Anthropic Soars, Then Wobbles, in Historic Debut

Also: House Records Show Commerce Cleared 880,000 AI Chips for China

D.A.D. joke of the day: I won't let AI do my taxes. It makes too many deductions it can't back up.

What's New

AI developments from the last 24 hours

Anthropic's Market Debut: A Success With An Asterisk

AI Business & Money Anthropic

By the usual yardstick, Anthropic's first day as a public company was a smashing success. Its shares closed at $97.17, up 17% from the $83 offering price, valuing the company at about $1.76 trillion on a fully diluted basis, CNN reported. That dwarfs the debuts of Facebook and Uber and makes Anthropic San Francisco's most valuable company, the San Francisco Chronicle noted. At the close, CNBC said, only six companies were worth more: Apple, Microsoft, Nvidia, Alphabet, Amazon and SpaceX. Anthropic raised $81.3 billion, more than it first planned, after increasing both the number of shares and the price.

The rest of the day told a more cautious story. The stock opened 34% higher and briefly pushed the company's value past $2 trillion, then faded as the broader market fell, leaving anyone who bought at the opening bell down 13%. Nor did the debut lift other AI stocks, which fell sharply, CNBC reported. Hedge fund manager Dan Niles sees two warning signs. One is the business: Anthropic and OpenAI are locked in a price war, and their combined revenue growth slowed from 75% in the second quarter to 48% in the third. The other is the trading: a strong debut usually closes above its opening price, a sign that eager buyers outnumber sellers, and this one didn't. The real test, he says, is whether the stock stays above its $83 offering price for the next 30 days. He has his doubts. NYU finance professor Aswath Damodaran, a widely followed authority on valuation, put Anthropic's worth at about $76 a share, below even the offering price, once employee stock options are counted.

The bulls see plenty of room to run. Analysts at William Blair called the price reasonable at about nine times Anthropic's expected 2027 sales, and Melius set a two-year target of $155 a share. Their case rests partly on scarcity: Anthropic is the first company built purely around frontier AI models that ordinary investors can buy. History offers a caution, though. Analyst Ryan Detrick found that most major listings rise on their first day, but the median one later falls by more than half from its first-year peak. The next big test arrives when OpenAI, Anthropic's chief rival, goes public, which is expected next year.

On the day of the debut, CEO Dario Amodei made a pledge that speaks directly to his critics: "And to state it unequivocally, 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. We will not bend to either business or political pressure." He also laid out the upside he is selling: "I believe sufficiently powerful AI, working in concert with humans, could compress a century of scientific and medical progress into a decade."

Why it matters: Until now, investors mostly judged the AI boom through chipmakers like Nvidia and a handful of tech giants. Anthropic's stock is the first daily public verdict on what a frontier AI lab itself is worth, and on whether AI's promise justifies its enormous cost. Amodei's pledge to slow down for safety will be tested against quarterly expectations.


TSMC Cuts Off Chinese Chip Designers From Its Most Advanced Factories

China & Geopolitics Chips, Data Centers & Energy Regulation & Policy

TSMC, the world's largest contract chipmaker, is telling Chinese customers to stop ordering chips made on its 16-nanometer process or anything more advanced unless they obtain a U.S. export license, Reuters reports, citing people familiar with the matter. Most will be limited to older 28- and 40-nanometer technology. It is the latest step in a crackdown that has gathered speed since the scandal that ended Jensen Huang's tenure at Nvidia, over his ties to an executive at Megaspeed, a Singapore data-center company, put the flow of American chips to China under a spotlight. The move follows a Commerce Department rule unveiled Tuesday that ends a workaround: Chinese designers could have TSMC, or independent packaging and testing firms, certify that their chips fell outside U.S. controls. Now such chips are presumed to be controlled, including orders routed through subsidiaries or intermediaries abroad. The Financial Times, which first reported the halt, says Samsung has stopped projects for Baidu's AI chip unit Kunlun and TikTok owner ByteDance, both of which have been trying to diversify away from scarce Nvidia processors, and that both chipmakers have suspended shipments to consumer electronics customers, putting production of Xiaomi's chips at TSMC on hold.

Why it matters: China's push to build alternatives to Nvidia depends on getting those chips manufactured, and TSMC was one of the last routes to cutting-edge production. Together with Monday's crackdown on renting chips abroad, Washington is closing the side doors one at a time, and the pressure to go further is building in Congress, as the next item shows.


House Chairman Reveals Commerce Cleared Nearly 880,000 AI Chips for China

China & Geopolitics Regulation & Policy Nvidia

As Washington tightens the rules on chipmakers, Congress is also turning up the heat on the agency that enforces them, the Commerce Department, and the push is coming from the president's own party. Rep. Brian Mast, the Republican chairman of the House Foreign Affairs Committee, used his powers under the Export Control Reform Act to publish details of Nvidia's and AMD's export license applications for China, a disclosure his committee calls the most detailed in the law's history. The documents show requests to sell 3.4 million AI chips to China, of which the Commerce Department approved more than 879,000, including sales to Alibaba and Tencent, which the Pentagon lists as Chinese military companies. The committee also says some of Nvidia's applications relied on possible chip-smuggling fronts. Chris McGuire, who oversaw chip export controls at the National Security Council, pointed out that every application from Nvidia and AMD rated the risk of the chips being diverted as low, and that more than 30% of the approved chips went to companies on the Pentagon's list. "There is simply no way that the Department of Commerce did a serious national security review of these licenses," he wrote. Separately, The Information reported this week that Commerce had approved some licenses for Nvidia's H200 chips over the objections of the State, Energy and War departments. Commerce has not responded publicly.

Why it matters: This week's crackdowns target the side doors: rented chips abroad and chips made to order at TSMC. These records suggest the bigger gap may have been the front door: Commerce's own approvals, granted largely on the companies' assurances.


Claude Helps Solve Long-Standing Black Hole Physics Problems, Anthropic Says

Science & Discovery Anthropic

Anthropic says Claude cracked two unsolved problems in theoretical physics, including a calculation first attempted in 2003. Working like a small research team of self-copies, the model finished the math showing that a sharp transition tied to black hole formation—first demonstrated by Hawking and Page in 1983 using general relativity—also holds in a quantum regime where spacetime has no fixed shape. That suggests black hole formation may be baked into the deeper quantum theory, not just a quirk of relativity.

Why it matters: It's an early, concrete example of an AI model producing genuinely new results in a field defined by decades-old unsolved math, not just summarizing or checking existing work.


What's Controversial

Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community

Wave of Cyberattacks, Many AI-Powered, Hits the European Commission, JPMorgan Chase and More

Cybersecurity Europe & UK Finance & Banking

A string of new breaches adds to a wave of attacks, many of them AI-assisted, that has unnerved security officials around the world. The European Commission confirmed that attackers used artificial intelligence tools in a cyberattack it detected on October 22, forcing it to take several staff systems offline, including those used to check payslips, log working hours and request leave. POLITICO, which first reported the breach, linked it to ARTEX, the free AI security-testing tool that the FBI recently warned criminals and state-aligned hackers have weaponized (D.A.D., November 18). A breach of the system that manages staff phones may have exposed employees' names and numbers, but spokesperson Thomas Regnier said there was no evidence of a breach affecting staff as a whole, and that he was not aware of any ransom demand. Officials say most of the attack has been contained.

The FBI has tied ARTEX to October's breach of ConnectWise's cloud backup service and to the theft of data on about 67,000 people from South Korean financial institutions. The speed is what alarms defenders. David Kennedy, a former NSA hacker, told CNN about a recent case in which more than 30 AI agents worked in parallel to break into a company: they used stolen passwords, found a weakness in its two-factor login protection and finished the job in 14 minutes, while the company didn't notice for about four and a half hours. He did not name the company.

Separately, JPMorgan Chase said outside vendors that provide services to the bank had been hit. "Based on our investigation so far, we believe the attackers accessed certain marketing data held by those vendors, which may include some customer names and email addresses," the bank said, adding that its services weren't disrupted and that the attack had no connection to outages some customers saw earlier this week. Newsweek reported that a hacking group had published about 16,000 email addresses of customers tied to the bank's Reserve card program for wealthy clients; Chase hasn't said whether that leak is part of the same incident, and no AI link has been reported.

Why it matters: AI is shrinking the time between break-in and damage from days to minutes, faster than most organizations can detect an intruder. The Commission attack shows that even major government bodies are targets, and the Chase breach is a reminder that a company's security is only as strong as its vendors'.


What's in Academe

New papers on AI and its effects from researchers

Study Finds Most AI Companion Chats Show Overreliance, Few Cause Harm

Kids, Teens & Mental Health

Researchers built a framework to separate emotional overreliance on companion apps like Character.AI and Replika from actual relational harm, then tested it on 110 adults' real conversations. Overreliance behaviors—like treating the AI as a primary emotional support—showed up in 63% of conversations, but measurable harm appeared in just 10.5% of cases. When overreliance patterns recurred, though, they tended to coincide with harmful outcomes. Open-source AI models could flag most of these warning behaviors automatically.

Why it matters: As companion apps scale, this gives platforms and regulators a way to distinguish normal attachment from dangerous dependency instead of treating all heavy usage as a red flag.


AI Chatbots Show Fewer Local News Sources in Latin America, Study Finds

A study analyzing nearly 50,000 search queries across nine Latin American countries found that AI chatbots with web search surface fewer local news sources than traditional search engines—and the gap isn't small. Where Google and similar engines already return local news only 41% of the time for in-country users, language models did 3.6 to 11.6 percentage points worse, while also pulling from a narrower pool of domains overall. The pattern favored sources from wealthier countries over regional outlets.

Why it matters: As AI search tools replace traditional search for news discovery, this suggests they may be quietly narrowing information diversity and underrepresenting local voices in ways users don't notice.


Structured AI Chatbot Improves Accuracy in Contact Tracing Interviews

A research team built Epira, an AI interview system for public health contact tracing, pairing a chatbot with a rule-based controller that forces it to ask every required question and flag safety issues. In testing on measles and gonorrhea cases, Epira captured 88.9% of required fields—open-ended chatbots using the same underlying models missed 31 to 42 percentage points more. Crucially, performance held steady whether the system ran on a local model (Qwen2.5) or a cloud model (GPT-5.4), and 72 volunteers rated it highly usable and moderately trustworthy.

Why it matters: It suggests that for high-stakes, compliance-heavy tasks, how you constrain an AI matters more than which AI you pick—a lesson that extends well beyond public health to any workflow requiring consistent, auditable interviews or intake.


Medical AI Chatbots Need Better Reasoning Checks, Not Just Right Answers, Researchers Say

Health & Medicine

A review by researchers at Saudi Arabia's University of Prince Mugrin and Prince Sultan University, submitted for publication but not yet peer-reviewed, argues that medical AI chatbots are being graded the wrong way: regulators and benchmarks focus on whether the final answer is correct, not whether the reasoning that produced it was sound. The authors warn this misses real failure modes—hallucinated facts, flawed cause-and-effect logic, and reliance on demographic shortcuts that happen to produce right answers for wrong reasons. They point to hybrid systems pairing AI with rule-based medical logic checkers as a promising but thinly tested fix, with most positive results coming from single, small-scale studies.

Why it matters: As hospitals adopt AI diagnostic tools, a model that's right 95% of the time for the wrong reasons could fail unpredictably on the 5%—which is exactly the case a doctor needs it to catch.


AI Is Finding Far More Software Holes, but Fixes Aren't Coming Faster, Study Finds

Cybersecurity AI Agents

A new academic framework borrows driving-automation terminology—levels of autonomy—to rate how independently AI agents can find, confirm, and fix software vulnerabilities. Applying it to a decade of security data, researchers found AI agents reached full autonomy in vulnerability discovery and patching faster than expected, and disclosed vulnerabilities more than doubled in 2026, with a bigger share rated critical. But the time to actually fix bugs, or for attackers to exploit them, hasn't budged.

Why it matters: AI is getting dramatically better at finding security holes in software, but that flood of discoveries isn't translating into faster patches—meaning the gap between knowing about a vulnerability and closing it may be widening, not shrinking.


Get tomorrow's briefing

Share this edition Subscribe