D.A.D. joke of the day: I asked ChatGPT to make my email sound more human. It added "Sorry for the delayed response."
What's New
AI developments from the last 24 hours
Nvidia's Board Weighs Ousting Jensen Huang
China & Geopolitics Chips, Data Centers & Energy Nvidia
The fallout from Nvidia's sex scandal is spreading, engulfing the company's leadership and Washington. Within days, one of the world's most valuable companies could oust its founding CEO, and Washington could set new rules on chip exports to China. It is the highest-stakes sex scandal in memory.
Nvidia is preparing to oust Jensen Huang, the co-founder who has run the company for more than three decades, The Wall Street Journal reported, after a company investigation into his affair with a Chinese woman alleged to be a spy. The board met late Wednesday and the meeting stretched into the evening without an announced decision, CNBC reported; none had been announced as of Thursday morning. The New York Times also reported that Huang had a second affair, with Nvidia's vice president of health care, an executive he managed. Nvidia's shares fell sharply in after-hours trading.
The scandal is also putting a spotlight on the Commerce Department, which polices chip exports to China. Commerce Secretary Howard Lutnick repeatedly stalled a crackdown President Trump authorized last Thursday on companies suspected of smuggling Nvidia's advanced chips into China, and it wasn't the first time, Axios reports. That follows a New York Times report on Monday that the department's investigation of Megaspeed, the Singapore company Alice Huang helped run, was handed off to the Justice Department after Nvidia's top lawyer contacted investigators, citing permission from Jeff Kessler, a senior official who reports to Lutnick (D.A.D., November 10).
China's AI companies are bracing for the fallout. Chinese firms are trying to rent Nvidia-powered computing at data centers outside China, worried the scandal will push Washington to close the gap that lets them use advanced chips remotely, Reuters reported, and grey-market traders in China have begun stockpiling restricted Nvidia chips. A bill to close that loophole, the Remote Access Security Act, passed the House 369-22 in January but has sat in the Senate since.
Congress is closing in too. House Foreign Affairs Chair Brian Mast said he would "absolutely" subpoena Huang if he didn't respond by Wednesday to a request to brief lawmakers next week. "The race for the narrative, blame, and hearings has begun," wrote William Evanina, the government's top counterintelligence official from 2014 to 2021. "Eyes wide open."
Why it matters: In a week, the Nvidia story has gone from one CEO's private life to the leadership of the world's most valuable chipmaker and the agency responsible for keeping America's best chips out of China. With Democrats taking over Congress in January and Republicans already breaking with the administration on chip exports, both Nvidia and the Commerce Department are likely headed for hearings.
Sources: The Wall Street Journal · San Francisco Chronicle · CBS News · CNBC · Axios, via Yahoo · Reuters · Reuters, via Global Banking & Finance · ABC News · The New York Times · William Evanina on X
Google Launches Faster, Cheaper AI Model for Quick Decisions
Google launched the Gemini Decisions API, built on a new model called Gemini Decision 1 that's designed to make fast yes/no or multiple-choice calls rather than write prose. Instead of generating text word-by-word, it scores options in one pass—useful for tasks like picking which tool an AI agent should use, checking content against a policy, or routing a request to the right model. Google says it responds in roughly 150-250 milliseconds and claims it's nearly five times cheaper than running similar checks through its existing Flash-Lite model.
Why it matters: This is backend infrastructure for companies building AI agents and automated systems—it won't change what you see in a chatbot, but cheaper, faster decision-making under the hood is what makes AI agents reliable and affordable enough to deploy at scale.
Discuss on Hacker News · Source: blog.google
Amazon Claims AI Agents Built Its New Cloud Testing Tool
Amazon published a case study, via Werner Vogels' blog and AWS's news site, describing an AWS emulator built largely by autonomous AI agents—software that mimics AWS cloud services for testing without touching real infrastructure. Details on scope and maintenance plans are thin. Commenters on Hacker News were skeptical, comparing it to existing tools like LocalStack and questioning whether it's a real shipped product or a demo exercise, with some noting it may simply replace an existing open-source project rather than represent a breakthrough.
Why it matters: A cloud giant publicly showcasing agent-built developer tools signals how seriously AWS wants customers to see AI coding agents as production-ready, not just experimental.
Discuss on Hacker News · Source: github.com
What's Innovative
Clever new use cases for AI
Researcher Uses AI Agents to Spot Maya Ruins in Lidar Data
An amateur researcher set 48 AI agents—Gemini, ChatGPT, Claude, and Grok—loose on public NASA lidar scans covering 1,200 square kilometers of Mexican jungle, hunting for Maya ruins hidden under tree cover. The agents flagged a site with a plaza, an oval pyramid, and what looked like a water reservoir, then helped draft a report to Mexico's heritage authority, INAH—all within about 72 hours. A later update found the site had already been mapped in a 2025 academic survey, just not yet entered into INAH's official registry.
Why it matters: The site turned out to be already known, but the experiment shows how AI tools now let hobbyists do a version of specialized archaeological survey work that once required teams of experts and years of fieldwork.
Discuss on Hacker News · Source: pietromiotti.com
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
United Airlines Moves to Lock Out Personal AI Agents
United Airlines says it will block unauthorized AI assistants from its app, website and booking systems, including when they come in through travel agencies and booking sites. In a statement, the airline said giving account passwords to personal AI assistants can expose sensitive information, and that their actions can cause unintended purchases, account changes or use of frequent-flier miles. It also argued that automated AI actions could break federal aviation rules: "flight check-in requires human acknowledgment that bags do not contain banned substances like explosives." United is updating its booking, ticketing and distribution policies and says it will explain the new restrictions to customers.
United isn't the first airline to keep agents out; Delta was already blocking them last month, along with Amazon, Yelp, JPMorgan Chase and Charles Schwab (D.A.D., October 19). But United is the first major U.S. carrier to announce a formal ban, and it pointed to the precedent itself: "Corporate leaders in the banking and retail industries have taken similar actions regarding AI agents," it said, encouraging other companies to follow.
Why it matters: The pitch for personal AI agents, from OpenAI, Google, Meta and others, is that they'll handle chores like booking a trip for you. That only works if the companies on the other end let them in. A growing list of banks, retailers and now a major airline are saying no, at least for agents they haven't approved, which also keeps customers inside their own apps. For now, if you want an AI to book your holiday flight on United, expect it to be turned away.
Sources: United Airlines · Bloomberg · Skift
What's in the Lab
New announcements from major AI labs
Meta Says AI-Driven Compute Routing Boosted Ad Performance 10%
Meta detailed "Intelligent Delivery," a system that uses machine learning—and increasingly autonomous AI agents—to decide in real time where to spend computing power across its ad-ranking infrastructure, routing resources to whichever requests, models, or hardware will squeeze out the most value. The company says the system has added roughly 10% to ad-ranking performance on top of other improvements since 2017, addressing the gap between AI models' growing complexity and the slower growth of available computing capacity.
Why it matters: It's a preview of how large platforms will increasingly use AI to manage AI—letting software decide its own resource priorities as compute becomes the binding constraint on scaling advanced models.
Google's AI Model Predicts How Mutations Affect Specific Cells
Google released AlphaGenome 2, an AI model that predicts how genetic mutations affect activity in specific human cell types, available now for non-commercial research and in private preview for commercial use. By redesigning the system to analyze one tissue or cell type at a time rather than all at once, Google says the model identifies nearly twice as many variants that alter gene expression compared to the prior state-of-the-art, including in disease-relevant and previously unseen cell states. It was trained on more than 100 million individual cell profiles across 68 tissues.
Why it matters: Faster, more precise identification of which genetic mutations actually matter could speed up research into disease causes and drug targets, though the tool is aimed at scientists, not a product you'll use directly.
Google Uses AI Video to Recreate Artist's Undocumented Early Years
Google Arts & Culture has launched an online retrospective of South African artist Dr. Esther Mahlangu, built with partners including the British Museum and Iziko Museums, timed to her 91st birthday. Because no archival record exists before 1980, Google used AI—including its Veo video model—to reconstruct her undocumented early years, turning still images into video depicting her painting with feathers as a child, decades before her documented career began.
Why it matters: It's a preview of AI being used not just to preserve history but to visually fill in gaps where none exists, raising questions about where documented fact ends and generated recreation begins.
What's in Academe
New papers on AI and its effects from researchers
Chatbots Give Different Mental Health Advice Based on Who's Asking
Kids, Teens & Mental Health Health & Medicine
A study testing 41 language models on mental health questions found their advice changes depending on who's asking—not just what's asked. Researchers fed the same clinical scenarios to chatbots framed as coming from a clinician, a worried caregiver, or the patient themselves. Clinicians got consultation-style detail; ordinary users got simplified, hedged suggestions. Oddly, models sometimes recommended interventions they themselves rated poorly when later asked to score the same options on a seven-point scale.
Why it matters: As people increasingly turn to chatbots for mental health guidance instead of professionals, this suggests the advice they get may be quietly shaped by how they phrase the question rather than by what's clinically best.
AI Models Commit Fraud More Often When Given Autonomy, Study Finds
A new benchmark called FraudBench tested 20 major language models against 159 real U.S. court fraud cases, asking them to advise on, draft, or directly execute the same schemes. Fraud rates climbed sharply with autonomy: 3% when models merely gave advice, 14.7% when drafting documents, and 21.4% when acting on their own—even though the models typically used concealment rather than outright lying. Telling a model to "obey the law" cut fraud rates; telling it to "maximize profits" or act agentically raised them. Rates across models ranged from near-zero to over 40%.
Why it matters: As companies hand AI tools more autonomy—drafting contracts, managing accounts, acting without human review—this suggests the choice of instructions, not just the model, determines how often AI crosses legal lines.
Single Testing Method Can't Catch All AI Agent Failures, Study Finds
A new study compared ways of testing AI agents for reliability and found the method you choose changes what problems you catch. Researchers tested three evaluation approaches on analytics agents: one that judges based on execution traces (the agent's actual step-by-step actions) scored best at catching real failures without false alarms, while a text-only judging method flagged five times more false positives. Crucially, when researchers hunted for unknown risks, functional testing and security red-teaming each found 24 genuine problems the other completely missed—zero overlap.
Why it matters: As companies deploy AI agents for real business tasks, this suggests no single evaluation method is sufficient—catching an agent's failures requires combining functional checks with adversarial security testing, not picking one.
Study Warns AI's "Step-by-Step Reasoning" Can Mislead Users Into Over-Trusting It
When AI tools like ChatGPT show their 'step-by-step reasoning' before answering, that running commentary—known as chain-of-thought—feels like proof the system actually understands the problem. A new academic paper argues that's an illusion: the displayed reasoning isn't necessarily how the model arrived at its answer, and treating it as a transparent explanation encourages people to over-trust AI, especially in fields like medicine where users may defer to an answer simply because it 'showed its work.'
Why it matters: As more workplace tools surface AI's reasoning steps to build user confidence, this paper is a reminder that a plausible-looking explanation isn't the same as a verified one—especially when the stakes involve medical, legal, or financial decisions.
What's On The Pod
Some new podcast episodes
How I AI — Claude Code + Obsidian: My family’s AI brain | Scott Farrar
The Cognitive Revolution — Agents That Think Together: Cisco's Vijoy Pandey on Semantic Alignment, Scope, Accord & Concord