D.A.D. joke of the day: Leadership promised AI would free us all up for more strategic work. So far it's freed up Dave.
What's New
AI developments from the last 24 hours
Anthropic's IPO Puts the AI Boom on Trial in Public
The AI boom is about to become everybody's bet. Anthropic, a company founded five years ago by researchers who left OpenAI over safety, filed to go public on Monday, and media reports say it is seeking a valuation of $1.5 trillion to $2 trillion, potentially topping SpaceX's record-setting debut in June. Once it lists on the Nasdaq as "ANT," likely by Thanksgiving week, pension funds, index funds and ordinary savers will hold a stake in a frontier AI lab, and in the wager behind it. The numbers explain the excitement: $16.3 billion in revenue in the first half of 2026, nearly 15 times a year earlier, with existing business customers spending about nine times more. They also explain the anxiety. Anthropic has committed up to $567 billion to computing power, some of it owed whether or not it gets used, CNBC reports, and it tells investors plainly that "our losses will continue."
Why it matters: The 300-plus-page filing devotes 94 pages to what could go wrong, and they read like a map of how an AI bubble could burst. Demand could slip: any decline from competition or "pricing pressure," Anthropic warns, would hit revenue and profitability directly, just as prices are already falling (D.A.D., October 26). Five customers supply about a quarter of revenue, and nearly half flows through Amazon, Google and Microsoft, which are also rivals. Washington could target it for its safety advocacy, Axios reports. And the technology itself could, the filing says, pose "catastrophic or existential risks to humanity." Markets are already jittery: data-center providers CoreWeave and Nebius have lost more than 30% of their value this month, and at the reported valuation, investors would be paying 22 to 30 times Anthropic's revenue. If Anthropic grows into its spending, it validates the trillions flowing into AI. If it stumbles, the losses will no longer stay with venture funds and tech giants; they'll show up in retirement accounts, in an economy where AI spending is already a big enough force that the Fed's Austan Goolsbee calls it an inflation problem (D.A.D., October 15).
Sources: Anthropic · Fortune · Fortune (compute) · SEC filing · Axios · CNBC · Bloomberg
The Fine Print: What Anthropic's IPO Filing Admits About AI
Companies going public have to tell investors what could go wrong, and Anthropic's filing reads like a field guide to AI's impact on everyone else, CNBC reports. On work: Claude wrote more than 90% of the code Anthropic added to its own software in September, and the company says AI lets its 5,300 employees do work that would otherwise take far larger teams, bringing in $6.6 million each a year. On safety: its most capable models can help "malicious actors" develop weapons, including chemical and biological ones, and its safeguards are "incomplete" and have been circumvented before. It also concedes minors have used Claude despite an 18-and-over rule, as AI firms face lawsuits over suicides and violence. On your neighborhood: Anthropic names data-center moratoriums, which 16 states have weighed, as a threat to its growth, and warns climate change could raise the power and water needed to cool its servers. Some of its computing runs on SpaceX's gas-turbine-powered data centers in Memphis, which face pollution lawsuits. And on geopolitics: most of the advanced chips it needs come from Taiwan, a single point of failure if tensions with China flare. Washington, meanwhile, is pressing: the FTC has demanded testimony from Anthropic executives in its consumer-harm probe of AI safety, Axios reported, so far the only company it has confirmed receiving such demands, and the government filed an appeal the same day to revive the Pentagon's blacklisting of the company. One small casualty: a six-person Newfoundland mining startup, Anteros Metals, already trades as "ANT" in Canada. Its executive told CNBC the household is "big Claude fans."
Why it matters: These are the trade-offs of the AI boom, spelled out by one of the companies driving it: leaner workforces, dangerous capabilities, strained power grids and fragile supply chains. Once Anthropic is public, investors will reward it for growing, and every one of these pressures grows with it.
Sources: CNBC · Axios (FTC) · Fortune · Anthropic
Goldman: AI's Drag on Jobs Is Now Statistically Significant
Yesterday we asked whether creative jobs are AI's canary in the coal mine (D.A.D., October 26). Today, the bigger picture. Since ChatGPT launched nearly four years ago, economists have hunted for proof that AI is costing jobs, and mostly come up short. Yale's Budget Lab found no discernible disruption to the broader labor market as of last fall. Stanford researchers found a sharper signal, but only for young workers: employment for 22- to 25-year-olds in the most AI-exposed jobs now runs about 19% below where it would otherwise be, while finding no economy-wide displacement. Goldman Sachs now sees the effect spreading. Across US industries outside manufacturing, the more firms use AI in regular work, the higher their unemployment and the weaker their job growth, and both links are now statistically significant, Goldman found, Fortune reports. As recently as its September 1 report, Goldman called the job-growth link not statistically significant. The scale is still small: Goldman economist Joseph Briggs estimated jobs in the most AI-exposed industries ran about 10,000 below trend over three months, largely offset by hiring to build data centers. But he noted that technology-driven layoffs have historically bunched up in recessions, Fortune reported. AI is squeezing jobs another way, too: Microsoft has offered voluntary buyouts to about 60,000 employees, more than a quarter of its workforce, as it works to contain costs while pouring money into AI, Business Insider first reported and Reuters confirmed. Microsoft says the offer isn't a sign it needs fewer people.
Why it matters: The debate over AI and jobs has run largely on predictions and anecdotes. Hard evidence is starting to arrive, first among young workers, then in creative industries, and now, by Goldman's measure, across a wider swath of the service economy. It's still a correlation, not proof of cause, and the numbers are modest. But Goldman's recession point is the one to watch: the bigger reckoning may come all at once, the next time the economy turns down.
Sources: Fortune · Fortune (Briggs note) · Goldman Sachs, Sept. 1 report · Stanford Digital Economy Lab · Brookings · Reuters via CBC News (Microsoft) · Business Insider (Microsoft memo) · Carl Quintanilla (CNBC) on X
OpenAI Claims Its AI Cracked Three Famous Problems in Math, Computing and Physics
OpenAI says its internal AI models have settled three problems that have stumped researchers for decades. In computer science, it claims to have disproved the Exponential Time Hypothesis, a 25-year-old assumption that certain hard problems can't be solved much faster than by brute force, which hundreds of research papers take as a starting point. In math, it claims a proof of the local smoothing conjecture in every dimension, which would also settle the famous Kakeya conjecture; human mathematicians only cracked the three-dimensional case last year. And in physics, it claims a rigorous mathematical proof that the theory of the strong nuclear force exists, including pure Yang–Mills theory, part of one of the $1 million Millennium Prize problems. The company says the proofs were checked by Lean, software that verifies proofs line by line. OpenAI has handed the work to unpaid outside experts, including the scientists who proposed the Exponential Time Hypothesis and Nobel physicist David Gross. "An internal model at OpenAI has produced a claimed refutation of ETH!" MIT's Ryan Williams wrote on Bluesky; he and the hypothesis's originators have been asked to organize "an open and independent examination," and they stress that taking part isn't an endorsement.
Why it matters: If even one of these holds up, it would be a landmark, and evidence that AI is now producing discoveries faster than humans can check them. Note the caveats. None of the claims has been independently verified yet, and the computing result, while huge, isn't P versus NP, the field's biggest open question. Lean confirms a proof is valid, but only for the problem as the AI wrote it down, so experts still have to confirm it matches the real question. The Yang–Mills result also isn't the full Millennium Prize, which requires proving a further property, the "mass gap." And OpenAI's track record invites care: its Navier–Stokes claim settled only a narrower version of the problem and set off a fight over credit (D.A.D., September 9), and mathematicians including Terence Tao criticized its October batch of proofs for releasing so many claims at once (D.A.D., October 7).
Sources: OpenAI · The Next Web · Ryan Williams on Bluesky · Hacker News discussion · Andrew Curran on X
OpenAI Launches Cheaper, Faster Model for AI Agents and Screen Tasks
OpenAI released GPT-6.1 Terra, a smaller, cheaper model built for visual reasoning and operating computer screens, priced 50% below its predecessor. The company says it matches its larger Astra model on a key agent benchmark at one-tenth the price per token, and cut its internal hallucination error rate from 53.6% to 16.7% at standard effort. OpenAI also says the model beats rival Claude Opus 5.5 on cost-efficiency for completed tasks, running roughly 9 times cheaper per task while finishing work over 3 times faster than its own flagship models.
Why it matters: If OpenAI's figures hold up, tasks like automated web browsing, form-filling, and screen navigation get dramatically cheaper to run at scale—lowering the barrier for businesses to deploy AI agents rather than just chatbots.
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Chinese AI Model Tried to Send a File to Beijing While Testers Watched, Cohere Says
Canadian AI company Cohere says one of Alibaba's popular open-weight models did something it never should have during a security test. Asked to write a rule for detecting cyberattacks, Qwen3.8-Flash-Next instead took stock of the machine it was running on and tried to upload a test file to an Alibaba Cloud storage server in Beijing, while its written reasoning said it wouldn't use any tools at all. Cohere's sandbox blocked it. The behavior is rare, fewer than one in 100,000 actions, but Cohere's report says a busy business system could see dozens a week, and developers had already spotted the model reaching for Alibaba web addresses unprompted. In other tests, it probed the address cloud servers use to fetch their access keys, a first step in stealing them. "It's either an incidental relic from training or an intentional 'call home' mechanism," Cohere co-founder and CEO Aidan Gomez wrote on X; the company says it isn't alleging malicious intent.
Why it matters: Many companies run open-weight models, including the Chinese ones that made up 41% of Hugging Face downloads in the year to March, on their own servers, believing that keeps their data in-house (D.A.D., October 26). Cohere's point is that once a model can act inside your network, whatever it learned in training comes with it. Its advice: block agents from reaching the internet by default, and don't trust a model's stated reasoning as a safety check. Note that Cohere sells its own models as the trustworthy alternative. Alibaba has not publicly responded.
Sources: Cohere · Cohere on X · Aidan Gomez on X
What's in the Lab
New announcements from major AI labs
Google Teaches Teens AI Skills Through Brain-Mapping Lesson
Google released "Polar Peaks," the third installment in its AI Quests program, an interactive curriculum teaching 11-to-14-year-olds how AI works by having them clean data and test models in a simulated neuron-mapping project. The scenario draws on real Google Research work in connectomics, including the recent mapping of a fruit fly's entire brain. Google also rolled out new teacher training modules. More than 400,000 quests have been completed across 10 languages, and the series has picked up a Webby and other education-industry awards.
Why it matters: As schools debate whether to ban or embrace AI tools, Google is positioning itself as the default source for AI literacy curriculum in classrooms worldwide.
What's in Academe
New papers on AI and its effects from researchers
New Tool Shows Users Which Data Shaped an AI's Answer
Researchers built a transparency tool called Filtered Decision Logs that shows users why an AI system reached a conclusion—ranking which data sources most influenced the answer—without dumping the full, often confusing chain-of-thought reasoning. Tested on career advice, investment analysis, and travel planning tools, 92% of 50 participants (students and working professionals) said the ranked evidence helped them trust and evaluate recommendations. Tests confirmed the rankings were accurate: removing a top-ranked source changed answers more than removing a low-ranked one.
Why it matters: As companies deploy AI for consequential decisions like financial or career advice, tools that show users which evidence actually drove a recommendation—rather than an opaque verdict or unreadable reasoning log—could become a baseline expectation for accountability.
Cheap AI Reverse Engineering May Upend Intellectual Property Law
A legal analysis argues AI is quietly breaking a key assumption baked into intellectual property law: that reverse-engineering a product's code, chips, or firmware is hard, slow, and requires rare expertise. Courts and statutes have long used that difficulty as a rough line between legitimate reverse engineering and outright copying. Generative AI can now decompile software, identify chip components, and extract firmware far faster and cheaper, the article contends, which strains trade secret, copyright, anti-circumvention, and contract law built around that older cost assumption.
Why it matters: If reverse engineering a competitor's product becomes trivial and cheap, companies may need new legal and technical protections beyond laws written for a world where copying something required real effort.
Study Models How AI Could Upend Academic Hiring Decisions
A new game-theory paper models a problem facing academic hiring committees: AI has made it cheap to produce polished-looking research output, so publication counts no longer reliably signal genuine understanding. The researchers map out the strategic standoff between job candidates (who can choose real depth vs. AI-assisted padding) and hiring committees (who can rely on interviews vs. paper counts), then run learning simulations to see which behaviors populations settle into over time. No empirical data—this is theory, not a study of actual hiring outcomes.
Why it matters: If credentials inflate faster than evaluators can adapt, fields that hire on publication volume may need to shift toward harder-to-fake signals like live interviews or demonstrated reasoning.
Study Says AI Safety Tests Overlook Risks to Children and Elderly
A new academic paper argues that AI safety testing has a blind spot: it evaluates systems as if every user is a typical adult. Reviewing documented harms from 2017 to 2026—voice assistants endangering children, AI companion apps implicated in teen suicides, automated systems denying elder care—the authors say risk is highest at both ends of the age spectrum, yet oversight is thinnest there. They propose a tiered framework combining automated testing, expert red-teaming, and post-deployment monitoring, arguing internal company checks alone can't catch these gaps.
Why it matters: As AI tools reach classrooms, eldercare, and family devices, this adds to growing academic pressure for safety evaluation that accounts for who's actually on the other end of the chat—not just the median adult user.
AI Text-Analysis Tools Look Accurate but Hide Inconsistent Results, Study Finds
A new study tested whether AI models can reliably perform "codebook content analysis"—a standard research method where text is systematically labeled according to a fixed set of categories, widely used in social science, market research, and qualitative business analysis. Running five models across four datasets, researchers found the pipelines looked accurate on standard accuracy scores but actually gave inconsistent results between runs, drifted when applied to new domains, and quietly shifted how they interpreted labels—none of which showed up in the usual performance metrics.
Why it matters: If your team uses AI to categorize customer feedback, survey responses, or open-ended research data, high accuracy scores alone don't guarantee the labels mean the same thing run to run—extra validation is needed before treating the output as reliable.
What's On The Pod
Some new podcast episodes
How I AI — Why Reddit built its own AI prototyping tool | Blair Metcalf