Microsoft’s CEO says the ad-funded web has an agent problem, and Anthropic promises to slow down on the day it sells shares. Meanwhile, an FAA upgrade managed to take down the very network it was upgrading.
1. Nadella says the ad-funded web has an agent problem
Nadella’s argument is that most of the web’s business model assumes a human is looking at the page. At the NYT DealBook Summit, he said aggregators built for the “human web” will get “reshuffled.” If you were monetizing eyeballs, he said, you may have to settle for cost-per-click or something more transactional. His summary: “if you don’t accept the agent, somebody else will.“
His example was Amazon, which blocks AI agents from browsing its catalog and buying on users’ behalf. Nadella suggested a platform like Shopify could let agents in and win the business. That scenario is his hypothetical, and he runs a company that sells agents, so he benefits from a web that welcomes them. TechCrunch is the only account here.
The catch is that Amazon’s choice isn’t just stubbornness. An agent that skips the page also skips the ad, and ads are what Amazon sells. Blocking agents may be a bargaining position as much as a defense.
Nadella applies the same logic inside Microsoft. He offered buyouts to 60,000 employees in October, saying every job will become managing agents. His example was an 11-person team that builds a “software factory” rather than the product itself.
TechCrunch · TechCrunch · TechCrunch
2. Uncensored models spread faster than their warning labels
Researchers at UIUC and Nanyang Technological University went through 22,096 candidate models across five hosting platforms and identified more than 11,000 uncensored LLMs, with Hugging Face the dominant source. Only about 13 percent carried moderation warnings. The authors’ point is that a model can be copied from platform to platform while the information about its risks gets lost along the way.
The growth is steep: roughly 100 a month in May 2023, over 1,000 a month now. One uncensored model has more than 53 million downloads. Twenty-seven had security-scan flags, such as malware signatures or unsafe deserialization, and those were downloaded 136,000 times.
One limit matters: the method finds models that disclose being uncensored in their metadata. The honest ones are countable, and the rest are not. The Register also notes legitimate uses, including security analysis that commercial models refuse. Hugging Face, which is being acquired by Nvidia, did not comment. The researchers’ own proposed fix is more documentation and auditability, which helps only where uploaders cooperate.
3. An FAA upgrade took down the network it was upgrading
The FAA’s 4.5-hour nationwide ground stop on November 13 appears to have started with a router. Per WIRED’s reporting, an L3Harris contractor was moving a site in El Dorado, Arkansas, from copper lines to fiber, part of a project about 70 percent complete. The new router then began making 1,000 connection attempts every four to five seconds, and the flood overloaded other nodes and knocked out radio frequencies.
Six days later, Transportation Secretary Sean Duffy said the root cause was still unclear. Pilots and aviation experts say flights stayed safe, but one controller called it “the hairiest 90 minutes of my career.“
Some selection notes tie this to an AI rollout. The reporting here doesn’t mention AI at all. It is a story about the ordinary risk of cutting over a live legacy system. Duffy’s response is to ask Congress for $50 billion, $30 billion of it for air traffic control modernization. That is the same kind of work that preceded Friday’s outage, which doesn’t make it wrong, but it leaves open how this particular cutover failed.
4. Accused spy was a paid amateur with a laptop and Telegram
According to an FBI affidavit reported by The Register, Igor D’iu allegedly crossed into the US illegally and took orders over Telegram from people he only suspected worked for Russia. His targets were employees of a Palo Alto defense-tech company that is unnamed in court documents and widely reported to be Palantir, though that isn’t confirmed. The allegations are not proven.
The tradecraft was low-tech. He allegedly photographed employees’ laptop screens in coffee shops, left a remotely controlled laptop outside an apartment complex, and was handed addresses to check while dressed as a delivery worker. Agents say they seized profile packages on at least three employees. He was allegedly paid $15,000 three times.
One episode is hard to explain away: he allegedly followed a woman for nearly 30 miles until she drove to a police station. He says he wanted to ask her out. The affidavit says he came to suspect he was doing something illegal only in October. The lesson for companies is that people, not systems, were the target. He faces two spying-related counts and one fake-ID count, and is due in court November 24.
5. Anthropic’s Nasdaq pledge to go slow has no measurable trigger
On the day Anthropic went public in what TechCrunch calls the largest IPO in history, Dario Amodei told the Nasdaq crowd: “we will pace the training and deployment of AI systems, taking as long as is necessary to have confidence that they are safe and will remain under human control.” He added that the company would “not bend to either business nor political pressure.“
It is a striking thing to say while taking public money. It is also hard to test. The speech names no threshold, no delay that would count as proof, and no outside party with authority to say the company has fallen short. In September Amodei proposed third-party evaluators embedded to verify safety commitments. Altman, Hassabis and Musk said they agreed, though TechCrunch notes few concrete commitments have followed.
A pledge like this is credible only if it eventually costs something, such as a release held back or a model shelved. Until then, shareholders and critics are left reading a speech. The first real data point will be a visible delay.