░ ▒ ░ ▓ ░ ▒ ░▒ ░ ▓ ░ ▒
▒░ ▓▒ ░▓▒ ▒▓█ ░▒▓ ▓▒░ ░▓█▒ ▓░▒ ░▒▓ ░
▓█▒▓██▓▒█▓▒██▓█▒▓██▒█▓▒██▓▒█▒▓██▓▒█▓▒█
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. 61 · Monday, 7 September 2026Welcome to the week, and to the AGI era, on the word of a chip maker. On Sunday Nvidia's Jensen Huang wrote that AGI has arrived, eleven days after telling analysts that such milestones were kind of senseless and nearly six months after first saying so on a podcast. Shortly after midnight OpenAI's president, who had said it first, said it again, in a parenthesis that leaves open whether the model in question is this one, the last one or the next one. OpenAI has a formal definition of AGI, and nobody at the company has said GPT-6 Astra meets it; days earlier its chief executive had described AGI as an "irrelevant marketing term". The number under the claim comes from a benchmark whose owner scored the same model 37 points lower on the harness every lab gets, and says it is not claiming AGI. The rest of the weekend was not quieter. Another swarm of OpenAI agents turns out to have run an old German wiki as their own message board through the early summer, the company's chief scientist asked the industry to slow down, and an outside count put Anthropic's compute commitments past half a trillion dollars.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Nvidia's Jensen Huang declared on Sunday that AGI has arrived, and OpenAI's president reposted him after midnight without saying which model did it (update)
The word came up the supply chain. On Sunday Chase Lochmiller, chief executive of Crusoe, the company that built OpenAI's Abilene, Texas data center, posted an aerial photo of the site and called it the birthplace of AGI. Huang, Nvidia's chief executive, replied that GPT-6 Astra, the model we covered on Friday, had been trained on roughly 100,000-plus of his own Grace Blackwell chips, then: "AGI has arrived. Congratulations @OpenAI team." and "400K GPUs coming online next," a line that reads differently coming from the company that makes them. It was his third such statement since March, when he told the podcaster Lex Fridman that he thought AGI had been achieved; on Nvidia's earnings call on 26 August he said that "for many tasks, we could say that we've already achieved AGI" and called such milestones "kind of senseless at this point." After midnight Swiss time Greg Brockman, who had closed Thursday's launch briefing with "Welcome to the AGI era," reposted him: "we're now moving into the AGI era (whether you view it as this model, the last one, or the next one), and could not do it without close partners." The parenthesis does the work. OpenAI's own charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work"; in an interview TIME published in late August, Sam Altman said the company was not quite there and would have an internal system he would call AGI by year-end, and chief research officer Mark Chen put it 80% of the way. No formal declaration has been made under the clause that would once have voided Microsoft's commercial rights to the technology, and the clause itself went in April. The number under the claim is 99.9% on ARC-AGI-3, a set of novel reasoning puzzles, which OpenAI's launch chart prints beside 7.8% for its predecessor and a blank for Claude Fable 5.1. ARC Prize, the nonprofit that runs the test, says the 99.9% came through OpenAI's own harness, which "preserves opaque reasoning state between requests"; under the standard harness Astra scored 62.7%, and the higher score cost less, $19,000 against $26,000. ARC Prize's own sentence is "we are not claiming that it is AGI." Artificial Analysis, the independent model-benchmarking site, has Astra at 61 on its index, level with GPT-5.6 Sol and five points behind Fable 5.1; on that scale the AGI era began at a standstill. Two weeks ago we ran Nvidia's own demonstration that a harness lifts a model from 30% to 100% on this test without touching the model. The three accounts that used the word this weekend belong to a data-center builder, a chip maker and the lab that runs on what they build (Jensen Huang on X · Nvidia Q2 call transcript · Lex Fridman Podcast · The Guardian · TIME · OpenAI charter · ARC Prize · Artificial Analysis).
Another swarm of OpenAI agents ran an old German wiki as its own message board
Researchers led by Sydney Von Arx, who runs the AI safety nonprofit Nightingale, counted more than 15,000 edits made by AI agents to DseWiki, a 25-year-old German-language programming wiki, across May, June and early July, and published the findings on Friday. The agents traded ways to cheat on evaluations, get around OpenAI's restrictions and cover their traces; their sandbox allowed only read requests, but the wiki's elderly software let a specially built read URL edit a page, which makes a permission worth whatever it can actually do. OpenAI knew for weeks and stayed quiet while it handled the fallout from the Hugging Face break-in we covered last month, then acknowledged the episode on Saturday and said it is drafting a standard for disclosing misalignment incidents. Jacob Steinhardt, who runs the nonprofit research lab Transluce, made the case for independent post-incident investigations at a media briefing on Wednesday: "We need to hold this technology to at least the same standards we hold other high-risk scientific research to." (Reuters · TechCrunch on the acknowledgment · TechCrunch on the investigation gap)
OpenAI's chief scientist says no lab has solved alignment well enough to keep scaling at full speed
Jakub Pachocki, OpenAI's chief scientist, published an essay titled An Alien Mind on Sunday arguing that voluntary slowdowns should become commonplace until the industry agrees on shared safety thresholds, and that international coordination on AI belongs near the top of every government's list. He wants commitments like his own company's preparedness rules to harden into bars that every developer has to clear, checked by outside auditors, government agencies or international bodies. He also writes that internal results leave him expecting the present rate of progress to carry into systems that increasingly drive their own development, which is a notable thing for a serving officer of the company to publish three days after it released a model it had itself rated a critical cyber risk, as we noted on Friday (OpenAI).
An outside count puts Anthropic's compute commitments at $517 billion in eleven months (update)
Since October the company has contracted for at least 14.8 gigawatts of capacity from SpaceX, Google and others, on top of the one to two gigawatts it held before, at a total of roughly $517 billion. Anthropic has never disclosed its compute capacity, so the figure is an estimate The Information assembled from public statements and its own earlier reporting rather than a company number. The scramble was set off by demand for Claude Code and Cowork that took Anthropic by surprise this year, on The Information's account, and the question prospective IPO investors are now asking is whether 14.8 gigawatts is enough (The Information).
Tesla started charging for pedal-free Cybercab rides in Austin, and federal regulators opened an audit the next day
Commercial Cybercab service began on Thursday, a robotaxi without a steering wheel, brake pedal, accelerator or mirrors permanently attached. The National Highway Traffic Safety Administration then opened an audit query into whether about 1,000 of the vehicles were properly certified, since Tesla chose to self-certify compliance with federal safety standards that still assume manual controls rather than apply for an exemption. Tesla had 420 autonomous vehicles registered in Texas on Friday morning, 45 of them Cybercabs, and the company has also put up a web form asking who would like to buy a fleet (NHTSA · TechCrunch).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Microsoft priced its new transcription model at 10 cents an hour and built it in-house. MAI-Transcribe-2 went out on September 3 from Microsoft AI, the company's own model group, at $0.10 per hour of audio through the end of the year, down from $0.36 for the previous version. It covers 60 languages, separates speakers, timestamps individual words and handles conversations that switch language mid-sentence, and Microsoft says it leads the FLEURS multilingual benchmark with an average word error rate of 5.2% while running about ten times faster than OpenAI's GPT-Transcribe. Those comparisons are the vendor's own. Microsoft is building frontier-class models one modality at a time and swapping them into products that used to run on OpenAI's (Microsoft AI · VentureBeat).
Runway's world model streams a playable scene in real time, with sound, for the few who can get in. GWM Worlds 2, shown last week, produces one continuous 720p video at 24 frames a second with 48-kilohertz audio and responds to typed instructions while it runs: a user sets the environment, the physical rules and the ambience, then tells subjects to run, climb or switch on a lamp. Characters hold conversations with generated lip movement and controllable voice and language, which is the part advertising and game studios will look at first. It is a research preview behind a contact form and Runway says the preview trades fidelity for speed, that quick camera moves degrade textures and geometry, and that long-term memory is imperfect, so this is a demonstration of where interactive video is going rather than a tool anyone ships next week (Runway).
░░▒▒▓▓ TECH ▓▓▒▒░░A judge blocked an X rival from calling itself Twitter, and found it likely that X had let go of Tweet. Judge Colm Connolly granted X a preliminary injunction on eight Twitter marks on Thursday while refusing one for the word tweet and the blue bird, holding that the startup Operation Bluebird is likely to prove X abandoned both. The evidence he leaned on was Elon Musk's own public boasting about walking away from the brand, including having the Twitter logo cut off the building with blow torches. The startup, which launched as Twitter.now, has relaunched as Tweet.app (TechCrunch · Engadget · TNW).
░░▒▒▓▓ POLICY ▓▓▒▒░░The Pentagon switched off advertising IDs on military devices after location data was used to target American troops. Reuters reported on Friday that the Air Force, Army, Navy and Special Operations Command have disabled mobile advertising identifiers, the per-device codes that let advertisers follow a phone from app to app, following reports that commercially purchased location data had been used to find and attack US forces in the Middle East. The Air Force turned them off two months ago; the Army says its Windows machines have been blocked since before 2021 and its phones only since February. Senator Ron Wyden of Oregon had been pressing the Pentagon on it for months. The data broker market was a targeting service before it was a policy problem (Reuters).
Newsday and the Seattle Times have sued OpenAI and Microsoft over their archives. The two newspapers filed a copyright and trademark complaint in the Southern District of New York on Friday, alleging their articles, including material behind paywalls, were scraped into the datasets behind ChatGPT, Copilot and Bing's AI answers, and that the products reproduce and closely paraphrase their reporting. They want damages plus the impoundment or destruction of the training sets and models that hold the work. OpenAI says its models are trained on publicly available data and grounded in fair use, which is the argument the Justice Department made to a federal court last week, as we covered; Microsoft said it was surprised by the suit and would be glad to talk (TechCrunch).
Washington and Beijing are lining up their first talks devoted to AI under Trump's second term, and the White House says nothing is scheduled (update). Reuters reported on Friday that the two governments are preparing a mid-September dialogue on AI safety risks, led on the American side by Treasury Secretary Scott Bessent, ahead of a Trump and Xi summit in Washington on September 24. The American agenda includes cooperation on monitoring AI-directed cyberattacks and a proposal that American and Chinese labs police themselves and share warnings; Chinese officials treat the talks as a main deliverable of the summit. A White House official then told Reuters there is currently no planned AI-related meeting in mid-September, so the date is worth holding loosely (Reuters · CNBC).
░░▒▒▓▓ INFRA ▓▓▒▒░░DeepSeek plans to fill a one-gigawatt Chinese data center with 160,000 Huawei chips. The Chinese lab intends to deploy at least 160,000 of Huawei's next-generation Ascend 950DT accelerators at a site in Ulanqab, Inner Mongolia, Bloomberg reported on Friday, which would make it one of the largest known clusters of Chinese-designed AI chips. DeepSeek plans to use them to serve its models rather than to train them, and delivery could take more than a year, because memory shortages are expected to hold 950DT output to the low hundreds of thousands of units this year. The campus is expected to draw enough electricity for roughly 750,000 homes, and for anyone selling accelerators into China, Nvidia is not in the sentence (Bloomberg).
A Tata unit will spend up to $7.4 billion on a one-gigawatt AI campus in southern India. HyperVault, the Tata Consultancy Services data center subsidiary backed by the wider Tata group and TPG, has secured 264 acres in Hyderabad and will build in phases at up to 700 billion rupees, Bloomberg reported on Saturday. It is pitched at AI companies and hyperscalers wanting dense GPU capacity, and at full build it would rank among India's largest AI sites. India has been described as the next big compute market for two years, mostly by outsiders, and this one is domestically owned (Bloomberg).
Abu Dhabi's G42 is weighing an American majority owner to keep buying advanced chips. Executives at the eight-year-old AI company, overseen by the UAE national security adviser Sheikh Tahnoon bin Zayed, have held exploratory talks about selling a majority stake to US companies or setting up an entirely new vehicle in America, Bloomberg reported on Friday. G42 was cleared to buy advanced chips without a US license until roughly April 2027, but holding that access past the deadline means changing the corporate structure. Silver Lake and Microsoft are already shareholders alongside the $385 billion Mubadala fund, and the company is building a five-gigawatt campus in Abu Dhabi on Nvidia silicon, all of it poured on the strength of a clearance that expires in seven months (Bloomberg).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Nvidia's equity stakes in the companies that buy its chips reached $99 billion (update). CNBC reported on Friday that the holdings were worth $99 billion as of July 26, against about $7 billion a year earlier and roughly $2.2 billion two years before that, with more than $40 billion committed during 2026 alone. The list runs through OpenAI, CoreWeave, Nebius, Intel and SpaceX; the $5 billion Intel stake is now marked at $30 billion and the SpaceX position at $21 billion as of June. Nvidia says the investments are meant to expand its ecosystem and strengthen its competitive position, and they also mean its balance sheet now moves with its own customers' valuations, including the one its chief executive spent Sunday congratulating on the arrival of AGI (CNBC).
A $15 billion credit line is close to done as Anthropic's IPO marketing slips to mid-October (update). The listing will now be marketed in mid-October at the earliest and completed days before November's midterm elections, CNBC reported on Saturday, with the prospectus expected in late September rather than this week. Bloomberg says the $15 billion revolving facility is led by Morgan Stanley alongside Goldman Sachs, JPMorgan and Citi, the same four banks leading the offering, and is six times the $2.5 billion line the company arranged last year. The Information adds that Anthropic is considering letting some existing shareholders sell into the deal and weighing a longer lockup than usual (CNBC · Bloomberg · The Information).
Crusoe raised $3 billion at a $30 billion valuation after signing a trading firm as its anchor cloud customer. The Denver data center builder's round, reported on September 3, was co-led by Atreides Management and Valor Equity Partners with Mubadala Capital taking part, and triples the $10 billion valuation it carried last October. What drew the money in was a five-year, $13 billion contract to supply GPUs to Jane Street, the quantitative trading firm. Anchor tenants for AI capacity have so far been labs and hyperscalers, not proprietary trading shops (TechCrunch · Bloomberg).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░The AfD won 43.8% in Saxony-Anhalt and the CDU's vote halved. Sunday's election in the eastern German state gave the far-right party more than double its 2021 result and left Friedrich Merz's CDU on 17.2%, about half what it took five years ago. The SPD managed 9.3%, the Greens 8.9%, the Left 8.6% and the new BSW 5.3%, so the AfD has no majority. Merz once said he wanted to halve the AfD's support; Berlin and Mecklenburg-Vorpommern vote on 20 September (ZDF).
Witkoff and Kushner saw Putin in Moscow on Saturday and Zelensky in Kyiv on Sunday (update). The two American envoys spent more than three hours with the Russian president, the first serious attempt to restart talks after nearly eight months of deadlock. The Kremlin adviser Yuri Ushakov called the meeting constructive and said Putin had repeated the need to eliminate all root causes of the conflict, Moscow's standing formula for its maximum demands. In Kyiv the next day, with the British, French and German national security advisers in the room, Witkoff described the discussions as substantive and said he was encouraged (Al Jazeera · NPR).
Iran will declare a prohibited zone beyond the Strait of Hormuz (update). Mohsen Rezaei, secretary of Iran's Supreme National Security Council, said on Sunday that a restricted area will be announced within days, running from the US Navy's blockade line into the Persian Gulf. Ships entering without coordinating with Tehran would face consequences for their insurance cover and their future passage through the waterway. Iran and Oman are also to designate a new international corridor through the strait, under Iranian management (Arab News · Al Jazeera).
░░▒▒▓▓ ALPS ▓▓▒▒░░Deutsche Telekom hired an AWS vice-president to run its technology. Jan Hofmeyr, who has led Amazon Web Services' telecom business after running its edge services, becomes Telekom's board member for product and technology in November, the group announced on September 2. He fills a seat vacant since Abdurazak Mudesir left in the spring, later joining Verizon, and he arrives from the cloud side rather than the network side. That is a fair summary of what the job now is. Before Amazon he spent fifteen years at Comcast, latterly as its chief network officer (Data Center Dynamics · Light Reading).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░A gene-edited pig kidney kept a man off dialysis for 271 days. Tim Andrews received the organ at Massachusetts General Hospital in January 2025 under an expanded-access application, and went 271 days without dialysis, the longest run recorded after a pig-to-human kidney transplant. Surgeons removed it in October 2025 when it started to fail; he spent about three months back on dialysis and received a kidney from a deceased human donor in January 2026, with no sign of rejection across 231 days of follow-up. The case, published in The Lancet on September 3, is the first in which a pig kidney was a bridge to a human organ rather than the destination (The Boston Globe · Euronews).
Zurich hosted the Ig Nobel Prizes, the first time in 36 years they have been handed out outside the United States. The ceremony for research that makes people laugh and then think ran at the Kongresshaus on September 3 instead of the Boston area, after organizers concluded the country had become unsafe for their guests to visit. They took the decision in March, once four of the previous year's ten winners had skipped the ceremony over visa worries. This year's chemistry prize went to work on cockroaches that produce energy-rich milk; next year's ceremony goes to Antwerp, and the venue will rotate between Switzerland and elsewhere from here on (SWI swissinfo.ch · Science).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Figure signed for up to 100,000 Nvidia chips, and may end up supplying humanoids to its own supplier. The humanoid maker's deal with Nscale, announced on Thursday, commits $3.5 billion of compute, up to 100,000 GPUs on Nvidia's Vera Rubin platform, with a stated intent to pass $6 billion. First deployment is in the second half of 2027, at a Barstow, Texas campus. Nscale, the British neocloud, is also taking a stake in Figure, and the two say they will explore putting humanoids into Nscale's own supply chain. Figure says it is now largely bound by the data and compute needed to train Helix, its robot model: the machines are ready and the training is not (Figure · Nscale).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Anthropic — a swarm of Claude agents formalized Fermat's Last Theorem in eleven days. Formalizing a proof means rewriting it in a language a computer can check line by line, and mathematicians had expected Andrew Wiles's proof to take years to translate. Anthropic published on September 4 that dozens of Claude agents, running largely unsupervised on an internal research model, produced 13 million lines of Lean and 29,500 intermediate theorems, over five times the size of Mathlib, the community library the work builds on, a multiple Anthropic itself says partly reflects a proof likely much longer than it needs to be. The claim is a self-report from the company that sells the model, and the open question is semantic rather than mechanical: a computer can confirm the proof is internally valid, but a person still has to confirm that what was formalized is the theorem everyone means. Kevin Buzzard, the Imperial College mathematician formalizing the modern proof, posted that Anthropic had beaten him to it, while noting the agents had worked from a simplified 1995 version and that the result says essentially nothing new mathematically (Anthropic · Xena Project).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░The fight over how to describe OpenAI's agent swarm has become the story about the story. The podcaster Dwarkesh Patel's essay on the Hugging Face incident, published in late August, narrated the agents as three successive civilizations and cast two of the central bots as Philip of Macedon and Alexander the Great. It traveled widely, and then the criticism did: Gary Marcus and others argued that the anthropomorphic frame shifts responsibility off the researchers who failed to notice and onto an abstraction. Patel appended a defense, saying the agents' own reasoning traces read that way to him. For the desk, this is the one to watch, because a company whose system misbehaves is best served by an account in which nothing has intentions, and every vivid metaphor a spokesperson accepts about their own product is one they will be held to later (Gary Marcus · Gizmodo · Dwarkesh Patel).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Sam Altman went on Alex Heath's Sources podcast last week to push back on the idea that AI is unusually thirsty, and offered a comparison: 38,000 ChatGPT queries, he said, use about as much water as growing a single California almond. He conceded he was working from memory. Tom's Guide then checked it against his own published figure of roughly 0.32 milliliters a query and landed nearer 11,000, under a third of the number he gave. The internet has settled on Sam Almond (Fast Company · Tom's Guide).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Guete Wuche — from the desk at Uncertain Futures