D.A.D. joke of the day: I asked AI to summarize my 40-page report. It came back with one sentence: "This could have been an email."
What's New
AI developments from the last 24 hours
One Founder, Claude and 11 Labs: Anthropic Buys a One-Person Drug Company
Anthropic has bought Mithril Medicine, a drug startup with one full-time employee. Founder Cody Tranbarger used Claude to run contracts, logistics and experiments with 11 outside labs, and handed Claude much of the rest of the company's operations. The result, Anthropic says: a possible treatment for a rare childhood muscular dystrophy went from an idea to promising results in mice in under 18 months.
The disease is LAMA2-related congenital muscular dystrophy. Most children show symptoms at birth or within months. It brings severe weakness and breathing problems, and there is no approved treatment. Tranbarger now works at Anthropic and plans more programs. If the candidates keep holding up, Anthropic expects human trials could start as soon as late 2028. Its name for the model: "Pipeline-in-a-Person."
The deal closed in July, R&D World reports, and no price was disclosed. Tranbarger wasn't entirely alone: he directed part-time consultants, and Atlas Venture put up seed money. Endpoints News adds a twist: he is the person behind "Professor Oak," a popular anonymous biotech account on X, and has now revealed himself.
The deal came with an update on Anthropic's drug research with six partners, including Bristol Myers Squibb. One example: at the drug developer Enveda, Anthropic says, Claude caught every documentation problem its human reviewers found, plus seven they missed, and cut review time by 89%.
Why it matters: It's the latest sign that AI labs see medicine as their next big frontier. A week ago, OpenAI bought Orchestra Bio, whose software coordinates drug development (D.A.D., October 21). Anthropic bought the biotech startup Coefficient Bio in April and just invited outside life-sciences researchers to work with its staff (D.A.D., October 26). Mithril goes a step further: Anthropic now runs drug programs of its own. Look at what Claude actually did, though: the coordination, the paperwork, the project management. That's the same job OpenAI just paid for, and the part of the work most professionals would recognize from their own jobs. Sam Altman has long predicted that AI would let one person build a billion-dollar company. This isn't that, but it shows the shape. The caveats are big. The results are in mice, not people. The technical report isn't peer reviewed. And the claims come from the company that bought Mithril and sells the AI.
Sources: Anthropic · Mithril technical report (PDF) · R&D World · Endpoints News · Tranbarger on X · Andrew Curran on X
Google Previews AI Tool That Builds Explorable Virtual Worlds With Sound
Google unveiled Genie 4, its latest "world model" that generates interactive, navigable environments in real time—complete with physics, human movement, and now native audio like music, ambient sound, and dialogue. Users can steer these simulations with plain-language instructions as they unfold. The system is launching as a research preview to select testers, with Google planning wider access through Project Genie for AI Ultra subscribers and early partners using the Gemini API and AI Studio.
Why it matters: World models like this underpin next-generation training grounds for robotics, game design, and simulation-based planning, and audio-and-physics realism is the detail that decides whether these tools feel like gimmicks or genuine testing environments.
Discuss on Hacker News · Source: deepmind.google
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Trump Says He'll Ban the Term "Artificial Intelligence" — in "Everything Else," Too
President Trump went after The Wall Street Journal on Truth Social on Tuesday for still saying "artificial intelligence." He called it "a very bad and inaccurate name" and asked when the paper would switch to Super Intelligence, "the name which virtually everyone else is using, or soon will be." His promise: "In short order, we will be officially BANNING the term, 'AI/Artificial Intelligence,' in all relevant documents, and everything else, due to its inaccuracy."
This is the latest round. Trump announced the rename at the UN in September (D.A.D., September 23). A week later he signed an executive order telling federal agencies to write "Super Intelligence" and "SI" (D.A.D., September 30). In early October he posted that anyone who says "Artificial Intelligence" is "THE ENEMY" (D.A.D., October 12). The new part is "everything else." The September order covers only the government's own documents, and Trump didn't say how a wider ban would work.
The rename has also set off a speculation boom. Registrations of Slovenia's .si web addresses jumped nearly a hundredfold after the executive order, the security firm Netcraft found. Critics, led by commentator Adam Cochran, point to thousands of .si addresses registered months before the rebrand and now listed for resale, some for millions of dollars, and allege that insiders knew in advance. No one has shown who bought them. Netcraft puts the rush down to speculators: a .si address costs about $11 a year.
Why it matters: The label is the sideshow. The same executive order gives the White House science adviser until late November to propose a new legal definition of Super Intelligence. That definition decides which systems federal rules cover at all. Watch that deadline, not the Journal's style guide.
Sources: Deadline · Mediaite · IBTimes · Donald Trump on Truth Social · Andrew Curran on X
Authors Reject Anthropic's Book Piracy Settlement, Demand More
Some authors say Anthropic's $1.5 billion book-piracy settlement, covering nearly 500,000 works at roughly $3,000 each, isn't enough. Novelists Tobias Wolff and Chang-rae Lee have opted out of the Bartz v. Anthropic deal to pursue a separate suit, Cruz v. Anthropic, seeking up to $150,000 per infringed work under federal copyright law. Judge William Alsup's underlying ruling still stands: training on pirated books is illegal, but training on legally acquired books counts as fair use (D.A.D., July 22).
Why it matters: The breakaway lawsuit tests whether the industry's largest copyright settlement actually sets the price of literary piracy—or just the floor.
Discuss on Hacker News · Source: stanforddaily.com
What's in the Lab
New announcements from major AI labs
AI Handles Most Customer Service Calls at Fast-Growing Voice Platform
Voice AI platform Retell AI says OpenAI models now generate responses for 80% of the 100 million-plus customer service calls it handles monthly for over 4,000 businesses, up from 30 million calls at the start of the year. In Retell's own benchmark, published in an OpenAI case study, OpenAI models beat Google's Gemini and Anthropic's Claude on following instructions and avoiding made-up answers, and customers report resolving more than 80% of calls.
Why it matters: It's a real-world signal that OpenAI's models are winning the unglamorous but massive market of automated phone support, where reliability and cost—not flashy capability—decide which AI vendor gets picked.
ChatGPT Will Now Run Oracle Business Apps Directly
OpenAI and Oracle are deepening their partnership, building Oracle applications and data directly into ChatGPT while using OpenAI models to power new Oracle products. The package includes an Oracle AI Workspace inside ChatGPT, plus security and small-business tools. Oracle says it already has 130,000 active ChatGPT users and over 95,000 using OpenAI's coding tool Codex internally, and that its recruiting team cut market-research prep time from days to under 20 minutes using ChatGPT Work.
Why it matters: This extends OpenAI's push to embed itself inside the enterprise software millions of office workers already use daily, following similar bets with AT&T and Google Cloud (D.A.D., October 22).
What's in Academe
New papers on AI and its effects from researchers
Users Credit AI for Execution, Themselves for Creative Direction, Study Finds
A study of 20 people using an AI image-editing tool called Roop found a consistent pattern in how they divvy up creative credit: they see themselves as the directors who decided what should change, while crediting AI for the visual execution—the actual pixels. Roop tracked edits with a "heat map" showing which tool touched which part of an image. Reviewing that map helped people remember what they'd done, but rarely changed who they thought deserved credit.
Why it matters: As AI co-editing tools spread into design, marketing, and publishing workflows, questions of authorship and credit aren't just philosophical—they'll shape copyright claims, portfolio attribution, and disputes over who actually made a given piece of work.
Study Finds Many AI Chatbots Leak Your Conversations to Trackers
A new academic audit of 42 AI chatbot platforms—general-purpose and specialized ones covering health, legal, and education—found widespread leakage of conversational data to third-party trackers. Researchers intercepted network traffic and browser activity to find that 21 platforms leaked conversation titles, 9 leaked verbatim user prompts, and a notable number sent user emails to trackers, potentially letting outside companies link what you typed to who you are.
Why it matters: If you or your employees discuss sensitive health, legal, or business matters with AI chatbots assuming privacy, this research suggests that assumption is often wrong, with real exposure risk for regulated industries.
AI Models Still Trust What People Say Over What They Do
A new study tested whether giving AI models access to conversation, not just past behavior, improves their ability to predict what a person will actually do next. It doesn't. Researchers had eight open-weight language models forecast cooperation decisions in a repeated negotiation game. Past decisions alone predicted outcomes well (AUC .80, a strong score). Adding dialogue didn't raise overall accuracy—instead, models took people at their word: when someone said they'd cooperate, predicted follow-through jumped to 69%, but people actually followed through only 40% of the time. Models also largely ignored which personality type of counterpart users were facing, despite that mattering a lot to real behavior.
Why it matters: As companies deploy AI to anticipate customer or employee behavior from chat transcripts, this suggests the models may be trusting what people say over what they're likely to do—a costly blind spot for anything from sales forecasting to fraud detection.
AI Usage Varies Widely Across Europe, But Trust Levels Match
A study comparing Denmark and Romania—the EU's highest and lowest users of generative AI—found a surprising disconnect: 48.4% of Danes used tools like ChatGPT in the past three months versus just 17.8% of Romanians, a massive usage gap. But trust in AI was roughly equal between the two countries, even though Danes trust their government and legal institutions far more than Romanians do. Researchers say trust in AI tracks digital skills and confidence in AI governance, not actual usage or broader institutional trust.
Why it matters: For policymakers and educators pushing AI adoption, the finding suggests that using AI more doesn't build more trust in it—meaning governance and skills training, not usage mandates, may be the real levers for responsible adoption.
AI Text Floods Internet's Hidden Pages, Spares Popular Sites
A study of more than 5 million webpages found AI-generated text has spread unevenly across the internet. English-language sites carry more AI text than other languages, and .com domains show far more than .gov or .edu sites. The sharpest growth is on obscure, poorly-linked pages—the ones fewest people see—while heavily-trafficked, well-connected pages remain the least likely to contain AI-written content. Researchers used multiple detection tools combined statistically to correct for false positives.
Why it matters: If AI content is concentrating in the web's periphery while trusted, high-traffic sources stay more human-written, it complicates assumptions about how much AI text ordinary searches and AI models themselves are actually absorbing.
What's On The Pod
Some new podcast episodes
AI in Business — How Security Leaders Are Responding to Machine Speed Threats - with Sridhar Boddula of MassMutual
AI in Business — Measuring Marketing When AI Makes the Click Disappear - with Gillian Hinkle of Salesforce
How I AI — Why Reddit built its own AI prototyping tool | Blair Metcalf