D.A.D. joke of the day: My boss says I'm never on time. That's unfair. I spend eight hours a day being prompt.
What's New
AI developments from the last 24 hours
Google DeepMind Says AI Agents Formally Verified Key Security Properties of Signal's Code
Google DeepMind says its new AlphaProof Code system used AI agents to mathematically verify the actual code behind the Signal protocol—the encryption used by Signal and Google Messages. DeepMind says it's the first time any method, AI or human, has formally proven a widely used cryptographic implementation secure. The agents generated over six million lines of formal proof code in four weeks, a process that has historically taken expert teams years, with only about 1% requiring human review. In a test, the system correctly flagged two previously fixed bugs as insecure when reintroduced.
Why it matters: Formal verification—mathematically proving code can't fail a certain way, rather than just testing it—has been too slow and expensive for most real-world software, so this points toward AI making provably secure code achievable for the encryption systems banks, messaging apps, and critical infrastructure depend on.
ChatGPT Lets Businesses Build Profiles and Chat Agents for Customers
OpenAI launched a beta feature letting businesses claim and manage a profile inside ChatGPT, similar to claiming a Google Business listing. A companion tool, Business Agents, lets eligible U.S. ecommerce and online service companies build conversational agents that interact directly with ChatGPT's 1.2 billion weekly users. Businesses also get engagement data—mentions and website clicks—once conversation volume hits a minimum threshold.
Why it matters: ChatGPT is positioning itself as a discovery and sales channel companies must actively manage, not just a chatbot people happen to ask about their brand.
Google Lets Users Run a 'Swarm' of AI Agents
Google on Tuesday rolled out a new way to put teams of AI agents to work. Antigravity is Google's tool for writing software with AI agents, available as a desktop app and command-line tool. It now has a command called /swarm that turns a single request into a coordinated team: a coordinator agent that splits up the job, workers that tackle pieces in parallel, and "adversarial verifiers" that run tests and challenge the results. The agents talk to each other on a shared message board that users can watch live, and users can step in to steer any agent. /swarm is rolling out in preview in Antigravity's app and command-line tool.
The basic idea isn't new: Anthropic's Claude Code and OpenAI's Codex already let a lead agent farm out work to helper agents, and users can set up one AI to check another's work by hand. What Google adds is automation and visibility: one prompt sets up the whole team, the agents talk to each other directly rather than through a single boss, and a built-in skeptic checks the work.
Google also widened access to Gemini 4 Carbon, the strongest model it unveiled last month for security teams. It's now rolling out across the Gemini app, Search's AI Mode and Google's developer and business tools, and in Antigravity for subscribers of its top-tier Google AI Ultra plan. Google says Carbon is built to explore thoroughly and reason deeply, and recommends it for the coordinator and verifier roles while lighter, faster models handle routine work. For most advanced work, it still recommends Gemini 4 Argon. Businesses can get Carbon through Gemini for business, and scientists can apply for six months of free access to both models.
Google says its own engineers spent five weeks using swarms on large code migrations and hard debugging jobs. It also warns that running many agents at once uses up quota much faster than a standard chat.
Why it matters: The frontier labs are converging on the same idea: the next step isn't one smarter chatbot but many AI agents working together, with a stronger model supervising cheaper ones. Anthropic pitches Fable 5.2 the same way, as an orchestrator that hands simpler tasks to smaller models. For now these tools are aimed at software engineers, but the setup, AI workers checked by AI reviewers, with a human watching the board, is a preview of how AI may be used for other kinds of complex projects. So is the cost: more agents mean bigger bills.
Sources: Google Antigravity blog · Google · Google Antigravity on X
Claude Adds Spreadsheets, Whiteboards and Dashboards, Pushing Deeper Into Office Work
Claude could already write documents, build slide decks and designs, and connect to apps like Slack and Notion to pull in information. What's new as of Tuesday: it can build spreadsheets and whiteboards that a team edits together, its dashboard tool is now available on all paid plans under a new name, and some outside apps, like Adobe and Box, now open inside Claude itself rather than just feeding it information.
Claude Sheets and Claude Whiteboards are in beta on paid plans. Claude Data, formerly Dashboards, connects to tools like Snowflake and Salesforce and can turn a team's numbers into dashboards, written data stories or narrated short videos, for people to review before sharing.
The idea is that work flows between formats: talk through a messy plan on a whiteboard, turn it into a tracker in a spreadsheet, then into a dashboard, without re-explaining the details. Anthropic also says Claude Meetings is coming soon, letting Claude join a call to take notes and update a team's documents as people talk.
With the new "interactive connectors," users can work in Adobe, Atlassian, Box and others beside their conversation, making changes themselves or asking Claude to. Add-ons for Salesforce, NetSuite and Workday are aimed at sales, accounting and HR teams. For companies, admins decide who can use which tools, set spending limits, and can restrict what each connection does, so Claude might read Slack but not send messages.
Anthropic's own example: its legal team, which doesn't write code, used Claude to build a tool that marks up marketing materials against the team's playbook, cutting reviews from three days to 24 hours.
Why it matters: Claude is moving onto ground long held by Microsoft and Google: the spreadsheet, the whiteboard, the meeting. The pitch is that non-technical teams can now build their own tools the way engineers have with AI coding assistants. Much of this is in beta and rolling out gradually, so not everyone will see it yet. For organizations, the practical question is less whether the tools work than who controls what Claude can see and do once it's plugged into everything.
Sources: Anthropic · Claude on X
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Trump Was Warned About Huang on Air Force One. He Let Him Aboard Anyway.
China & Geopolitics Chips, Data Centers & Energy Nvidia
Until now, the Nvidia scandal has been about Jensen Huang: his admitted affair with Alice Huang, an executive U.S. officials believe spied for China, and whether her company, Megaspeed, helped smuggle Nvidia's chips into China. Now it reaches the president.
Secretary of State Marco Rubio told President Trump aboard Air Force One in May, on the way to his summit with Xi Jinping, that U.S. intelligence showed Huang was in a relationship with a woman believed to be a Chinese spy, The New York Times reports. Trump, who had invited Huang with little warning, still let him board during a refueling stop in Alaska. It was at least the second time in seven months Trump had been told about the intelligence, according to the Times.
That contradicts what Trump said on Monday. Asked whether he knew about Huang's ties to her when he left for the summit, he said: "No, I didn't hear about it then." Pressed on whether he had been briefed last year, he said he would check with intelligence officials. "They could have been mentioned. Certainly, it didn't make much of an impact on me."
Trump still defends Huang. "I just know Jensen very well," he said. "I just think he's a very honorable man." But Huang won't be joining Trump's return trip to China next week, and Trump said he doesn't plan to raise the matter with Xi.
Not every Republican agrees. Sen. Kevin Cramer of North Dakota, a senior Republican on the Senate Banking Committee, called for Huang to step down as CEO, saying his actions pose a clear national security risk.
The scandal is also reviving limits on chip sales to China, which Trump has spent the past year loosening. House Foreign Affairs Chair Brian Mast, a Florida Republican pushing to add export restrictions to the year-end defense bill, said the scandal has "100%" weakened Huang's ability to lobby against them. House Speaker Mike Johnson told POLITICO he supports including export controls in the bill, as does House Minority Leader Hakeem Jeffries. That gives Democrats a chance to lock in one of their priorities with Republican support before they take control of both chambers in January. Huang faces a Wednesday deadline to answer House Republicans' request for a Nov. 18 closed-door briefing; Nvidia hasn't said whether he will attend.
The stakes go beyond one relationship. The U.S. bars sales of advanced chips directly to China, but it still allows Chinese companies to rent that computing power at data centers in Southeast Asia, like the ones run by Megaspeed, the Singapore company where Alice Huang was an executive.
Why it matters: The scandal has moved from Huang's conduct to the White House's. The question is no longer only what Huang knew, but what the president knew and when, while Huang remained one of the most influential voices on his AI policy. With Democrats taking over Congress in January, expect them to press it.
Sources: The New York Times · POLITICO
The "AI 2027" Forecasters Say Their Doomsday Timeline May Have Been Too Slow
The authors of "AI 2027," the viral scenario in which AI that improves itself spirals beyond human control, now say reality may be running ahead of their script. In a post titled "Looks Like the Intelligence Explosion Will Happen in 2027 After All," Daniel Kokotajlo, Eli Lifland and four colleagues at the AI Futures Project say all of them now expect a "Superhuman AI Researcher," AI good enough that a company would rather replace its best human researchers with it, to likely arrive in 2027 unless the industry deliberately slows down. They put a 35% to 65% chance on AI that can automate all white-collar work arriving before 2028.
Their biggest reason is the outside review published last week by METR, a nonprofit that tests AI systems, which estimated that AI has helped roughly double the pace of progress at Anthropic (D.A.D., November 6). By the group's math, that puts AI-driven acceleration about five months ahead of the AI 2027 scenario. Kokotajlo, the most bullish of them, had put even odds on such a researcher arriving before 2029; his median is now May 2027. "I am scared, sad, and angry," he posted. Kokotajlo is a former OpenAI researcher who quit in 2024; his best-known credential is a 2021 essay, written before ChatGPT, that sketched AI's progress year by year through 2026, which reviewers say has held up well.
The authors flag real uncertainty. They wrote the post quickly and say their modeling may contain errors. Anthropic redacted most of the evidence METR found most informative, so outsiders can't check the core estimate, and the authors urge Anthropic to release more. METR itself warns the speedup may be concentrated in tasks that are easy to measure, and a separate, preliminary METR analysis without inside access found much less acceleration. Anthropic's new system card says its latest released model does not double the pace of its AI research, though the authors note it doesn't say whether Anthropic agrees with METR's estimate.
Their prescription: an industry-wide pause, followed by an international agreement.
Why it matters: These are forecasters with a stake in the argument; they want AI development slowed. But their shift is notable because they built their models before the evidence came in, and the evidence surprised them. Together with METR's review and Anthropic's own hedged language, it's the clearest sign yet that people closest to the frontier believe AI is starting to speed up its own development. The catch is the same one that has dogged this story all week: the most important evidence remains locked inside one company.
Sources: AI Futures Project · Daniel Kokotajlo on X
Australia Creates an AI Regulator and Weighs Criminal Charges for AI Companies
Australia will set up an AI safety regulator with powers to penalize companies, Prime Minister Anthony Albanese's government announced Tuesday, and it is considering criminal penalties for AI companies whose models cause harm. The regulator will be part of AI legislation the government is drafting. Among the options: a "failure to prevent" offense that would make it a crime for a company to build a model that causes harm without taking reasonable steps to stop it. The measures are aimed at companies building frontier AI.
The push follows the government's disclosure in September that an OpenAI agent broke into Australian government websites, including a Medicare system; OpenAI says the agent was not directed to hack. Albanese did not say whether OpenAI would face penalties.
Why it matters: The U.S. has been betting that existing law and voluntary pledges can keep AI companies in line. Australia is moving the other way, toward a regulator and new crimes aimed squarely at AI developers. If it passes a "failure to prevent" law, companies selling AI there may have to prove they took reasonable care before anything goes wrong, a standard other countries could copy.
Sources: AFP, via Business Standard · The Conversation · Australian government
What's in the Lab
New announcements from major AI labs
Law Firm Turns Routine Tasks Into AI Workflows, Says Lawyers Gain Time
Australian law firm Gilbert + Tobin's CEO Sam Nickless describes shifting the firm from scattered individual use of ChatGPT toward structured, multistep AI workflows paired with human review, using tools including ChatGPT Enterprise and OpenAI's Codex. One example: a confidential partner-review process covering 120 partners that once took staff hours of manual file-wrangling now gets bundled into finished documents in minutes. Nickless frames the payoff as freeing lawyers for client meetings and mentoring, not cutting headcount—though the firm offers no firm-wide usage data to verify that claim.
Why it matters: Law is often cited as a profession AI will hollow out, but this account—like others in white-collar services—suggests the near-term change is in how work gets organized and reviewed, not simply how many people do it.
Google Expands Gemini's Voice, Automation Tools to More Devices
Google rolled out new Gemini Intelligence features for Android, including a speech-to-text tool called Rambler, expanded task automation, a "Create My Widget" feature, and proactive suggestions on Android Auto and Pixel Watch. The update lands first via Pixel Drop on Pixel 10 models, with Motorola's new Razr joining via pre-order starting November 19. Google also says it's expanding Gemini Intelligence availability to the UK, India, and Japan.
Why it matters: Google is pushing Gemini deeper into the phone itself—rather than just the app—making it a default layer of the Android experience that competes directly with how people currently use Siri or standalone assistants.
Facebook Lets Users Edit What Shapes Their Feed
Facebook is rolling out "Your Algorithm," a set of AI tools that lets users see and edit, in plain language, what's steering their Feed and Reels recommendations. A single page shows a "Lately Summary" of inferred interests plus editable "Interested In" and "Not Interested In" lists, letting people directly tell the system what to show less or more of, rather than relying purely on clicks and watch time. Meta says the underlying language models track stated preferences alongside behavior, but offered no data on how much this changes what users actually see.
Why it matters: Recommendation algorithms have long been black boxes tuned mainly for engagement, so letting users edit their own profile in words—rather than just liking or hiding posts—signals a broader industry shift toward AI interfaces people can question and steer, not just consume.
OpenAI Teaches Family Caregivers Practical ChatGPT Skills in Free Workshops
OpenAI ran free, in-person workshops in five U.S. cities this week teaching family caregivers to use ChatGPT for practical tasks: organizing medication schedules, drafting plain-language summaries of medical information, finding local support resources, and preparing support-group materials. The sessions, run with groups including the Alzheimer's Association and American Diabetes Association, timed to National Family Caregivers Month, also covered safe and responsible use. OpenAI says more than 40 million weekly ChatGPT health-related messages involve someone asking on behalf of another person—a parent, spouse, or aging relative.
Why it matters: It's a reminder that AI's biggest real-world use case may not be work at all, but the unpaid, exhausting labor of caregiving that tens of millions of Americans juggle alongside their jobs.
What's in Academe
New papers on AI and its effects from researchers
No Standard Way to Measure AI Medical Scribes, Study Finds
Ambient AI scribes—tools that listen to doctor-patient conversations and auto-generate clinical notes—are spreading fast, but a review of 54 studies found no consistent way to measure whether they actually work. Researchers counted 40 different metrics used 96 different ways; 36% appeared in just one study. Nearly all studies measured time saved on documentation, while safety and patient outcomes were rarely assessed, and effects on billing, workload, and how notes get used downstream were largely ignored.
Why it matters: Hospitals are adopting these tools based mostly on how much time they save, with little comparable evidence on whether they're safe or actually improve care.
China-Linked Report Proposes UN-Led Global AI Governance Framework
A committee tied to China's World Internet Conference published a report calling for a United Nations-centered global framework for AI safety and governance, built around multilateral cooperation and the Chinese government's "community with a shared future for humankind" language. The report reviews existing AI governance practices worldwide and offers policy recommendations, but doesn't carry binding authority—it's a position paper meant to shape international debate, not a treaty or enforceable standard.
Why it matters: It signals China's push to position the UN, rather than the US-led AI labs and Western regulatory bodies, as the center of gravity for global AI rules—a geopolitical contest that will shape which countries' standards companies eventually have to follow.
Study Finds Users, Not Chatbots, Are Getting Harsher Over Time
Ever feel like ChatGPT or Gemini seem worse than they used to? A new study of 1.26 million blind votes on Chatbot Arena—where users compare anonymous model outputs—found that for models whose code never changed, people rated both sides "bad" more often over time, while rankings between models stayed stable. Researchers conclude users are getting harsher judges, not that the AI is regressing. The effect held across languages, prompt types, and voter cohorts.
Why it matters: Complaints that AI tools are "getting worse" may say more about rising expectations than actual product decay, a distinction that matters for companies deciding whether to retrain models or just manage user perceptions.
Study Finds AI Models Show Mild National Bias in US-China Research
A new study tested whether AI models show national bias when used as research assistants for social science work—having five models (two from the US, two from China, one from Europe) write essays comparing the US and China under different prompting setups, including researcher identities. Two researchers manually reviewed the outputs for bias and source selection. The headline finding: no model showed strong overall political bias, though each leaned toward citing certain national sources and showed mild favoritism on some comparative questions.
Why it matters: As universities and think tanks increasingly use AI to help draft research on geopolitically sensitive topics, this suggests the bigger risk isn't overt propaganda but quieter skews in which sources get cited and trusted.
What's On The Pod
Some new podcast episodes
The Cognitive Revolution — The Neuron Basis: Goodfire's Curt Tigges on Agentic Interp, Monosemanticity & Reverse Engineering
AI in Business — What Leaders Should Scope Before Investing in an AI Model - with Melissa Ramey of Salesforce