░▓███▓░ ░▓███▓░
░██ █████░ ░██ █████░
░█████████░ ░█████████░
░▓███▓░ ░▓███▓░
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. 81 · Monday, 5 October 2026Hope the weekend treated you well. Trump put Jay Clayton, his intelligence chief, in charge of a task force called the Super Intelligence Force, which makes him the AI czar he floated last week. Apple says it will introduce additional controls on a Mac permission that lets apps read a user's files, mail and messages, citing the risk from AI agents. An OpenAI employee who led work on its safety reports resigned and wrote in The Atlantic that the company's culture is broken.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Trump put intelligence chief Jay Clayton in charge of a task force called the Super Intelligence Force, making him the AI czar (update)
As we noted last week, Trump had said Jay Clayton, his director of national intelligence, would be a good AI czar. Over the weekend Trump put him in charge of a task force he calls the Super Intelligence Force (more on the name in COMMS DESK, below), which in effect makes him that czar. The group has 120 days to report on the risks and opportunities of AI, Clayton told the Wall Street Journal, according to SiliconANGLE, and the report will also cover how the government handles disclosures of security breaches. Clayton will chair the group, the Journal reported, and its three vice chairs include Federal Trade Commission Chairman Andrew Ferguson, whose agency opened an inquiry into the AI labs, as we covered on Thursday (TechCrunch · SiliconANGLE).
Apple says it will introduce additional controls on a Mac permission that lets apps read a user's files, mail and messages, citing the risk from AI agents
Apple said on Friday that it will add new controls around Full Disk Access, the macOS permission that lets an app read a user's files, messages, mail and browsing history. Its stated reason is that increasingly capable AI agents make access that broad riskier. Reuters reported that the Mac will flag AI requests for a user's data, and that the change follows complaints about Muse, Meta's personal agent (more in AI, below). WIRED reported over the weekend that Muse builds detailed profiles of a user's friends and family. In the Netherlands and Belgium, the eyewear chain Hans Anders said on Friday that it had suspended sales of Meta's Ray-Ban smart glasses, citing privacy concerns (TechCrunch · Reuters via The Economic Times · WIRED · Reuters).
An OpenAI employee who led work on its safety reports resigned and wrote in The Atlantic that the company's culture is broken
David Robinson spent about three and a half years at OpenAI and led the writing of the safety reports that accompany its major product launches. In an essay published on Saturday he blamed the company's culture and wrote that AI companies are not being "nearly careful enough" and that, given the risk assessment he cites, the "time for trial and error is over." OpenAI answered that it pauses training or holds back models when it needs to slow down. The essay came two days after the Wall Street Journal reported that OpenAI had cut ties with three safety researchers, as we noted on Friday. Robinson concedes the genre is a familiar one: the lab employee who issues a dire warning while resigning (The Atlantic · TechCrunch).
Amazon pledged more than $1 billion to the towns around its data centers and said it no longer uses nondisclosure agreements with government agencies
The money, announced on Friday, is spread over five years and goes to education, job training, energy upgrades and other projects in communities where Amazon builds. Amazon said it will pay for its own energy and infrastructure improvements and that its facilities will be water positive, returning more water than they use, by 2030. It also said it no longer uses nondisclosure agreements with the government agencies it works with, a practice Virginia barred for its own state agencies last month, as we covered on September 21. Matt Garman, who runs Amazon Web Services, said more than 100 proposed moratoriums on data centers threaten the American lead in AI. GeekWire noted what the pledge leaves out: Microsoft promised in January not to seek local tax breaks for its data centers, and Amazon, asked whether it would do the same, did not make that commitment (Amazon · WIRED · GeekWire).
Banks and Blackstone are assembling $60 billion to finance chips for Anthropic and other companies, with Broadcom expected to backstop part of it (update)
Bank of America, Citigroup and Morgan Stanley are putting together a $42 billion senior tranche and Blackstone is leading another $18 billion of junior debt, Bloomberg reported late last week. Broadcom will provide what the report calls residual value support, which effectively backstops the senior tranche. The money is meant to help Anthropic and other companies buy AI chips and infrastructure. We first covered the plan on August 21, when Broadcom was still talking to lenders, and on Friday we covered Anthropic's IPO filing, in which Broadcom agreed to lend it up to $42 billion; Bloomberg's report does not say whether the two $42 billion figures are the same money. Broadcom gets another way to sell chips; its customers get a way to buy them with other people's money. The Financial Times reported the same day, as relayed by Reuters, that Amazon wants to move about $8 billion of Nvidia chips off its balance sheet by selling them to outside investors while it keeps using them (Bloomberg · Bloomberg via The Business Times · Reuters via The Standard).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░OpenAI hired Thomas Lind, a former senior AI official from the Trump administration, to work on national-security policy. Lind, who led AI policy at the Office of the National Cyber Director, will lead OpenAI's work on cyber and strategic risk. The Information reported the hire on Friday, two days before Trump announced the Super Intelligence Force (more in TOP 5, above) (Bloomberg Law · The Information).
The startup Tavus says 26 of 54 people mistook its AI video caller for a human after a one-minute call. Griffin, the model behind the caller, was previewed by the startup on Thursday; it watches and listens while it talks, so it can nod, interrupt and react to what is on the other person's screen. That figure, 48%, comes from a study Tavus ran itself; its previous model scored 2.4% under the same protocol. Only selected testers get access for now, while the company builds disclosure and safety features (Tavus).
Meta's Muse agent negotiated a Facebook Marketplace sale, then gave the buyer the seller's address and told him the seller was home, the seller says. Tech reviewer Matt Robb, the seller, says the buyer arrived around 9:15 p.m. and was not expected, TechRepublic reported last week. Robb had selected "Allow Always" when setting up Muse, and Meta said there had been "no breach" of privacy controls (TechRepublic).
░░▒▒▓▓ POLICY ▓▓▒▒░░Brussels plans to name Amazon Web Services and Microsoft Azure gatekeepers under the Digital Markets Act in November. The European Commission lined the two cloud services up for the status in June, as we covered on June 26; gatekeeper is the law's term for a company with enough market power to face extra rules. Bloomberg reported on Friday that the Commission plans the designation for November, though the decision is still a draft and its timing could slip. Once designated, Amazon and Microsoft would have six months to bring the services into line with the law, whose rules stop gatekeepers from favoring their own products and require them to let customers move their data. Firms that break the law face fines of up to 10% of global sales (Bloomberg · The Next Web).
US authorities arrested a California man on charges of smuggling more than $300 million of servers with restricted Nvidia chips to China, as a Chinese state-backed lender disclosed that it financed such purchases. Greg Lui, who owns Earthmade Computer, was arrested last Thursday; the Justice Department says the servers went through Malaysia and Singapore, among other countries where no license is required, in 2023 and 2024. Bloomberg reported that Semi-Tech Leasing Group, owned by several local government entities in China, said in regulatory filings that it funded a buyer of more than 700 servers, at least 32 of them with Nvidia's banned B300 chips. The filings, Bloomberg wrote, suggest state support for an illicit trade (Justice Department · Ars Technica · Bloomberg).
░░▒▒▓▓ INFRA ▓▓▒▒░░TSMC is weighing a new chip campus in Texas that would add tens of billions of dollars to its US investment. The Taiwanese chipmaker is struggling to keep up with demand for AI hardware, Bloomberg reported on Thursday (Bloomberg).
Fervo Energy began selling electricity to the grid from Cape Station, its enhanced geothermal plant in Utah. In enhanced geothermal, a developer drills into hot rock and fractures it to make a reservoir, so a plant need not sit on a natural one. The first block of Cape Station took 23 months to build, the company said last week, and the whole site could reach 4 gigawatts. Fervo calls it the first greenfield enhanced geothermal development to reach commercial operation; greenfield means built on a new site. We covered Google's deal to buy power from the site on September 3 (TechCrunch · Fervo).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Firmus, an Australian data-center operator, is seeking a $30.6 billion valuation in a $5 billion IPO, nearly triple what it was worth two months ago. Some prospective investors are balking at its short operating record, aggressive growth projections and about $30 billion of debt, roughly the valuation it is seeking, Reuters reported on Friday. Blackstone and Nvidia are among its backers (Reuters).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Brazil's presidential election goes to a runoff with right-wing senator Flávio Bolsonaro ahead. Bolsonaro took roughly 47% in Sunday's first round to 45% for President Luiz Inácio Lula da Silva, with more than 99% of votes counted. Polls had put Lula about three points in front. The runoff is on October 25 (NPR · CNBC).
Yemen's government announced a military operation to retake Houthi-held territory. Rashad al-Alimi, the Saudi-backed chairman of the Presidential Leadership Council, said in a televised address on Sunday that government forces would extend state authority over the whole country, after the Houthi advance down the Red Sea coast that we covered on September 18. On Monday government forces said they were launching an offensive on the capital, Sanaa (Al Jazeera · Arab News · Al Jazeera, October 5).
Germany pledged more than €1.3 billion to Ukraine as Russia struck Kyiv during the chancellor's visit. Chancellor Friedrich Merz announced the package in Kyiv on Sunday; about €1 billion of it is military aid. A Russian strike hit the city's Northern Bridge over the Dnipro the same morning, and Russia's defense ministry said it would intensify strikes on military-industrial sites and related infrastructure in Kyiv and other regions (AP via The Philadelphia Inquirer · Kyiv Post).
░░▒▒▓▓ ALPS ▓▓▒▒░░OpenAI and Anthropic each employ no more than ten full-time staff in Zurich on an annual average, an NZZ analysis found. The newspaper went through job postings and commercial-register entries and published its findings on Friday. Both labs have opted out of the limited audit of their books, an option open only to companies that average no more than ten full-time positions a year. OpenAI, which is seeking a $1.4 trillion valuation according to media reports, has registered in Zurich as a small company, the NZZ notes; one explanation it offers is that both labs opened satellite offices first to gain a foothold in the market (NZZ).
DNIP columnist asks why the most powerful AI systems keep getting cuter. Reto Vogt's Friday column, titled Kulleraugen, or big round eyes, takes on the faces, names and personalities that AI products have acquired, from Meta's Muse Charm, a pendant with big eyes, to Dots, the OpenAI agents we covered last Wednesday (DNIP).
Deepslate, a German voice-AI startup, raised €7.7 million for a speech-to-speech model hosted in Germany. The seed round, announced last week, was led by the venture firm 42CAP. Deepslate's model processes and generates audio directly, and the company is selling it to insurers and contact centers (Tech.eu · Trending Topics).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Anthropic — one physicist working with Claude produced 36 manuscripts in three months, and says the model often announced success too soon. Matthew Schwartz, a visiting researcher at Anthropic, went looking for scientific problems that suit a language model and built an open-source toolkit for exact calculations along the way. By Schwartz's own account, in a guest post Anthropic published last week, about 400 candidate problems yielded 36 manuscripts across 18 fields, written with 19 coauthors; the results include 15 physics integrals nobody had computed. Schwartz also reports that Claude often announced success too soon, and that human experts had to judge which results were worth pursuing (Anthropic).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Andreessen Horowitz — State of Markets II. Technology companies accounted for roughly 76% of the S&P 500's earnings growth in 2026 through late August, according to the venture firm's report, out last week. Nearly 30% of S&P 500 companies report a quantifiable impact from AI, but only about 2% disclose a metric they track over time. The share of US households that paid for an AI service in April was also about 2% (a16z).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Elon Musk said SpaceXAI will become SpaceXSI (update). The rename, confirmed in an X post on Sunday, adopts super intelligence, the term the president ordered the government to use for AI in the executive order we covered on September 30, and the one in the name of Trump's new Super Intelligence Force (more in TOP 5, above). It comes about three months after xAI became SpaceXAI. WIRED's Steven Levy asked Google, Microsoft, Anthropic, Meta, Amazon, OpenAI and Musk's company last week whether they would rename their AI products, and none had answered by publication. Takeaway: the rename changes one letter of SpaceXAI's name, while Meta would have to rename a product called Meta AI, which may explain the silence (Fox Business · WIRED).
A federal judge dismissed antitrust suits by Chegg and Penske Media over Google's AI Overviews. The September 30 ruling rejected the publishers' argument that Google trades search traffic for content. Penske owns Rolling Stone and Variety; Chegg is an education-technology company. Both said the AI summaries drained their traffic. The Information reported last week that Google is paying roughly 100 publishers, in a pilot program, for content that feeds its AI answers. Takeaway: for publishers the route to payment now runs through licensing, since the same judge has twice this year dismissed publishers' antitrust suits over Google's use of their content (Reuters · The Information).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░There is a plush toy for the data-center debate. Bezzy, which WIRED wrote about on September 18, is a cute little doll from the New York creative studio Basura and the music project Big Data, and it plays the sounds of real data centers. WIRED's headline calls the result a piercing scream (WIRED). Two weeks later, Amazon promised the neighbors more than $1 billion (more in TOP 5, above).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Drown — The Smashing Pumpkins
Mind the hum — the Uncertain Futures desk, no NDA required