Good day . Mark Zuckerberg says Meta was ready to write a $200 billion check before it had seen the models, and Palantir’s CEO warns that AI can predict what you’re about to think. Plus Google Play now runs apps you vibe-coded for yourself, dead startups’ Slack logs are being sold to AI labs, and brands are auditing what ChatGPT says about them. Let’s get into it.
Zuckerberg Weighed a $200B Blind Bet Business Insider
What happened: On the Cheeky Pint podcast, released Monday, Mark Zuckerberg said that at the end of last year, before Meta had seen its new models, he was ready to “write this check for, like, I don’t know, like $100 billion, $200 billion,” and asked chief AI officer Alexandr Wang, “Are you sure this is going to work?” He didn’t say what timeframe or scope the commitment would cover.
Why it matters: Meta’s own guidance puts 2026 capital spending at $130 billion to $145 billion, and Cantor Fitzgerald analysts expect $200 billion in 2027, per forecasts Business Insider saw. Yesterday we covered Meta saying Muse Wonder puts it back at the frontier — this is what it cost to find out.
What everyone’s saying: Wang said Meta’s CFO, Susan Li, pulled him aside “many times” to say that if it didn’t work, “it’s going to be very, very bad in financial terms.” Investors have already shown their limits: Meta shares fell more than 6% in after-hours trading after April’s earnings, when the company raised its spending forecast.
My read between the lines: Zuckerberg’s defense is “in order for it not to work out, it’s not just that we have to fail, AI overall has to fail,” and even then Meta would have “other uses to productively use compute.” That’s a hedge dressed as confidence. The bet is less on Meta’s models than on a mountain of GPUs staying useful to somebody, and the check was sized before anyone had seen the product.
📖 Further reading: I Make AI Versions of Myself for a Living. This One I Didn’t Agree To. — my paid piece on Meta’s Muse Image, pulled three days after launch, and who holds the leash when a model can build a working copy of your face. Worth a read now that you know how big a check sits behind the engine.
Meta weighed a $200 billion bet on AI before it knew whether it would work. Your bet is smaller: can an AI coworker take the weekly report off your plate? Viktor is an AI agent that lives in your team’s Slack, connects to 3,000+ of your tools, and does real work: reports, dashboards, code and campaigns. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor →
Karp Warns AI Can Predict Your Thoughts The News International
What happened: In an interview with Louis Robson, who leads Palantir Sovereign, Palantir CEO Alex Karp said the technology used to predict actions on the battlefield can also predict people’s behavior at home. “Which may mean you are predicting their actual thoughts one is going to have in a domestic context,” he said. “Now this is something we must look at controlling it.”
Why it matters: Palantir is the U.S. government’s primary enterprise software provider for defense logistics, target acquisition and real-time operational analysis, per The News International, so this is a vendor describing the limits of its own category. Karp wants tighter limits at home to protect privacy, but argues the same restrictions on the battlefield could leave America unable to compete with adversaries predicting its moves.
What everyone’s saying: Karp warned that “you could comply with the letter of the law and totally eviscerate the spirit of the law here,” and the clip is circulating on X as a bombshell. It also fits a pattern: in September Karp told CNBC that leading AI labs may need to be nationalized because of the technology’s “unlimited risks,” and that builders of harmful tech should be held liable.
My read between the lines: Should this surprise anyone? Facebook used to brag that it could tell when couples were about to break up before they changed their relationship status. That was a social network guessing from likes and photos, and this is the same prediction business with far more data and far better models. And the warning comes from a company that sells prediction for a living.
The Brief is free and stays free. But a story like what happens to a company’s data after it dies deserves more than four bullets. Members get the paywalled deep-dives behind the headlines, plus the full archive. Become a member →
Google Play Opens to Vibe-Coded Apps TechCrunch
What happened: Google said Tuesday that people can now build an Android app in Google AI Studio and launch it on their own device through Google Play using a standard Google account. The apps stay private: they don’t appear in Play search or listings and aren’t shared with anyone but their creator. At launch it’s limited to users 18 and older in the U.S.
Why it matters: Until now, running your own native Android app meant publishing it to Play’s roughly 3 billion users, so many people stuck to web apps. Now describing an app in plain language and installing it on your phone is a one-step path. The apps still have to follow Google Play’s existing policies.
What everyone’s saying: Google’s Sam Bright framed it as moving from “there’s an app for that” to “there’s an app for me.” ZDNET’s headline put it more bluntly: Google lets you make personal Android apps for free now, with just prompts.
My read between the lines: “Private” is the starting point. Google says that in the months ahead you’ll be able to share these apps with a network of trusted friends, and an app that takes off can graduate to the public store after trust and safety reviews. Google is building a farm system: your weekend prompt becomes the next app in a store it controls.
📖 Further reading: I Pay $20,000 a Year for a Database. I Talk to Something Else. — my paid piece on the friction tax clunky software charges and why I stopped opening the dashboard. Personal apps push the same idea one step further.
AI Labs Are Buying Dead Startups’ Slack Logs MIT Technology Review
What happened: As startup shutdowns rise, firms like SimpleClosure and Sunset are selling dead companies’ code and workplace data (emails, chats, documents) to AI labs, training-environment vendors and mature tech companies. SimpleClosure’s Asset Hub has handled more than 150 deals, and 47% of its deals now include code or data. One founder netted about $25,000 for his code base and expects roughly $15,000 more for his workspace data.
Why it matters: Labs are running out of public and synthetic training data, and real archives avoid errors synthetic worlds make. In August, Google outbid Mercor to buy Spirit Airlines’ data for $10 million, including 100 million emails, 500 million Teams messages and 176,000 employee records. Bloomberg reported in September that SpaceX’s AI unit is weighing buying data from struggling startups.
What everyone’s saying: Lawyer Greg Ewing calls this data “legally radioactive” and describes a “Wild West of AI training,” noting emails and calendars hold private details like doctor appointments and school pickups. Spirit’s flight attendants’ union objected on privacy grounds. Emory law professor Ifeoma Ajunwa calls the payouts “captured capital,” since proceeds typically go to creditors and investors, not employees.
My read between the lines: Nobody has yet published tests comparing real-world data against the alternatives, so the prices rest on what buyers call “research theses.” Labs are paying for evidence they haven’t shown works, with data from people who weren’t asked and won’t be paid. Your old Slack history is a trove of how people disagree, which is the stuff agents need to learn.
Brands Are Auditing What ChatGPT Says About Them MIT Sloan Management Review (paywalled)
What happened: MIT Sloan Management Review published research from the University of Auckland and HEC Montréal on brand visibility in AI search. One case, a New Zealand investment platform given the pseudonym Stakewise, was getting about 700 ChatGPT referral sessions a month by mid-2025. When its team queried ChatGPT with real customer questions, they found answers that were “confident, fluent, and frequently wrong.”
Why it matters: ChatGPT repeatedly said the platform’s KiwiSaver product offered five funds when it has seven, mischaracterized its fees, and described a protective investment safeguard as restrictive. For a retirement product, that’s misinformation at the exact moment someone is deciding. Interviews with 15 senior marketing leaders found big gaps in AI visibility metrics, and leaders underestimated how fast errors spread.
What everyone’s saying: The authors argue that SEO is now entwined with generative engine optimization, or GEO (also called AEO): brands must manage how visible AI answers make them and how accurately they describe them. As zero-click AI responses replace links, new metrics replace traditional traffic analytics.
My read between the lines: Stakewise only knew it had a problem because it audited the answers, not the analytics, and its analytics looked great. The page you’re ranking for can be fine while the summary the model gives is wrong, and there’s no appeals desk for a chatbot’s opinion of your business. The fix on offer is to check regularly and feed the model better sources.
📖 Further reading: Google’s Invisible Axe: The Silent Killer of Small Businesses — my account of Google unlisting a business of mine and how little recourse there was when a platform’s version of us was wrong. AI answers are the next platform with that power.
That’s your AI Brief for Wednesday.
—Nicholas from Artificially Intimidating
