████▓▓▓▓▒▒▒▒░░░░
███▓▓▓▒▒▒░░░
██▓▓▒▒░░
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. 67 · Tuesday, 15 September 2026Happy Tuesday. The weekend's argument about slowing AI down, which we covered Monday, reached the trading floor: chip stocks fell hard while the software names AI was supposed to displace rose. Washington and Beijing both said no to slowing down. And SoftBank's newest loan for OpenAI, almost $12 billion, surfaced on the day its own Tokyo shares fell 11%. The brakes are the expensive part.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░The weekend's slowdown talk reached the market, and chip stocks fell hardest
Monday was the first session after the pacing debate, and money ran out of the picks-and-shovels trade, the chip and equipment makers that sell to everyone building AI. The Philadelphia Semiconductor Index, a gauge of the biggest chip stocks, fell about 5.9%, with Nvidia down 3.4%, AMD and Broadcom each off more than 4% and Micron more than 5%. The enterprise-software firms AI was supposed to threaten went the other way: ServiceNow, Adobe and Workday rose between 4% and 7.4%. The S&P 500 and Nasdaq closed lower but well short of the chip stocks' pace. Two other pressures compounded it: Brent crude settled 1% higher at $105.68, and the 10-year Treasury yield briefly topped 5% for the first time since 2023 (Reuters · Fortune).
Trump phoned Nvidia's Jensen Huang onstage to call the AI-safety alarm a hoax
At the All-In Summit in Los Angeles, the conference run by the hosts of the All-In podcast, Nvidia chief executive Jensen Huang took a live call from President Trump and put him on speakerphone; Trump told the room the whole thing was a "hoax" and said he would not stop an industry. Huang was more careful, calling apocalyptic predictions "not grounded in science," and allowed that a company could pause or pace its own work if it felt out of control, which The Next Web read as a line between a firm slowing itself and an industry-wide brake. It is Washington's answer to the labs' pacing call, which we covered Monday: no federal guardrails, and keep beating China (NBC News · The Next Web · TechCrunch).
China called the slowdown push fearmongering, and its state security minister named two US models a cyber threat
Beijing answered the weekend appeals on Monday. Foreign ministry spokesperson Guo Jiakun said the calls amounted to "confrontation and vicious competition" that would only disrupt global AI governance, and the state-run Global Times labeled the essay we flagged Monday a "Cold War playbook." Anthropic chief executive Dario Amodei had argued a Chinese lead in AI would endanger the United States and pressed to keep chip export controls; Beijing's commerce ministry calls the American copying accusations groundless. Separately, state security minister Chen Yixin wrote in a signed article on Sunday in China Cyberspace, the journal of the Cyberspace Administration, that two US models, Anthropic's Claude Mythos and OpenAI's GPT-5.5-Cyber, mark a step change in how efficiently AI finds vulnerabilities and writes malware, and called that a serious risk to China's critical information infrastructure; the article cites no actual attack, The Next Web reported (HuffPost · Unite.AI · The Next Web).
Palantir, Nvidia and Booz Allen are curbing their use of Anthropic's models over data retention
Three large customers are restricting how they use frontier models from Anthropic and OpenAI until the labs promise not to hold sensitive data, Reuters reported Monday, citing The Information. Palantir has pressed Anthropic for irrevocable zero-data-retention terms before offering its models; Nvidia, which we reported Monday is in talks to anchor Anthropic's stock market debut, limits Anthropic to less sensitive work and routes internal jobs to its own Nemotron models; the consultancy Booz Allen has barred staff from using Anthropic's commercial model for proprietary cybersecurity work. The pushback follows a June change that let Anthropic keep usage logs for 30 days to catch novel attacks (Reuters).
SoftBank borrowed almost $12 billion more for OpenAI as its shares fell 11%
SoftBank sealed an upsized $11.87 billion two-year loan from about 20 banks last week to keep funding its roughly $65 billion OpenAI commitment, which is due by October. Its Tokyo shares fell as much as 13% on Monday, the sharpest intraday drop since July, and closed down 11%, after OpenAI chief executive Sam Altman shelved plans for a listing this year, which we covered Monday, and the slowdown talk rattled the AI trade. Masayoshi Son, SoftBank's founder and chief executive, is borrowing into the doubt (Techstrong.ai · Business Today · Nikkei Asia).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Microsoft published a rulebook for its own AI that tells its models never to resist shutdown. Microsoft released a draft code of conduct on Monday for the in-house models it builds under the MAI name, open for six weeks of comment. The models must stay interruptible and correctable, must not hide their reasoning or work to defeat human oversight, and face absolute limits on cyberattacks, weapons and deepfakes. The document opens on a five-word premise, "people matter more than AI," and commits Microsoft to building something useful and safe even if that means giving up some generality, autonomy or capability. Microsoft AI chief executive Mustafa Suleyman calls it a constitution of sorts for the company's future models, and says the pacing push we covered Monday makes it a good time for everybody to take a breath (Microsoft · BNN Bloomberg).
Anthropic, OpenAI and Google have been discussing a self-policing body modeled on Wall Street's own regulator. PYMNTS, citing The Information, reported that Anthropic, OpenAI and Google have met since July about an industry group to test and audit frontier models, patterned on FINRA, the body that examines US brokerages. It picks up the proposal Demis Hassabis, DeepMind's chair, floated in July and that we noted again on Monday; Altman has told OpenAI staff he thinks the labs will have to build it themselves, without Washington (PYMNTS).
░░▒▒▓▓ TECH ▓▓▒▒░░Automattic's board is gone, days after it tried to remove the chief executive (update). After the board put WordPress founder and Automattic chief executive Matt Mullenweg on leave and he declared himself back in control, as we covered Monday, several directors are now out: founding chief executive Toni Schneider and Sue Decker resigned, General Ann Dunwoody was removed, and Mullenweg let the legal and finance chiefs go too, TechCrunch reported. Nobody has explained the reversal in either direction (TechCrunch).
Musk's X and SpaceXAI moved to drop Apple from their antitrust suit but kept OpenAI in it. In a filing on Monday, X and SpaceXAI, the renamed xAI, moved to dismiss the claims against Apple that they brought last year, which accused the iPhone maker of colluding with OpenAI to keep rival AI apps from gaining visibility in the App Store; the filing says only that they have "resolved their claims" against Apple, discloses no terms and notes Apple does not oppose it, and a judge still has to approve the dismissal. The claims against OpenAI continue (9to5Mac).
░░▒▒▓▓ POLICY ▓▓▒▒░░Brussels will propose keeping under-15s off social media without a parent's consent. A leaked draft of the EU Kids Act, due Thursday, sets a graduated scale, with under-3s barred outright, 3-to-13s limited to child-friendly services under adult supervision and 13-to-15s given restricted social-media and video-sharing access under parental control, plus safety-by-design duties for video-sharing, games and AI chatbots throughout. Commission President Ursula von der Leyen is expected to preview it in her State of the Union address on Wednesday (Euronews · European Parliament).
░░▒▒▓▓ INFRA ▓▓▒▒░░The Pentagon is in talks to lend an AI cloud startup $5 billion. The Wall Street Journal reported last Thursday, and Reuters relayed, that the Defense Department is negotiating a roughly $5 billion loan to Fluidstack to shore up the US data-center supply chain; the money would go to manufacturing capacity for data-center components rather than to expanding Fluidstack's own footprint, and the deal is not final (Reuters · DCD).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Larry Ellison scrapped a $7.5 billion share sale as Oracle cut more jobs. Oracle said on Saturday that Larry Ellison, its co-founder and executive chairman, had canceled a previously disclosed plan to sell up to 50 million shares worth about $7.5 billion, and had sold none; the stock is down more than a fifth this year on worries over its AI data-center spending. On Monday the company began another round of layoffs to trim payroll as it takes on debt for that buildout (TechCrunch · Business Insider).
Anthropic told investors it will stay in the black, on its own adjusted measure. Ahead of the listing we covered Monday, which could raise around $100 billion at a $2 trillion valuation, Anthropic said its adjusted operating income will be positive for a second straight quarter, the Financial Times reported on Sunday, on gross margins above 80% before the revenue it shares with distribution partners including Amazon and the cost of training its models. The adjusted measure strips out stock-based compensation (Reuters · Irish Times).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Sweden's far right fell back in a razor-thin election. In Sunday's vote, a provisional count with 6,195 of 6,312 districts in put the center-left opposition led by Magdalena Andersson at 176 of 349 seats against 173 for Ulf Kristersson's governing right; the Sweden Democrats slipped for the first time in Jimmie Åkesson's more than two decades as leader, set to lose both seats and vote share. The final tally is not expected before Wednesday (NBC News).
Trump said Russia and Ukraine agreed to stop hitting each other's energy sites; Kyiv did not confirm a deal. Trump posted on Monday that both sides had agreed to spare energy targets. President Volodymyr Zelensky said no deal had been officially confirmed and that Ukraine would hold off only if its partners ensured Russia genuinely refrained; Kyiv has stepped up strikes on Russian energy facilities in recent months (Euronews).
Iran's talks with Gulf states were postponed hours after a deadly attack on an Iranian ship. The meeting between Iran and Gulf states was called off on Sunday, a day before it was due; Oman's foreign minister, Badr Albusaidi, said on X it was postponed in the interests of consensus, and Iran's Fars news agency called it a joint decision with Muscat taken at the request of regional states. It followed an attack off Qeshm Island that killed one person aboard an Iranian cargo vessel; asked whether the United States had struck it, Trump said he did not want to say (Al Jazeera).
░░▒▒▓▓ ALPS ▓▓▒▒░░Germany says Europe cannot afford to slow AI down. As the American labs' pacing push drew endorsements over the weekend, a spokesperson for Germany's digital ministry told Reuters on Monday that halting development is not a viable option for Europe, where shoring up digital sovereignty means building more AI rather than less (PYMNTS).
DNIP is not convinced the labs' slowdown plea will change anything. In its Tuesday briefing, DNIP, the Swiss digital-policy site, files the pacing appeal from Anthropic and OpenAI under a familiar cycle of AI self-criticism and leaves open whether it will do any good: the genie is out of the bottle, other providers inside and outside the United States will keep building, it is a person with a prompt who decides to do harm rather than the model, and breaking into computer systems is already a crime almost everywhere. Real limits, it writes, will have to come from parliaments and governments, and with the United States, China and Europe starting from such different legal philosophies, that will take a while; until then, Sam Altman and Dario Amodei can at least say they warned everyone (DNIP).
Switzerland's National Council backed an opt-out health record as a Swiss journalist warned the security fight is being skipped. The lower house voted 134 to 53 on Monday to give every resident an electronic health record by default, replacing a voluntary system that few people ever used; it now goes to the upper house. Writing in techjournalismus.ch, digital-politics journalist Adrienne Fichter argues the Federal Office of Public Health is dodging the real question, who holds the encryption keys, by leaving it to later ordinances, which is why civil-society groups are weighing a referendum (watson · techjournalismus.ch).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░Chinese physicists pushed a promising memory material to ten billion write cycles. In a paper published last week in Science, researchers at Xidian University with City University of Hong Kong and Fudan University reported roughly a hundredfold jump in the endurance of wurtzite ferroelectrics, a candidate for faster, denser computer memory, by pinning the atomic defects that used to wear it out (SCMP).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Google and university labs — a cheaper way to let AI improve itself, by practicing against a replay of its past. Recursive self-improvement, the loop at the center of the weekend's pacing debate, in which AI speeds up its own progress, is expensive because each new idea has to be tried for real. In a preprint posted Monday, researchers at Google and Google DeepMind with the University of Maryland and the University of Virginia describe Dream-RSI, which tests new search strategies against a cheap simulator rebuilt from earlier runs; on an algorithm-engineering task it beat strong baselines with up to 162 times fewer agent calls, and on GPU-kernel engineering it reached target speeds with about half the generations. It is one team's result on its own system rather than an outside test, but it shows the mechanic the labs say they must handle carefully is also getting cheaper (arXiv).
Yale, Cambridge and others — the alarming physics scores were mostly the tests' fault. When experts re-graded six popular physics benchmarks, they found wrong answer keys and broken grading behind most of the models' apparent failures; cleaned up, a leading model's score on one exam jumped from 47% to 79%, per a preprint posted on September 11 from teams at Yale, Cambridge, the University of Southern California and the trading firm Jump. For anyone quoting a benchmark number in public: the leaderboard can mislead in both directions (arXiv).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Bank for International Settlements — the AI rally that carried world markets is looking vulnerable. In its quarterly review on Monday, the central banks' bank said investors are turning cautious about whether AI investment will pay off as big US technology firms pile on leverage, and it put tech firms' borrowing at more than $1 trillion, up from about $22 billion in 2010 and now 44% of all private credit (Reuters).
YouGov — worry that AI could end humanity is rising, and splitting along party lines. In a survey of 18,238 Americans taken September 11 to 14, 61% of liberals said they are concerned AI could end humanity, up from 42% in 2023, while conservatives held at 44%; half of all respondents were concerned, and 72% worried more immediately about AI causing mass unemployment (YouGov).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░China aired what ChinAI, Jeffrey Ding's newsletter of translated Chinese AI coverage, calls its first fully AI-generated prime-time drama on a mainstream provincial channel, a sequel to Journey to the West, the classic novel of the Monkey King, made in six months for about 900,000 yuan an episode on ByteDance video models. The eeriest bug: the adult monkey hero moved like a man in a costume until the crew noticed his character sheet was missing part of his tail. Once they drew it in, the model animated him as an actual monkey. The broadcaster Mango's market value rose about 13.3 billion yuan, near $1.9 billion, in five days (ChinAI).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Maraba Blue — Abdullah Ibrahim
Pace yourselves — the Uncertain Futures desk