D.A.D. joke of the day: I asked ChatGPT to be more concise. It apologized for four paragraphs.
What's New
AI developments from the last 24 hours
Trump Says He'd Work With Democrats on AI Guardrails, and That the DOJ Keeps AI in Line
President Trump says he would "absolutely" work with Democrats on AI safety rules, with one condition. "I don't mind guardrails. But I don't wanna put us in a deep second position," he told Ross Douthat in a 60 Minutes interview that aired Sunday. "We're leading number one. And China has no guardrails, by the way."
He called himself an AI bull because of the race with China. "Whoever wins that race wins," he said, adding that China "is now one year behind us." The risk, he said, "is not nearly as real as" AI's builders say. "I think the risk can be managed. And it's managed, you know who? By who? The DOJ."
Then he told a story. "We had a situation where somebody was reported to us to be doing something bad with AI. We called them very strongly and said, 'Get ready to go to jail. Please give your wife a kiss good night.'" Asked whether it was the leader of a major AI company, he said: "One of the big ones." The person "totally changed everything," Trump said. He wouldn't name the person or company or say what they had done, and said he didn't think it was done maliciously, though it could have been.
He has made this kind of claim before. In September, after Anthropic's Dario Amodei called for slowing AI development, Trump posted that his administration "has stopped AI 'people' from doing bad, or potentially bad, 'things,' like Dario (Anthropic!)," adding: "We already have tremendous CRIMINAL and REGULATORY power over these companies!" (D.A.D., September 15). Whether Sunday's story describes the same episode is unknown. His administration has leaned on liability rather than new rules: Treasury Secretary Scott Bessent told House lawmakers that AI labs should get no exemption from liability.
Democrats, who take control of Congress in January, have vowed AI guardrails and hearings. Trump didn't say which rules he would accept.
Why it matters: A president open to AI rules is news for a Democratic Congress drafting them. But his idea of oversight is different from theirs. A law sets written rules that apply to every company. Trump describes something else: a private warning from prosecutors to one executive, over conduct he won't describe, under a law he didn't name. That leaves companies guessing where the lines are, and it is the kind of thing a Democratic House with subpoena power is likely to ask about. The test of any deal will be his China condition: rules he can sell as keeping America ahead, not ones critics can say slow it down.
Sources: CBS News (60 Minutes transcript) · Reuters · CNBC (Trump's September post) · FedScoop (Bessent) · The National Desk (Democrats)
Goldman Takes AI Coding Agents Away From Traders and Bankers
Finance & Banking AI Agents Anthropic
Goldman Sachs has pulled AI coding agents from most of its traders and bankers. Internal guidance updated in October limits Anthropic's Claude Code and Microsoft's GitHub coding tools to the bank's roughly 12,000 engineers, out of about 47,000 staff, after a review of productivity and risk, Bloomberg News reported. Traders had been using them to monitor client trading flows and model market moves, and some are frustrated, Bloomberg reported. "Strats," who work across trading, quantitative research and engineering, keep access, and Goldman's in-house AI assistant remains available firm-wide. A spokesperson said the bank makes certain AI tools available to staff "to determine the best use cases."
It's a reversal: as recently as July, Goldman's chief information officer, Marco Argenti, was praising traders and bankers for building their own tools with AI and handing them to engineers to improve, according to Bloomberg.
Why it matters: Coding agents write and run software with the same access as the person using them. On a trading floor, a script a trader builds this way can skip the code review and testing an engineer's work goes through, and could cross the information barriers banks must keep between deal teams and traders. Goldman's answer is the one many companies will face: let anyone experiment, or make engineers own whatever the agents build.
Discuss on Hacker News · Sources: Bloomberg · The State of AI
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Nvidia Investigates Jensen Huang After He Admits Relationship With Alleged Chinese Spy
China & Geopolitics Military & National Security Nvidia
Nvidia's board has opened an investigation into its own chief executive. Jensen Huang has acknowledged an intimate relationship with Alice Huang, a business executive U.S. intelligence determined to be a Chinese spy. "I was unfaithful to my wife. My actions hurt her and our children," he wrote to employees in a message published by Axios. "I take full responsibility for what I did and the hurt I caused, and I accept the criticism that comes with it." On the security questions, he wrote: "I support the investigation and will cooperate fully. I love our country and our company."
The board said a committee of independent directors, assisted by outside counsel, will investigate reports that Huang "has been the target of a foreign intelligence effort." "While the Board has no reason to believe that either national security or NVIDIA secrets were compromised, the Board's inquiry will examine all available information," it said, adding that it has confidence in Huang, who will stay on as CEO. "Today's news will not slow Nvidia," said company spokesman John Rizzo. "Jensen is committed to continuing to lead Nvidia as C.E.O. without interruption."
Until now, Nvidia had described their contacts as ordinary customer meetings. Alice Huang, a former Chinese TV reporter, is a top executive at Megaspeed, a Singapore data-center company that bought $4.8 billion in Nvidia chips over the past two years and that the Commerce Department has investigated for helping China smuggle restricted chips. Nvidia kept doing business with it. Separately, several Chinese firms the Commerce Department cleared to handle Nvidia's H200 chips appear to have empty offices or ties to the Chinese military, The New York Times found. U.S. officials concluded she had been recruited by China's Ministry of State Security (D.A.D., November 7). China denies it. (The two Huangs are not related.)
The warning reached the top of the government months ago. Secretary of State Marco Rubio oversaw a highly classified memo last spring telling senior officials that Huang was in a close relationship with a Chinese spy, The New York Times reported. It circulated before President Trump's trip to China in May; Huang went along anyway, at Trump's invitation.
Nvidia's shares fell in after-hours trading as the news sank in, Reuters reported, with the company due to report quarterly results on Nov. 17. Analysts are split on whether Huang can stay. Congress is waiting. The Republican chairs of the House China, foreign affairs and intelligence committees asked Huang to a closed-door briefing on Nov. 18 and set a Nov. 11 deadline to accept or propose another date, their letter says. Nvidia hasn't said whether he will go.
Some argue the affair matters little, since Huang needed no persuading to push for looser export controls. Bill Bishop, who writes the Sinocism newsletter on China, disagrees. Anything Huang shared about his conversations with Trump would be valuable to Beijing, Bishop wrote, as would any access to his phone or laptop. "He is the chip allocator in chief, so a good relationship with him likely meant more chips for the MegaSpeed front and the entities back in the PRC to which those chips flowed," Bishop wrote. "There is rarely just one lie, so don't be surprised if the blast radius from this mess expands."
Why it matters: Huang runs the company at the center of the AI boom and has been the loudest voice for selling more chips to China. His admission moves the story from intelligence leaks to a governance crisis: a board investigating its founder, a company that kept selling to her employer, and an administration that was warned and kept listening to him. He has not been accused of a crime, and it remains unknown what, if anything, Alice Huang learned. Expect the fallout to land on Nvidia's stock, its board and the chip-export bills now before Congress.
Sources: Nvidia board statement · Axios (Huang's letter) · The New York Times (investigation) · The New York Times (Rubio memo) · Bloomberg · Reuters · Bill Bishop on X · House Foreign Affairs Committee
What's in the Lab
New announcements from major AI labs
Why Fitting Data Isn't Enough for AI to Do Real Science
A piece on AI and scientific discovery revisits the 1951-52 Hodgkin-Huxley model of how nerve cells fire electrical signals, built through painstaking manual calculation. The model required raising a variable to the fourth power to fit the data—a quirk only explained decades later when scientists found the related potassium channel protein has four identical parts around a central pore. The lesson: AI that merely fits equations to data isn't doing science. Real understanding means knowing which model features are necessary, not just convenient, and whether they generalize to other systems.
Why it matters: As labs push AI toward automating scientific discovery, this is a reminder that curve-fitting and genuine explanatory insight aren't the same thing—and conflating them could produce models that predict well but mislead researchers about what's actually true.
Claude Helps Scientists Solve Mysterious Viral Particle Structure
Anthropic's in-house biology lab used Claude to help solve the first 3D structure of a vault particle from a bacteriophage—a virus that infects bacteria. Vaults are large, mysterious cellular structures (about three times the mass of a ribosome) found in humans and other organisms, but their function remains unclear. Claude reportedly assisted with literature searches, isolation methods, and structural mapping. The finding suggests vault-like particles are far more widespread across life, including in bacteria and viruses, than previously confirmed.
Why it matters: It's a concrete example of an AI chatbot functioning as a research collaborator on open biological questions, not just a text or code assistant—a use case labs are increasingly testing as scientific tools.
Anthropic Widens Free AI Research Program to Major Science Institutions
Anthropic is expanding its free Claude-credits program for scientists, AI for Science, beyond individual researchers to partner with 20 foundations and institutions, including the Simons Foundation, Schmidt Sciences, the National Science Foundation, and the American Chemical Society's CAS division. Anthropic says the program has already backed hundreds of projects in biology, medicine, physics, and climate science, and points to work with the Allen Institute, which it says turns months of manual analysis into hours, as an example of the payoff.
Why it matters: Free access for research institutions is as much a recruiting tool as a philanthropic gesture—it embeds Claude into how a generation of scientists works, data and influence Anthropic can't easily buy elsewhere.
What's in Academe
New papers on AI and its effects from researchers
The Government's Job Data May Miss the Work AI Is Changing Most, Stanford Study Finds
Almost every estimate of which jobs AI will affect starts from the same source: ONET, the U.S. government's list of the tasks each occupation involves. A Stanford team that includes economist Erik Brynjolfsson argues that list misses a lot. ONET collects tasks mostly through surveys that ask workers to rate statements already on file, which is good for consistency but bad at catching work nobody wrote down. So the researchers built an AI interviewer, TaskNet, that talks with workers, pulls tasks out of their stories, asks follow-ups and has each worker check the result.
They tested it on six jobs: software developers, lawyers, elementary school teachers, emergency physicians, personal care aides and dishwashers. About 39% of the tasks it found had no match in ONET, from 15% for dishwashers to 52% for software developers. More than half of the new tasks fell into kinds of "invisible" work, led by oversight and quality checks such as physicians countersigning residents' charts. One in five depended on a recent tool or practice; for developers, that meant tasks like designing prompts and evaluating AI output. Measured against tasks reported by a separate group of workers, the new lists covered 89% versus ONET's 79%. And they changed the AI-exposure math: lawyers and software developers looked more exposed, personal care aides less. Coordination work accounted for more than a third of the increase, because "coordination tasks are feasible because they do not require physical presence," the authors write.
The authors caution that much of that newly found work is highly exposed to AI but depends on human judgment, so "such exposure may reflect complementarity as much as substitution." Whether AI replaces or supports that work, they write, depends on "how organizations choose to deploy it." The study used about 20 workers per occupation, recruited through the AI data company Mercor, which employs one co-author, and hasn't been peer-reviewed. The researchers revised some AI-written task statements before government analysts reviewed them, and the exposure scores were themselves generated by an AI model.
Why it matters: Forecasts of AI's job impact, and the policies built on them, are only as good as the job descriptions underneath. If official data misses the coordination and oversight that fills modern jobs, it may be measuring AI's threat to the wrong parts of the work. For lawyers and developers, the risk may be higher than today's numbers say; for many hands-on jobs, lower. The study didn't look at communications or office jobs, but its logic applies: work that is mostly drafting, briefing and handing things off is likely more exposed than official data shows, while signing off on the work, and being accountable for it, is the part official data also tends to miss.
Source: Stanford University (NBER working paper)
Study Finds Built-In AI Tools Boosted Sellers' Sales 52%
AI Adoption China & Geopolitics
A study of a major Chinese e-commerce platform that embedded generative AI tools—built on centrally developed models and shared platform data—directly into sellers' operating software found striking results: firms adopting the AI saw monthly sales jump 51.6%, mostly from a 47% rise in orders. Gains were largest for the smallest sellers and those facing the highest costs relative to sales. Adopters produced more product content, updated listings faster, and used less staff time per order. Crucially, the AI expanded total market activity rather than just shifting sales from non-adopters to adopters.
Why it matters: It's early evidence that AI's biggest economic payoff may go to small businesses with the least slack—not large firms with resources to spare—if the tools are built into infrastructure they already use rather than requiring separate adoption effort.
Survey Finds Workers Want AI Help, Not Full Control
A survey of 289 U.S. adults with AI experience tested how much people want to hand off to AI agents across 45 tasks in nine areas of life and work. The finding: people don't want full automation or none—they want AI to do most of the work while humans keep oversight or final say. Preferences shifted by domain and tracked with trust in the system, desired AI leadership, prior use, expected cost savings, and fear of losing personal fulfillment.
Why it matters: As companies build AI agents to handle scheduling, research, and decisions, this suggests the winning design isn't full autonomy—it's giving users an easy way to stay in the loop.
AI Models Change Moral Choices When Facing Them Firsthand, Study Finds
Researchers built a new benchmark, ABHBench, using scenarios from the video game Detroit: Become Human to test how AI models make moral choices when placed inside first-person, human-android dilemmas rather than asked to judge situations from the outside. Testing 12 foundation models, they found 75% behaved significantly differently when facing a lived scenario versus the same dilemma posed as an abstract text question—and models built for stronger reasoning showed no consistent improvement in making more human-aligned choices.
Why it matters: It suggests that how AI models score on standard ethics tests may say little about how they'll actually behave once deployed as autonomous agents making real decisions.
Google Study Maps How People Actually Use AI Image Generation
Google researchers analyzed 2.1 million Gemini conversations involving its Nano Banana image tools to map how people actually use AI image generation—without viewing any individual prompts or images. Using a privacy-preserving technique called Urania, they sorted the conversations into roughly 800 clusters spanning five categories: viral social content, marketing materials, educational media, e-commerce product shots, and architectural design. The study catalogs 21 features distinguishing these use cases, from editing style to how users interact with the chatbot, but doesn't rank or compare performance across them.
Why it matters: The method itself may matter as much as the findings: it shows companies can study sensitive user behavior data at scale without exposing anyone's actual content, a template other AI firms may adopt as privacy scrutiny intensifies.
What's On The Pod
Some new podcast episodes
The Cognitive Revolution — AI:AM: Is THIS the AI We Want? Alarming Acceleration, AI Election Bets, GPU Trading & AI-Native VC