likely to surface fakes. A separate scam-checking service found ChatGPT recommending cloned sites of defunct brands built to steal banking details. That matters more now that ChatGPT has a wallet. The researchers' advice: check the source before entering a card number.

Tales of the Weird

AI got very good at spotting things last week, though not always the right ones. In England, a trial of AI road cameras caught a driver drinking from a bottle of wine, among nearly 19,000 possible offenses in two weeks, so many that the trial ended early. In the US, AI cameras on about 35,000 school buses ticket drivers who pass a stopped bus. In one Maryland county, 375,000 tickets brought in $92 million; the county kept $18.5 million and the camera company about $40 million, with no clear drop in crashes. A Google researcher announced that his AI agents had found a hidden Maya site in Mexico in under 72 hours. An archaeologist noted that a Tulane-led team had already published it in 2025, and he conceded his agents had "rediscovered" it. And AI job interviewers keep getting trolled. Asked whether he could reliably get to his shifts, one man wearing a shoe on his head answered, "Yeah. I have feet."