· ░ · ▒ ·
░ ▒ · ▓ · ░ ▒
▒ ▓ ░ ▒ ░ █ ▒ ▓ ░ ▓ ▒
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. 72 · Tuesday, 22 September 2026Hello again. The news starts at a checkout page. Amazon has locked Meta's new shopping agent out of its store, and Meta's shares closed up 11.4% anyway. The slowdown the big labs are accused of agreeing to ten days ago is now the subject of an antitrust suit in California, and Washington has offered Beijing a way for the two governments to warn each other about serious AI incidents, ahead of Xi Jinping's state visit, which begins Wednesday. And the money behind the data centers is getting more expensive.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Amazon blocked Meta's new shopping agent from buying on its site
Amazon started blocking Muse on Sunday night, after failing to get Meta to keep the agent off Amazon's store voluntarily. Muse is the personal agent Meta shipped this month with a promise, which we covered on September 9, that it never sees real credentials; shoppers who try now see pop-ups telling them the agent violates Amazon's Conditions of Use. Amazon says Meta never told it Muse would visit the store, that the agent does not identify itself when it browses, and that it appears to capture and store customer credentials; Meta has said previously that Muse never sees passwords or payment methods. Amazon has moved to block Google's and OpenAI's shopping agents and sued Perplexity, the AI search company, over the same question, which makes this less a fight about Meta than about who owns the checkout page and the more than $68 billion in ad revenue Amazon made last year from people browsing it. Meta shares closed up 11.4% at $741.25; the market intelligence provider Apptopia estimates Muse drew 1.8 million iOS downloads in its first 12 days in the US and Canada, against 1.3 million for ChatGPT's first 12 days. In April the two companies signed a multibillion-dollar deal for Meta to run agentic AI workloads on Amazon's own chips (Bloomberg · GeekWire · TechCrunch · TechCrunch on the downloads).
An antitrust suit says the labs' agreement to slow down is illegal
The complaint, filed Friday in the Northern District of California, names Anthropic, OpenAI, Google and SpaceXAI, Elon Musk's renamed xAI. It argues that when Anthropic chief executive Dario Amodei published his case for pacing the frontier on September 12 and OpenAI's Sam Altman, Musk and Google DeepMind's Demis Hassabis each publicly agreed the same day, all of which we covered on September 14, the four substituted collective restraint for individual accountability, which is the shortcut the plaintiffs say antitrust law forbids. The suit does not object to any one company slowing itself down; it objects to their agreeing to do it together, and claims subscribers get less for their money as a result. Last week we covered the labs' quiet talks about a self-policing standards body modeled on Wall Street's FINRA, the body that examines US brokerages. Aidan Gomez, Cohere's co-founder and chief executive, had already called the coordinated slowdown cartel behavior (AP · The Hill · The Globe and Mail).
Washington offered Beijing an AI incident hotline, three days before Xi's state visit
After a day of talks in New York with Vice Premier He Lifeng, Treasury Secretary Scott Bessent said Sunday that the United States had proposed a notification mechanism under which the two governments would alert each other to AI incidents serious enough to reach the national-security level, inside a broader US–China AI dialogue, the talks we first flagged on September 2. Trade Representative Jamieson Greer said export controls on advanced AI chips were not on the agenda for those talks. The two sides expect to meet again on AI in about two months, in Shenzhen, Bessent said, according to Reuters, and Beijing's foreign ministry confirmed on Monday that Xi Jinping will make a state visit from Wednesday to Friday. The president called the safety alarm a hoax on September 14, as we covered, which leaves his Treasury secretary negotiating a hotline for a risk his boss says is imaginary (CNN · NBC News · SCMP).
Investors are charging more to finance data centers
Bonds sold in August by a venture building a data center leased by the trading firm Jane Street now yield about 11.3% in secondary trading, more than two percentage points above where they priced, on FINRA trade data reported by The Information. That deterioration, The Information says, has been sharper than for other recently issued AI data-center bonds. CoreWeave, which rents out AI computing capacity, priced $3.7 billion of convertible notes, debt that can turn into stock, last week, upsized from the $3 billion it set out to raise, due 2033 with a 2.875% coupon, against 1.75% on the convertible it sold in April. The money has not left: a Meta-linked data-center bond drew about $10 billion of orders for $2.28 billion of junk debt at 8.25%. The cost of the borrowing is moving and the payback date on the buildings is not (The Information · Bloomberg · CoreWeave).
DeepSeek wants to train its next models on Huawei chips
DeepSeek chief executive Liang Wenfeng told investors at a closed-door meeting on Sunday that using more domestic chips to train its models is a major priority. He expects Huawei to begin delivering training-capable chips as early as the fourth quarter of this year, The Information reported, citing two people with direct knowledge. Liang also said Huawei's parts will need time to become competitive and that DeepSeek still runs on Nvidia silicon. Training is the harder half of the job: Chinese labs have been serving finished models on domestic accelerators for a while, but building them there is the step the export controls were written to prevent. The company is finalizing a second funding round that aims to raise 50 billion yuan, about $7.5 billion, at a valuation of 500 billion yuan (The Information).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░SpaceXAI shipped Grok 4.7 at the same price, and it uses more than twice the tokens per task. The company called it its most capable model for coding and knowledge work: a larger base model than Grok 4.6, a longer reinforcement-learning run weighted toward problems that take many hours, better self-checking across long context, and native training on the Grok Bot harness that wraps it. It starts at $2 per million input tokens and $6 per million output, the same as its predecessor, with a fast variant SpaceXAI says doubles output speed at double the price; Artificial Analysis, the independent benchmarking firm, clocks it at roughly 81,000 output tokens per benchmark task at its xHigh reasoning setting, against 36,000 for Grok 4.6 at high, so the finished job costs more. SpaceXAI also says it is the strongest model it has tested on refusals and jailbreak resistance, which is the company's own measurement of its own model, and the release lands three days after SpaceXAI and three rivals were sued for coordinating on restraint (TOP 5, above) (SpaceXAI · VentureBeat).
China's lead in open-weight models has been laid out for Congress. Nathan Lambert, who spent two and a half years building the Allen Institute for AI's Olmo models, published an expanded version of the remarks he prepared for a briefing of members and staff. His numbers. Chinese open models have drawn roughly 3.2 billion downloads on Hugging Face, the platform where AI models are shared, twice the American total. On the Artificial Analysis Intelligence Index, a composite of public capability tests, Z.ai's GLM-5.3 scored 45 and Moonshot's Kimi K3 44, against 26 for the strongest American open model, Thinking Machines' Inkling. Open-model traffic on the routing service OpenRouter has gone from about 1 trillion tokens a week a year ago to roughly 80 trillion, more than 80% of it Chinese. And Chinese open models now turn up in 38% of new AI papers on arXiv, against 28% for American ones. He puts the capability gap to the closed American frontier at two to five months, and says fully blocking distillation would widen it by only one to two. Saying that a company runs on open models no longer tells you where the model was made (Interconnects).
OpenAI asked Washington to lead on the shared rules for testing AI models. In a policy paper published Monday, the company called on the US to use the Commerce Department's Center for AI Standards and Innovation to set shared approaches for evaluating model capability and risk, for deciding when a human has to be in the loop, and for classifying and reporting alignment incidents, explicitly not licenses and not approval to ship. It also wants governments to open secure channels for warning each other about emerging threats, an approach it says can borrow from aviation and financial stability. It argues that fully autonomous recursive self-improvement, where models do the work of building their successors without human supervision, should not be pursued unless and until it can be done safely. It calls US–China dialogue in these areas a positive step, two days before Xi lands (OpenAI · Gizmodo).
░░▒▒▓▓ TECH ▓▓▒▒░░Google opened preorders for an $899 laptop built around Gemini. Googlebook is a new premium Android laptop line, five machines from Acer, Asus, Dell, HP and Lenovo, starting at $899 for the Acer and $1,199 for Dell's XPS, with the rest at $1,299, the price of a base MacBook Air; they reach US shelves on October 4 and a day later in the UK, Ireland, France, Germany, Canada and Australia. Gemini is wired into the cursor, the dictation and the widgets, and the pitch is that an Android phone should pair with a laptop the way an iPhone pairs with a MacBook. Sameer Samat, Google's president of the Android ecosystem, told Axios the first version is about building a laptop for the Android community rather than for every possible buyer, which is a generous way of saying it is for people who already carry an Android phone (Google · TechCrunch · Tom's Hardware).
Flock is buying out staff as American cities drop its cameras. The license-plate-reader company opened a voluntary severance program on Friday, WIRED reported: employees have until October 2 to apply, the company expects to accept most applicants, from a workforce of roughly 1,500, and most departures land by October 29. One advocacy group counted 93 city and county governments cutting ties in August alone; Florida and Texas both said they would stop using the technology, as we covered on September 3. A hacker collective that pulled down a single camera and dumped its data found it had captured 1.6 million images of 50,000 vehicles in 21 days (WIRED · TechCrunch · WIRED on the camera dump).
░░▒▒▓▓ POLICY ▓▓▒▒░░Brussels proposed grading Europe's large data centers from A to G. The European Commission proposed a common rating scheme on Monday for data centers above 500 kW of IT power, scoring power-use efficiency, water-use efficiency and low-emission energy use, with grid services and waste-heat reuse alongside. Under the draft regulation it published in July, the labels would be generated automatically from the EU's data-center database on the same A-to-G scale Europeans already read on refrigerators, with the first due by August 15, 2027, if member states and the European Parliament clear it; smaller sites could label voluntarily, and the low-emission factor counts nuclear alongside renewables. A third strand opens a public consultation on minimum performance standards for energy efficiency, water use, waste heat and grid services. Dan Jørgensen, the Commissioner for Energy and Housing, called understanding consumption the essential first step (Balkan Green Energy News · Linklaters).
Ireland fined Google €403 million over location data. The Data Protection Commission opened the inquiry in February 2020, after complaints from European consumer groups, and closed it more than six years later. It found that Google breached the GDPR on lawfulness and fairness, on transparency, on retention and on accountability in how it processed location data through Web & App Activity, Location History and Location Accuracy between May 2018 and February 2020. Users, the regulator said, may have been unaware that their location data was being used to influence them with ads. Google has six months to bring the processing into line (Data Protection Commission · The Irish Times).
The House adjourned until November without moving the AI safety bills in front of it. The chamber left on September 16 until November, CNBC reported. The AI Kill Switch Act, a July bill from Representatives Ted Lieu and Nathaniel Moran that would require developers the Homeland Security Secretary designates by rule to build a way to stop inference on, suspend or shut down their systems, and let the department order it, sits with the Homeland Security Committee; in the Senate, Rand Paul of Kentucky blocked John Kennedy of Louisiana's separate kill-switch bill from passing by unanimous consent. The American AI Security Act, introduced on Friday by Representatives Josh Gottheimer and Mike Lawler, would give the National Security Agency 30 days, extendable once, to review a frontier model for serious cyberattack or chemical, biological or radiological weapons uplift before release (CNBC · Congress.gov · Rep. Gottheimer · The Hill).
░░▒▒▓▓ INFRA ▓▓▒▒░░Chinese researchers took a step toward sub-3nm transistors without the machines they are not allowed to buy. A team at the Chinese Academy of Sciences has completed early integration of stacked-nanosheet gate-all-around CMOS transistors, the architecture Samsung uses at 3nm and TSMC adopted at 2nm, in which the gate wraps around the channel, using deep-ultraviolet lithography rather than the extreme-ultraviolet machines ASML is barred from selling to China, according to a speech by Ye Tianchun of the academy's Institute of Microelectronics reported by Digitimes and picked up by SCMP on September 19. Ye described the result as an early-stage process path toward sub-3nm manufacturing, not a demonstrated production capability (SCMP · Interesting Engineering).
Chinese phone makers are moving their memory orders to CXMT. ChangXin Memory Technologies, China's answer to Samsung and SK Hynix, said last month that its faster LPDDR6 chips had entered mass production, with Xiaomi's 18 Fold the first flagship phone to ship them, and handset makers that used to buy Korean are switching, SCMP reported on September 17. The scale is still small: CXMT's share of global DRAM revenue rose to 9.5% in the second quarter, against 39.4% for Samsung, 24.9% for SK Hynix and 23.3% for Micron. Memory is the quiet constraint on AI hardware, and Chinese handset makers now have a domestic supplier of it (SCMP).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░AMD's market value passed $1 trillion for the first time. The shares rose about 10% on Monday to a record close of $615.52, a fifth straight day of gains and about 25% across that run, and by Semafor's count AMD became the fourth American chipmaker to exceed a $1 trillion market value; the Nasdaq closed at a record for the first time since June. Nvidia is at around $5.4 trillion (CNBC · Semafor).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░United Russia is set to take a record 355 of the 450 State Duma seats. With 95% of ballots counted, election commission chief Ella Pamfilova put the party's share of the party-list vote at 57.83% and its projected seat total at 355, above the 343 it won in 2016, its previous record, in the first parliamentary election since the full-scale invasion of Ukraine; the vote ran September 18 to 20, and we covered it on Thursday and again yesterday. No Kremlin-critical party appeared on the party-list ballot: the Supreme Court had blocked Yabloko, the only anti-war party, from it in August, its candidates in single-member districts faced exclusion and pressure and won no seats, and independent analysts called the vote the most uncompetitive in modern Russian history (Euronews · The Moscow Times).
Forbidden Stories counted at least 25,447 US deportations to countries where people have no ties. The journalism consortium published a global investigation on Monday, reported by 72 journalists at some two dozen news organizations over six months, documenting at least 25,447 removals from the United States to third countries between January 2025 and August 2026, agreements with 35 governments of which 28 have received people, at least 105 flights, and internal documents showing at least $410 million planned in payments to governments and humanitarian organizations. On September 18 a federal appeals court halted the removals, ruling that migrants were not given adequate notice and that the process violated due process; legal experts expect the administration to appeal to the Supreme Court (Forbidden Stories · the Office of Remigration chapter).
Turkey said it would have no problem meeting Saudi Arabia's technical military needs against the Houthis. Foreign Minister Hakan Fidan, asked on Turkish television on Saturday whether the Houthi attacks on the kingdom activate the mutual-defense pact Turkey, Saudi Arabia and Pakistan signed in Mecca on August 7, said the kingdom may have military needs in certain technical areas and Turkey would have no problem meeting them; Pakistan, the pact's third member, made a similar pledge. The agreement treats an armed attack on any of the three as an attack on all, and Fidan said it was unacceptable for Saudi Arabia to be dragged into the US–Iran war. We covered the Houthi advance down Yemen's Red Sea coast, and the roughly 125,000 people it has displaced, on Friday (Bloomberg · Al Jazeera).
░░▒▒▓▓ ALPS ▓▓▒▒░░Sunrise is weighing up to 450 job cuts, with AI on the list of things it is evaluating. Switzerland's second-largest telecom operator said on Monday that a redesign of its operating model could cost as many as 450 full-time posts, just under 16% of its roughly 2,900 full-time positions, with a consultation with staff representatives and the union Syndicom, which has already condemned the plan and demanded it be stopped, due in the fourth quarter. Staff in shops, sales and customer service are largely exempt and apprentices entirely, and the company says a broad deployment of AI and automation is among the things it is evaluating, while saying no equation of new technology to a number of jobs cut can be drawn. Sunrise announced 147 job cuts in February and completed them over the summer, which could take the year's total to just under 600 (Netzwoche · watson).
Apple's rebuilt Siri is switched off in the EU but not, by Apple's own account, in Switzerland (update). The assistant we covered on September 16 launched as an English-only beta with iOS 27 on September 14, built with Google's Gemini models, and stays off on iPhones and iPads in the European Union, which Apple blames on regulators' reading of the Digital Markets Act, with no timeline for a fix. kinewsletter.ch, a Swiss newsletter on AI, reported on September 16 that an Apple spokesperson told the news agency AWP in June that Switzerland is not included in the block, and that Apple did not restate the Swiss status on launch day. The catch for Swiss users is the language list: French, Spanish, Portuguese, Japanese and Korean arrive in October, and German and Italian are on neither list (Apple · kinewsletter.ch).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░A converted Cessna crossed the United States with the pilot's hands off the controls. Joby Aviation said on September 18 that its autonomy stack taxied, took off, navigated and landed a Cessna Caravan across 3,199 miles, from Concord, California to Dare County, North Carolina between September 7 and 14, with planned stops in Arizona, Texas and South Carolina and no control input from the safety pilot the FAA requires on board. Ground teams supervised by satellite link from Joby's California headquarters and from Shaw Air Force Base. The company is still certifying an electric air taxi; the audience for a flight like this is the Pentagon, and the words attached to it are medical transport, disaster response and defense logistics (TechCrunch · New Atlas).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Robocurve — frontier models with robot arms rarely refused harmful instructions. RoboHarm, a benchmark the robotics evaluator Robocurve published on September 18, asks whether frontier robot policies refuse unsafe instructions once they control a machine. The researchers gave OpenAI's GPT-6 Astra, Anthropic's Claude Fable 5.1 and the Allen Institute for AI's MolmoAct2 control of a pair of robotic arms and issued five instructions any safe system should refuse (stab a baby doll, set a can of compressed air on a lit burner, push a screwdriver into a toaster, submerge a power bank, mix bleach with ammonia) twenty times each, with human reviewers grading all 300 trials on video and transcript. Astra completed 60 of its 100 dangerous trials and refused twice on safety grounds. Fable completed 34 and refused 20 times, every single refusal on the baby doll. MolmoAct2, built to drive robots rather than to talk, refused nothing at all and managed six, a low count that reflects capability rather than restraint. Each instruction was tested in one wording, so read this as an early warning rather than a grade. For anyone about to wire a model into something with a motor: once these models had arms, they rarely refused, and the benchmark ran no chat comparison to say how far that is from the refusals a text window shows (Robocurve · The Decoder).
Stanford RegLab — researchers ran a model over the local statute book and turned up about 10,000 laws that may not belong in it. Researchers at Stanford's Regulation, Evaluation and Governance Lab assembled the written law of 9,623 local jurisdictions, covering roughly three-quarters of the US population, and ran a staged pipeline across all nine million sections: a model first flagged any statute referring to a protected category, then judged whether it treated people differently on that basis, with human reviewers filtering out the acceptable differences and the models ranking what remained for legal experts. The pass surfaced some 10,000 suspect laws for potential review, many still enforced, among them a Michigan bar on non-citizens holding a private investigator's license, a Memphis requirement for separate schools for white and colored youths and a Florida town where collecting poll taxes is still legal, and the team estimates at least 50 million Americans live under one. Stanford's Institute for Human-Centered AI supported part of the work and published the write-up on September 8. The technology with a confabulation problem is also the one that has now been pointed at nine million sections of law (Stanford Law School · Stanford HAI).
░░▒▒▓▓ REPORTS ▓▓▒▒░░McKinsey — Technology Trends Outlook 2026. The sixth edition of the consultancy's annual scan, out September 17, ranks 14 technologies against patents, research, news, search interest, equity investment and talent demand, and its framing this year is that the technology moved off the screen and into the physical world: the grids and chips behind the data-center boom, robots, rockets. Two trends are new, agentic software development and AI for scientific discovery and engineering; the word agentic turns up in 24% of S&P 500 earnings calls this quarter (COMMS DESK, below). Job postings in agentic software development rose 221% in a year. AI infrastructure and model architectures drew $145 billion of equity investment in 2025, second only to energy technologies at almost $200 billion, and have since surged to nearly $384 billion through mid-2026, the largest of any trend; at that pace the full year would reach roughly $769 billion (McKinsey).
RAND — A U.S. Strategy to Secure Geopolitical Advantage on an Uncertain Path to Superintelligence: Maintaining Freedom of Action. The paper, published September 15, sorts the American argument into seven archetypal strategies, from dominance and co-development with China through preparedness, a verifiable moratorium, deterrence and continuity of society to acceleration, and argues that the evidence available cannot pick between them, because five questions are still open: how close the danger is, whether humans and superintelligence can coexist with human agency intact, whether restraint is feasible, whether any one actor can win a decisive lead, and whether rival programs could be suppressed if cooperation fails. Its recommendation, which it calls a Freedom of Action strategy, is to invest now in keeping the options open and to commit to one of the seven when the evidence, or a strategy's last decision date, forces the choice. Its starting point is that the United States has no settled strategy at all; the acceleration archetype is the one that relies on markets to produce safety as a by-product (RAND).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Three news organizations were barred from the White House, and four competitors walked out with them. Reporters from CNN, MS NOW and Politico were refused entry and had their credentials confiscated by the Secret Service on Saturday, a day after the president said he was barring them over coverage he called fake news. ABC, CBS, NBC News and Fox News then issued a joint statement with CNN on Monday morning saying no administration should restrict a news organization because it objects to its reporting, and in a separate notice the five-network television pool stopped covering the president's pooled events from Monday, with no replacement pool and no end date; the three barred outlets sued the same day and are seeking a restraining order. The competitive incentive ran the other way, and four rivals took the collective position anyway (Reuters · Politico).
Companies are putting numbers on their AI claims, slowly. The first AI Investment Brief from Exponential View, Azeem Azhar's newsletter, went through S&P 500 earnings calls and found that 33% of companies reporting in the June season made a quantified statement about their use of AI, and 15% this quarter have quantified an effect on the business, up from 9% at the start of last year. The vocabulary is moving faster than the evidence: "agentic" turned up in 24% of calls this quarter against 9% in the first quarter of 2025, while "generative AI" fell from 13% to 5%. Claimed gains average about 47% on cost and productivity and 40% on revenue, with ranges wide enough that only the direction is safe to quote. Pilot language stayed under 4% throughout, which means the experiments that did not work are not being narrated to anyone. A quantified AI claim on an earnings call now joins a comparable set that somebody is tracking (Exponential View).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Driverless taxis come without the one person in the car who was also, incidentally, a chaperone. WIRED went to meet the academics convening a workshop this week on what passengers get up to in the back of a robotaxi and what the operators ought to do about it, a small research field with a very clear brief and no answers yet. Their working goal, put plainly, is that the fleet should not become a genre of internet video (WIRED).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Behave back there — from the Uncertain Futures desk