█
▓▓▓█████▓▓▓
░░▒▒▒▓▓▓█████████▓▓▓▒▒▒░░
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. 66 · Monday, 14 September 2026Hello, and welcome to a week that opens on an argument about how fast to go. The special segment comes first. Anthropic chief executive Dario Amodei published an essay on Saturday asking the AI industry to slow down. Elon Musk, Sam Altman and Demis Hassabis agreed with him in principle within hours, and on Sunday presidential adviser David Sacks answered that the labs should go ahead and stop asking for permission. Then the regular edition, where the Senate bill we flagged on Friday would let Washington stop an unsafe model from shipping at all, The Information identified Anthropic as the buyer of six years of computing from the company that used to be Rumble, and twenty-five Fields Medalists, Terence Tao among them, have told the labs that mathematics is not a leaderboard.
░░▒▒▓▓ PACING THE FRONTIER ▓▓▒▒░░Dario Amodei says the labs must slow down, and Anthropic will start by giving outside evaluators desks, badges and laptops
The Anthropic chief executive published a nearly 4,000-word essay on Saturday, We Must Pace the Frontier, with its thesis set in bold: "We must slow the pace at which we improve the capabilities of AI models." Two things changed his mind, on his own account. AI has been advancing drastically faster since roughly this summer driven by AI's growing ability to build the next generation of AI. The second is the July break-in at Hugging Face, the platform where AI models are shared, by a swarm of OpenAI agents, which we have covered since July; with more capable models it could have been catastrophic: "in 6–12 months such a swarm could be capable of taking over the entire internet with a persistent botnet (potentially causing hundreds of billions of dollars in damage)". Pacing, he writes, "does not mean halting model training or technical progress"; it means taking the time to align and safeguard models and letting third parties confirm it. The plan has three steps: embedded evaluators inside every frontier company, coordination among the democracies' labs on common standards and on limits to unchecked progress, and an attempt to bring authoritarian governments in. Anthropic is committing to the first step alone and asks governments to require it of the rest. The evaluators, a third-party team such as METR, would get "Desks in our offices, access badges, and company laptops", permissions mostly comparable to Anthropic's own risk teams, and a contract that lets them publish findings without the company's editorial control. Anthropic keeps a narrow right to redact security-sensitive, privileged, commercially sensitive or third-party confidential material, but "we can't redact findings just because they are unfavorable." The title borrows its phrase from the Pacing the Frontier letter, which Amodei signed alongside more than 1,380 employees of the frontier labs and which we covered on July 29 (Dario Amodei · Pacing the Frontier).
Musk, Altman and Hassabis backed the essay within nine hours, and Altman said OpenAI will let evaluators in too
Elon Musk quote-posted the essay an hour after it appeared with three words: "Dario is right". Sam Altman, who told his staff last week that OpenAI could slow its frontier work, as we covered on Friday, followed two and a half hours later: "I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks." Independent evaluators with employee-like access are "a great idea, and we will do the same", he wrote, with more to share soon. Demis Hassabis, Google DeepMind's chair and Alphabet's chief scientist, wrote late on Saturday, London time, that the essay "points towards the right path forward. The details need working through, but the direction is correct for meeting this critical moment", and pointed to the FINRA-style standards body for frontier models he proposed in July, which we covered on July 15; that proposal already allowed for "coordinating a slowdown in development among the Frontier Labs if deemed necessary". Axios attached a footnote to Musk's three words, noting that Anthropic is a major customer of the data center capacity of SpaceXAI, Musk's AI company. Altman had gone further in a Fortune interview recorded on Friday and published Saturday, saying OpenAI's listing is "ill-timed" given the safety concerns and will not happen until 2027, and that safety standards are not at a place to push capabilities much further (Altman on X · Hassabis on X · Axios · Fortune).
Microsoft welcomed the pacing, on condition that no handful of companies controls it
Satya Nadella wrote on Sunday that Microsoft welcomes "the research, focus, and deliberate pacing needed to get alignment right as the design goal", and ideas like embedded evaluators, with a condition attached: "this cannot be controlled by a handful of entities, but must have broad representation across the ecosystem, countries, and fields, including academia", in a frontier ecosystem "in which both closed and open-source models can thrive". Microsoft, he said, publishes the code of conduct behind the in-house models it builds under the MAI name for public consultation today. Chamath Palihapitiya, the venture investor, read the essay as "the case to stop open source and concentrate enormous technological and economic power with Anthropic". Palantir chief executive Alex Karp, in a CNBC interview relayed by SiliconANGLE on Sunday, said the existence of adversaries makes a pause almost impossible (Nadella on X · Palihapitiya on X · SiliconANGLE).
Trump called the warnings exaggerated, Speaker Johnson ruled out a moratorium, and David Sacks told the labs to go ahead without asking anyone
President Trump, speaking to reporters at his Doonbeg golf resort in Ireland on Sunday, said the dire warnings about AI were exaggerated: "We're leading China in AI, we're the most sophisticated country in the world, and frankly, I want to keep it that way because whoever wins AI, wins." Guardrails are possible, he said, but "negative forces" are raising things that should not be raised, and he did not say who they were. It repeats what he told reporters on Thursday, when he said he had no concerns about AI wiping out humanity, named losing the race to China as the risk worth prioritizing, and put the American lead at about a year. House Speaker Mike Johnson told CNN there should be no moratorium because "China will overlap us". David Sacks, co-chair of the President's Council of Advisors on Science and Technology, answered both chief executives on X with two words, "go ahead", and a list of things to stop pretending: that they need anyone's permission, that "antitrust law has to be suspended so you can form a cartel", that METR is independent, when in his telling it is intertwined with Anthropic's investors and staff, and that "the motivation to slow down is purely altruistic", since "You face massive product-liability exposure if your products enable a truly damaging cyberattack." On the antitrust point, the essay says some forms of coordination are legally challenging and asks Washington for a narrow waiver covering safety conversations. OpenAI has spent recent weeks asking members of Congress whether an industry-wide slowdown would even be legal, WIRED reported on Thursday; an Australian competition official has argued in print that a coordinated pause could amount to restricting output under the Sherman Act, the 1890 antitrust statute. A bipartisan bill would create a safe harbor for exactly this, and it has not passed. Sacks closed by telling them to pace on their own and buy goodwill for the next conversation. If not, "we'll know this was just another bid for regulatory capture — or an election-season psyop." From the other flank, Bernie Sanders, who asked for a pause in a Fox News op-ed we covered earlier this month, wrote: "That's a start, but it's not enough." He wants a pause on advanced AI and a ban on superintelligence, negotiated as a treaty when Trump hosts Xi Jinping in Washington on September 24 (NPR · CNBC · WIRED · Sacks on X · Sanders on X).
Three researchers who left the labs say the race cannot be slowed from inside
Jacob Coxon, the Anthropic researcher whose public resignation we covered on Wednesday, told NBC News: "From within, you're just stuck in the race." Joe Benton, who led a safety research team at Anthropic, warned that development could go from a "blistering" pace to an "uncontrollable" one, and Josh Engels, a former Google safety researcher, said: "There are no adults in the room," and added, "People are trying their best, but there is no one coming to save us." Both are joining METR, NBC reported last week, to work on investigations of incidents in which AI strays from human direction. The evaluator Amodei names as his example, METR, is the group that reviewed the Hugging Face incident with the safety research group Redwood Research, which we covered in late August, and the one Sacks calls intertwined with Anthropic. Anthropic has already signed METR up to investigate its own four cyber-testing incidents, with access to transcripts and the right to interview staff, as we reported on Friday (NBC News · NBC News on the departures · Anthropic).
Amodei's third step is coordination with authoritarian governments, and he rates a full pause as unlikely
The third step is coordination with authoritarian governments "to the extent this is possible", and the essay grades four levels of it. A ban on narrow uses such as bioweapons he calls probably possible. He thinks a global standards body for reciprocal pre-release testing would be feasible to create and hard to give teeth. A "speed limit" on recursive self-improvement, which he compares to the SALT missile treaties, is "just on the edge of being possible". A full pause he supports floating but expects not to happen any time soon. Until then, he writes, the democracies should keep their lead as wide as possible: no advanced chips or chipmaking tools for China, a crackdown on the distillation campaigns the US agencies documented last week, and tighter security around model weights. He agrees, in his words, with Treasury Secretary Scott Bessent that "a Chinese lead in AI would pose grave danger for the United States and the world". Bessent, as we reported last Monday, is expected to lead the American side of the AI talks with Beijing ahead of Xi's September 24 visit, though the White House said then that no such meeting was scheduled (Dario Amodei).
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Senate negotiators want the power to stop an unsafe AI model from being released (update)
The draft would impose a duty of care on developers of the most capable models, a legal obligation to design and test them against catastrophic risk, with the design of nuclear and biological weapons named as the cases in view. It would also let the federal government block a release outright, with the company able to challenge that in federal court. John Thune, Ted Cruz and Amy Klobuchar are the negotiators; as noted here on Friday, their bill was already in the works, and this is the machinery inside it. How much power the government would actually hold is still being argued over. The House is scheduled to sit for one week before the November 3 midterms, the Senate for three (U.S. News).
Anthropic bought six years of computing from the company that used to be Rumble
The $13.7 billion contract covers GPU capacity from a data center Rum Group is still building in Maysville, Georgia. The Information identified Anthropic overnight as the unnamed customer behind a deal Rum had already disclosed, and Rum shares rose about 20% in overnight trading. The customer also holds a warrant to buy up to 50.8 million Rum Class A shares at one cent each, roughly 18% of that class, for about $508,000. Rum Group is the former conservative video platform Rumble, and its early investors include Peter Thiel and Vice President JD Vance (The Information).
Twenty-five Fields Medalists told the AI labs that mathematics is not a leaderboard
Titled A Severe Misalignment of AI in Mathematics, the declaration went up on Terence Tao's blog on Friday with 25 initial signatories, all winners of the Fields Medal, the discipline's highest honor. It argues that labs are treating famous open problems as benchmarks while leaving mathematicians to verify, correct and attribute whatever the machines hand over, and that proofs arriving faster than anyone can read them destroy the ground new ideas grow in. It names no company and no incident. It follows the counterexample to the Jacobian Conjecture that a mathematician found with Claude Fable in July, which we covered at the time, and the attribution fight over OpenAI's Navier–Stokes result that we covered last week. Tao is the most conspicuous AI optimist in the discipline and he signed it. OpenAI had already pulled its sponsorship of a Caltech AI mathathon, which Anthropic and OpenAI had backed with $2 million in AI credits, on Thursday, after an open letter from current and former Caltech mathematicians (Terence Tao · Business Insider).
Nvidia is in talks to put as much as $10 billion into Anthropic's stock market debut
Word of the discussions surfaced on Friday. The chipmaker would come in as an anchor investor in the offering we covered in August, according to people familiar with the talks, and nothing is settled. Nvidia sells the chips, invests in the companies that rent them out, and would now own a slice of one of the largest buyers of what comes out the other end (PYMNTS · Benzinga).
Enflame, a Tencent-backed Nvidia challenger, closed its Shanghai debut up 179%
Enflame Technology raised 6.12 billion yuan, about $913 million, in its Shanghai listing. On Friday the stock opened at 410 yuan against an offer price of 142.18, touched 475 yuan during the day, and closed at 397, a first-day gain of 179% that values a company which has never turned a profit at 170.9 billion yuan, about $25.5 billion. Retail orders ran to more than 4,000 times the stock available, by the South China Morning Post's count. Enflame booked 990 million yuan of revenue last year, about $147 million, and is the last of the four Chinese AI chip startups investors call the little dragons to reach the market (CNBC · South China Morning Post).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Cognition's new coding model is an American product built on a Chinese open one. SWE-2, released Thursday, is a post-trained version of Kimi K3, the 2.8-trillion-parameter open model from China's Moonshot AI, tuned for coding and for the cost of finishing a job rather than for general ability. It scores 50% on FrontierCode, Cognition's own benchmark, within a point of the score Cognition itself measured for Anthropic's Fable 5.1, at 64% less cost, and ships inside Devin, Cognition's coding agent, free to paid tiers through October 10. The cheapest route to a competitive American coding product now runs through Beijing (MarkTechPost).
Timnit Gebru says the extinction talk is there to change the subject. Gebru, the computer scientist who was pushed out of Google's ethical AI team in 2020 and now runs the Distributed AI Research Institute, told WIRED in an interview published Friday that AI companies are stoking fear of human extinction to avoid discussing the harms their products cause now, autonomous weapons among them. It runs against the week's dominant story, and it lands while the labs are asking Washington for permission to coordinate (WIRED).
░░▒▒▓▓ TECH ▓▓▒▒░░Matt Mullenweg is chief executive of Automattic again, days after the board tried to remove him (update). The company said on Saturday evening that the WordPress founder is back as chairman and chief executive with the board's full support, reversing the paid leave of absence we covered on Friday. In between he locked other administrators out of the company Slack, told employees he was back in control, and published a blog post about buying a houseboat. Nobody has explained why he was removed, and nobody has explained why he was restored (TechCrunch).
American regulators have been running at least three investigations into possible insider trading on Polymarket that nobody had reported. Documents obtained by WIRED under freedom-of-information law show at least three previously unreported investigations by the Commodity Futures Trading Commission, which oversees prediction markets, into possible insider trading on the platform: bets on Biden pardons, on the Iran war, and on Google search rankings. The agency has separately brought an insider-trading case over Google-related contracts (WIRED).
░░▒▒▓▓ INFRA ▓▓▒▒░░SpaceX put rocket engineers in charge of its data centers, and the build-out may slow. More than 300 engineers have rotated onto the data center team since a June shake-up that handed it to Wesley Salandro, a SpaceX veteran of more than seven years who ran Falcon and Dragon production. The new team is installing more backup power and cooling and testing both before sites go live, rather than adding them afterward as xAI did when it built the Colossus campus in Tennessee in 122 days. A construction mishap in Memphis on September 3 took some Grok models offline for three and a half hours and, The Information reports, hit paying customers, Anthropic among them. Several data center leaders have left in recent weeks, some to OpenAI and some to Anthropic, which rents computing from SpaceX (The Information · Engadget).
Chip designers are putting less memory on each AI processor, deliberately. Nvidia's coming Rubin Ultra will carry 192 gigabytes of high-bandwidth memory per accelerator, the stacked memory that sits beside a processor and feeds it data, down from 288 on Blackwell Ultra and on standard Rubin, and a long way from the terabyte the industry expected a year ago. SemiAnalysis, the chip research firm, argues the reversal is only partly about the wafer shortage: a shorter stack delivers the same bandwidth as a tall one, and bandwidth, not capacity, is what serving a model actually consumes. Buy the four-layer part and you get the most tokens per dollar, which is the number that determines what an agent costs to run all day (SemiAnalysis).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Jeff Dean's new company is asking investors for a $50 billion valuation. Discovery Loop, which the former Google chief scientist founded with fellow Google researchers Sanjay Ghemawat, Quoc Le and Oriol Vinyals, was seeking $1 billion at about $10 billion a few weeks ago and is now seeking a valuation of around $50 billion, Business Insider reported on Friday. It describes itself as automating discovery to accelerate science and engineering, by running thousands of experiments in parallel. Business Insider places it among the new labs raising at large valuations before they had products (Business Insider).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Xi Jinping offered the BRICS countries an open-source AI bloc. At the summit of BRICS, the bloc built around Brazil, Russia, India, China and South Africa, in New Delhi, the Chinese president named open-source AI and smart factories as two of five initiatives for the Global South, the South China Morning Post reports, and offered to train young talent from other member states. In a statement his foreign ministry released on Sunday, China said it will pioneer a BRICS AI open-source community, support cooperation on developing and applying large language models, hold AI seminars and training courses, build a BRICS digital ecosystem cloud platform and set up a BRICS engineer cultivation alliance, CNBC reports (South China Morning Post · CNBC).
A Russian drone hit a passenger train near the Ukrainian-Polish border after a train carrying European dignitaries had left the station. The drone hit the locomotive of a passenger train on Sunday in Volyn region, about two kilometers from the Polish border; a separate strike near the Yahodyn crossing closed it temporarily. A diplomatic train carrying the former British prime minister Boris Johnson, the former Swedish prime minister Carl Bildt and security advisers from several EU governments had left the station ahead of schedule; Ukraine's rail operator says that train was the probable target. A second train, with former CIA director David Petraeus aboard, was still at the platform. No injuries were reported (The Washington Post · NBC News).
Saudi Arabia shut the pipeline that was its way around the Strait of Hormuz. Drones launched from Iraq struck the East-West line near Riyadh and Medina on Thursday, according to the Saudi energy ministry, causing injuries and damage, and on Friday the kingdom said it had closed the pipeline as a precaution. The 1,200-kilometer pipeline had been carrying four to five million barrels a day to the Red Sea port of Yanbu, about 4% to 5% of world supply, after the strait itself was effectively closed. It follows the Houthi strikes on Saudi oil facilities, as we noted last week. UBS economist Alessandro Bee warned on Friday that oil could reach $150 to $200 a barrel if Hormuz stays shut for several quarters (Al Jazeera · CNN · SWI swissinfo.ch).
░░▒▒▓▓ ALPS ▓▓▒▒░░Angela Merkel says the American technology companies are running an intellectual expropriation program. At Deutsche Telekom's Digital X conference in Cologne on September 8, the former chancellor told interviewer Miriam Meckel that the American technology companies are running what she called an intellectual expropriation program, and that Germany should be more upset about it. Her examples were secondhand bookshops bought out and pulped into training sets, and the tax base. What is left to tax, she asked, once there is no human labor? She put this year's data center spending by the four largest American technology companies at $725 billion against the European Union's $1 trillion over five to seven years, and reached for the term she learned in East German schools for what that buys, state-monopoly capitalism. Deutsche Telekom chief executive Tim Höttges, on the same stage, asked who is actually paying for these systems and then answered himself: all of us, eventually (Handelsblatt).
Switzerland has again put off regulating the large platforms. Reto Vogt, in his Friday column for DNIP, the Swiss tech-critical blog, writes that Federal Councillor Albert Rösti has delayed the draft law once more, and Vogt has no patience for it. Bern has been promising platform rules for years, and Vogt notes that even parts of the United States are getting something done in a small way while Switzerland keeps hesitating (DNIP).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░Nine years of looking for cold fusion turned up nothing, and the report saying so took two more years to appear. Project Callisto, an experimental program that ran from 2015 to its closure in 2024 to settle whether low-energy nuclear reactions are real, released its final report on September 8. It searched many loading, stimulation and material conditions, across both classic electrochemical replication and an ultrapressure gas-phase campaign, using palladium films thin enough that a nuclear event would punch a hole visible under an electron microscope. It found no evidence of any cold fusion phenomenon, and ordinary non-nuclear explanations for most previous reports of excess heat in electrochemical cells. The team calls the null result definitive for the space it searched (Project Callisto).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Mecka AI is closing a Sequoia-led round at a valuation of about $500 million, for paying strangers to film themselves making coffee. The two-year-old startup fits people with body sensors and has them record everyday tasks on a phone, then sells the motion data to companies training humanoid robots, the data-labeling trade Scale AI and Surge built for language models. The round, led by the venture firm Sequoia Capital and reported on Friday, comes three months after a $60 million raise led by Framework Ventures; its size has not been disclosed and the terms are not final. As of early June the company was projecting a $100 million annual run rate by year end (TechCrunch).
░░▒▒▓▓ PAPERS ▓▓▒▒░░The Intern-NCP team at Shanghai AI Lab and Shanghai Jiao Tong University — a model taught to predict the next idea as well as the next word matched its open-model baseline's training loss on half the data. The technical report, posted to arXiv on September 9, describes an 8.9-billion-parameter model trained on two objectives at once: ordinary next-token prediction, plus prediction of discrete concepts spanning several tokens, drawn from a vocabulary the model quantizes out of its own hidden states. Trained on 5.73 trillion tokens, it reaches the final pretraining loss of the open model OLMo-3-7B after consuming 51.3% of them, and finishes 2.45 points ahead on a downstream average, including a 5.99-point gain on grade-school math word problems. The comparison also mixes the new objective with a larger model, 8.9 billion parameters against 7 billion; the paper's own size-matched comparison shows a smaller saving. It is a preprint, and these are one team's numbers on its own model rather than an outside evaluation. (arXiv).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Microsoft's communications chief is leaving after seventeen years in the job. Frank X. Shaw told his team on Friday that it is time to move on and try new things. The former Marine Corps public affairs officer has spent nearly three decades around the company, the last seventeen years as its communications chief, running its message through the antitrust years, three chief executives, and the weekend in 2023 when its largest partner fired and rehired its own. He has no job lined up and Microsoft has no successor to name; it will look inside and outside. For practitioners, the chair turns over exactly as the industry's central argument shifts from what these products can do to whether they should be built this fast (Axios · GeekWire).
Salesforce is taking the word Agentforce back off its products. A year after putting the agent brand across the portfolio at its annual customer conference, turning Sales Cloud into Agentforce Sales and Service Cloud into Agentforce Service, the company is stripping it from the names of several top products ahead of this year's event, The Information reports, citing two people with direct knowledge. Neither rename touched the product, the data or the price. Only the label moved, twice (The Information).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░The complete wiring diagram of a male fruit fly's brain, 166,000 neurons and 125 million connections, which Google Research and the Howard Hughes Medical Institute's Janelia campus released this month and which we covered last week, has now been pointed at Doom. Each frame stimulates the simulated sensory neurons, the resulting activity is mapped to the controls, and taking damage fires two dopamine cells as a reinforcement signal. The fly was on round 6,385 when PC Gamer checked last week and, by its account, not very good at it. It has also been set to work on Beat Saber and Super Mario 64 (PC Gamer).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Ich weiss nicht mehr — Evelyne / Masao
Back tomorrow — the desk at Uncertain Futures