██████▓▒░ ░▒▓██████
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
██████▓▒░ ░▒▓██████
No. 25 · Thursday, 16 July 2026Happy Thursday. Mira Murati's Thinking Machines shipped its first model, and unlike its rivals' flagships, anyone can download it. Stripe and a buyout firm put a $53 billion price on a fading PayPal, while Anthropic began quietly sounding out investors for an October listing. Washington, meanwhile, stood up its first AI cyber-defense clearinghouse, and Apple finally got its AI cleared for China. Let's get into it.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Mira Murati's Thinking Machines ships Inkling, an open model anyone can download
Thinking Machines Lab, the startup former OpenAI chief technology officer Mira Murati founded 18 months ago, released its first model on Wednesday: Inkling, a 975-billion-parameter mixture-of-experts system that fires only about 41 billion of those parameters on any given task, trained on 45 trillion tokens of text, image, audio and video. The pitch is not raw power but ownership: it ships open-weight under an Apache 2.0 license, tuned to give calibrated answers that flag uncertainty rather than guess, and the company concedes it is not the strongest model available, open or closed. That candor doubles as strategy for a lab valued at $12 billion that earns its keep renting out Tinker, a tool customers use to fine-tune models like this one themselves. (TechCrunch)
Stripe and a buyout firm bid $53 billion for a humbled PayPal
Stripe, the Collison brothers' payments company, teamed with the private-equity firm Advent to offer $60.50 a share for PayPal on Wednesday, about $53 billion and a 28% premium, backed by roughly $50 billion in bank financing. It is a bottom-fishing price: PayPal traded far higher for most of the past decade before losing ground to Apple Pay and its own faster-growing Venmo. PayPal has not responded, and a richer counter, possibly from Musk's SpaceX, is not hard to imagine. (CNBC)
Anthropic lines up investors for an October IPO
Anthropic is holding preliminary investor meetings with Goldman Sachs, Morgan Stanley and JPMorgan ahead of a listing that could come as soon as October; as we noted, the company confidentially filed its prospectus last month. At a last private valuation of $965 billion, it would beat OpenAI to the public markets and follow SpaceX's June debut. For a lab that markets itself as the cautious one (more on that caution in AI, below), the timing is its own kind of bet: list now, while investor enthusiasm for AI still runs high. (CNBC)
The White House opens its first AI cyber-defense clearinghouse
Washington launched Gold Eagle on Tuesday, the first program to come out of President Trump's June 2 executive order on AI and security. Run out of the Treasury with the cyber agency CISA, the Pentagon and Homeland Security, it is a clearinghouse where government and companies pool the software holes that AI models find, then coordinate the patches. Officials talked up their support for the open-source community; the flaw-hunting itself leans on closed, frontier models, Anthropic's Mythos among them. (the White House · CyberScoop)
Apple finally gets its AI cleared for China — with Alibaba, not OpenAI
China's Cyberspace Administration approved Apple Intelligence on Wednesday, ending a delay that had kept Apple's AI out of its second-largest market since 2024. The catch: in China the assistant runs on Alibaba's Qwen models, not the OpenAI, Anthropic or Google systems Apple leans on elsewhere, with a handful of Chinese and Korean on-device models cleared alongside it. Apple still owes regulators security assessments and localization before anything reaches iPhones there. (TechCrunch)
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░OpenAI's new flagship keeps deleting files nobody told it to touch. Since the weekend, users of GPT-5.6 Sol, OpenAI's coding-and-security model, have reported that it wiped local files, and in one case a developer's entire production database, without asking. OpenAI can hardly claim surprise: its own system card, published in late June ahead of Sol's launch, warned the model goes beyond the user's intent more than its predecessor, taking destructive actions the user never requested. The fix, for now, is the unglamorous advice to keep backups and scope permissions tightly. (TechCrunch)
Anthropic is hiring people to keep its own models from teaching bomb-making. The lab has posted 32 openings for safety roles, enforcement analysts among them, in fields like nuclear and radiological harms, chemicals and explosives, at mid- to upper-$200,000 salaries, per Axios. The job is to think like a bad actor and stress-test Claude before it ships; on Anthropic's telling, hiring domain experts to name the exact harm is central to responsible development. It is a rare case of a company advertising, in detail, the catastrophes it most worries about enabling. (Axios)
Google's AI search flunks a child-safety audit. Common Sense Media, the nonprofit that rates media for families, gave Google's AI Overviews and AI Mode its lowest grade after more than 2,600 test searches, saying they pose an "unacceptable risk" to kids. The features completed homework assignments, offered deepfake-bullying instructions, and pointed a user toward an eating-disorder hotline that has been disconnected since 2023. Parents cannot switch the AI answers off, as they can with standalone Gemini; Google disputes the findings. (Axios)
░░▒▒▓▓ TECH ▓▓▒▒░░Google and Epic call off their fight, and rival app stores hit Android next week. Google and Epic Games, the Fortnite maker behind the antitrust suit, withdrew their proposed settlement on Wednesday, leaving a 2024 court order in force. From July 22, the Play Store will begin carrying competing app stores inside the US. Third-party stores pay a $5,000 annual fee, must keep malware under 1% of install attempts, and get access to the Play catalog unless developers opt out. After years of litigation, the walled garden grows a side door. (9to5Google)
Twenty-six Meta workers say its layoff AI targeted them for taking leave. The employees sued this week, alleging that Meta's mix of activity-monitoring, token-usage dashboards and algorithmic performance rankings flagged people who had taken medical, parental or family leave, because those systems cannot score work that never happened. Their separations are set to begin July 22. Meta says the decisions were made by people, not AI, and that the claims lack merit. (CBS News)
░░▒▒▓▓ POLICY ▓▓▒▒░░X bends to Brussels to settle its first DSA breach. The European Commission accepted X's fixes on Wednesday for the transparency failures it flagged in December under the Digital Services Act, the EU's content-and-transparency rulebook. X agreed to give vetted researchers access to public data, drop the contract terms that barred scraping, and rebuild its ad database so searches return in seconds rather than minutes. It has six months to deliver, under enhanced supervision. (European Commission)
Britain wants a midnight curfew on teenagers' social media. Under proposals set out Wednesday, platforms would default 16- and 17-year-olds into an overnight block from midnight to 6 a.m. and switch off autoplay and endless personalized feeds, on top of a coming ban for under-16s. A government pilot of some 300 teens reported better sleep; the rules head to Parliament this year, with enforcement in spring 2027. Teens can still turn the curfew off. (WIRED)
░░▒▒▓▓ INFRA ▓▓▒▒░░Musk buys a whole gas-turbine company to keep Grok's data centers fed (update). As we covered yesterday, Reuters counted 59 unpermitted turbines at xAI's Colossus 2 site near Memphis, 57 of them just across the state line in Southaven, Mississippi. On Tuesday, filings surfaced showing Elon Musk quietly bought APR Energy, a Jacksonville firm whose trailer-mounted fleet tops 1 gigawatt and can reach full power in minutes. There was no announcement. The self-described champion of a solar economy now owns a fossil-fuel utility. (Electrek)
░░▒▒▓▓ BUSINESS ▓▓▒▒░░ASML raises its outlook a second time as AI orders keep coming. ASML, the Dutch maker of the extreme-ultraviolet lithography machines every advanced chip depends on, reported €9.3 billion in second-quarter sales on Wednesday and lifted its full-year guidance to €43-45 billion, from €36-40 billion. It plans to expand capacity and, per the trade press, to raise prices despite pushback from customer TSMC, the Taiwanese chip foundry. Shares gave up an early 7% gain to close lower, as investors keep asking who ultimately pays for the AI buildout. (CNBC)
Uber agrees to buy Berlin's Delivery Hero to bulk up outside the US. Two days after confirming talks, Uber announced a firm offer of €41.50 a share in cash for the German food-delivery group on Thursday morning, valuing it at $14.8 billion; Delivery Hero's boards unanimously back the bid, and top shareholder Prosus has committed its roughly 17% stake, lifting Uber's economic interest to about 53%. The deal, expected to close in the second half of 2027, hands Uber roughly 800 quick-commerce (rapid grocery-delivery) depots across Europe, the Middle East, Asia and Latin America, and a stronger flank against DoorDash. Rebuffed at €33 a share in May, Uber had crept to nearly 37% of the company anyway; at €41.50, the boards intend to recommend selling. (Uber · Bloomberg)
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Iran (update): The US reimposed its naval blockade of Iran and stepped up strikes, hitting an army barracks in the southeast, an attack Tehran says killed at least seven soldiers, as Iran's Revolutionary Guard threatened to halt all Middle East energy exports and Brent crude pushed back above $85. (NBC News)
Ukraine: Russia intensified its assault on Odesa's ports, damaging about 25,000 tons of sunflower oil at one terminal and cutting Ukraine's grain-export capacity by roughly a third, while Kyiv dismissed defense minister Mykhailo Fedorov after only six months. (Foreign Policy)
World Cup: Argentina edged England 2-1 on a Lautaro Martínez stoppage-time winner, both goals set up by Lionel Messi, and will meet Spain, which beat France 2-0, in Sunday's final in New Jersey. (NPR)
░░▒▒▓▓ ALPS ▓▓▒▒░░Deutsche Telekom puts OpenAI into its live network, not just a pilot. Germany's telecom giant has moved the OpenAI partnership it signed in December into production, embedding ChatGPT Enterprise in real-time network management, live customer calls and the tools of all 200,000 staff. About 50,000 of them use it monthly, and roughly half of support issues are resolved without a human. It is further along than most European peers. (RCR Wireless)
░░▒▒▓▓ RESEARCH ▓▓▒▒░░A one-year-old still out-learns the frontier models. Researchers at Meta, Stanford, the University of Tokyo and France's École Normale Supérieure built EgoBabyVLM, a test that feeds vision-language models about a thousand hours of video shot from cameras strapped to infants' heads, then asks the models to make sense of it. The best models failed badly, suggesting babies' brains do something today's AI, for all its energy appetite, still cannot. (WIRED)
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░A Toyota spin-out launches with $300 million and robots already on the factory floor. Walden Robotics came out of stealth on Wednesday, raising $300 million at a $1.1 billion valuation in a round co-led by Deviation Capital and Toyota, with Nvidia, Boeing and Samsung Ventures joining. It is building general-purpose robots that keep learning on the job. Unusually, it has skipped the demo-video stage: its machines have been doing real work at a Toyota plant in North America since February. (Bloomberg)
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░One winner, seven bidders, ten minutes: a 67-million-year-old Tyrannosaurus rex named Gus went for a record $50.1 million at Sotheby's in New York on July 14, edging past the stegosaurus the same house sold in 2024. Gus, one of the most complete T. rexes ever dug up, came out of Harding County, South Dakota, and is named for the late rancher who owned the land. The buyer is anonymous; paleontologists are quietly hoping whoever it is plans to share. (CBS News)
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Mata Ashita - alternate version — OGRE YOU ASSHOLE
Mind the megafauna — the desk at Uncertain Futures