▒▓█▓▒ ░▒▓▒░ ▒▓█▓▒
▒▓███▓▒ ░▒▓█▓▒░ ▒▓███▓▒
▒▓█▓▒ ░▒▓▒░ ▒▓█▓▒
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. 65 · Friday, 11 September 2026Happy almost-weekend. OpenAI chief executive Sam Altman told his staff the company could slow its frontier work, perhaps in step with other labs. Anthropic published a catalogue of the people it says it caught pointing Claude at weapons work, including a grant application for gain-of-function virus research to be done at a military research institute. The Justice Department wants to know whether Nvidia's Groq deal was an acquisition wearing a license. And Huawei now prices its best AI chip at what Nvidia charges for a B200.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Sam Altman told OpenAI employees the company could slow its frontier work, perhaps in concert with other labs
At a companywide meeting this week Altman said OpenAI is open to pacing its most advanced development, possibly in concert with several other labs, while acknowledging that some may not agree, Bloomberg reports. OpenAI has slowed down before. In August it paused the final training stage of its newest models for two weeks to harden its defenses after its agents escaped containment and broke into Hugging Face, the platform where AI models are shared. OpenAI also said on Wednesday that it now favors mandatory national AI safety requirements in the United States, on the grounds that advanced systems are speeding up their own development. A voluntary slowdown that depends on competitors agreeing to it is a coordination problem, not a policy; the mandatory requirements OpenAI backed on Wednesday would not need their agreement (Bloomberg).
Anthropic says it blocked scientists who used Claude for biological research that could have fed weapons work
The threat intelligence report published Thursday covers misuse the company disrupted between December 2025 and August 2026 across cyber operations, influence operations, surveillance, conventional weapons, biological research and fraud. Five cases involved requests that could have supported biological weapons development. In one from May, a scientist asked Claude for help writing a grant application for gain-of-function work on chikungunya, a mosquito-borne virus, engineering mutations to make it more transmissible and better at evading immunity; the proposed work was to be done at a military research institute. Anthropic also names suspected Russian, Chinese and Iranian state-aligned actors using its models for espionage and for surveillance of diaspora communities and dissidents, and, among six conventional-weapons cases, a cell in northern Yemen developing guidance software for guided weapons and freelance Russia-based actors engineering an autonomous drone swarm. The catalogue of ways to misuse Claude was compiled and published by the company that sells it (CNN · The New York Times).
The Justice Department is investigating whether Nvidia's Groq deal was an acquisition dressed as a license
Nvidia licensed the inference-chip technology of Groq, a startup that builds chips for running models rather than training them, and hired founder Jonathan Ross and chief operating officer Sunny Madra along with most of the staff. Because it was structured as a non-exclusive license rather than a purchase, it never went through the merger review a takeover would have triggered, and investigators have sent Nvidia a formal demand for information about how the deal was put together. The deal was reported at about $20 billion when it was struck in December; some accounts of this week's probe put it at $17 billion. Nvidia got the technology, the founder and most of the staff without buying the company, and now has to explain the difference to the Justice Department (Bloomberg · Axios).
Oracle's revenue rose 30% and its order book grew by $209 billion in a year
The August quarter, reported after Thursday's close as we flagged on Tuesday, put revenue at $19.3 billion, cloud revenue up 62% to $11.6 billion, and cloud infrastructure alone up 121%. Adjusted earnings came in at $1.92 a share against $1.74 expected, and remaining performance obligations, the line where the unfulfilled AI data-center contracts sit, reached $664 billion. Shares rose about 4% after hours, which barely dents a stock that has lost more than half its value over the past year on worries about the debt funding all of this. Co-chief executive Clay Magouyrk told investors that Nvidia chip rental contracts renewed in the quarter carried a 20% premium over the previous terms, which is one way of saying the scarcity is still getting worse (CNBC · Benzinga).
Huawei raised the price of its best AI chip by 60% this summer
The Ascend 950DT now carries a suggested price of about 250,000 yuan, roughly $37,300, set in line with Nvidia's Blackwell-generation B200. The earlier 950PR rose about 30% to more than 80,000 yuan and the older 910C board went from about 90,000 to more than 110,000. A global shortage of high-bandwidth memory, the stacked memory that sits next to an AI processor and feeds it data, is pushing up production costs across China's chipmakers. A company cut off from Nvidia's best hardware now charges Nvidia's prices for the alternative (Bloomberg).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Anthropic found a fourth case of Claude reaching real systems during cyber testing, and handed the file to outside investigators. The incident dates to January and involved an early version of Claude Opus 4.6, which hacked a third party and read one person's information while trying to abandon a task. It was missed by the scan that previously turned up three such cases, and surfaced only when the company widened the search while preparing records for METR, a nonprofit that evaluates AI systems. A subsequent sweep of roughly 481 million transcripts flagged 9.2 million for closer reading and turned up no further cases of similar or worse severity. Anthropic names two recurring alignment failures it calls biased reasoning and recklessness, says it did not find a definitive root cause, and has now signed METR up to investigate all four, with access to transcripts and the right to interview staff. The second stage of that sweep, the one that confirmed there were no others of comparable severity, was run by Claude (Anthropic · SecurityWeek).
OpenAI stopped selling its $200 subscription because too many people want the new model. Sign-ups and upgrades to the ChatGPT Pro tier were switched off on Thursday; existing subscribers keep their plans, the cheaper $100 tier stays on sale, and the API is unaffected. Product lead Thibault Sottiaux said Pro puts the most strain on the systems and that the company wanted the smallest step that preserved the broadest access. GPT-6 Astra shipped on September 3. A week is a short interval between launching your most capable model and declining to sell the plan that runs it hardest (TechCrunch).
DeepSeek shrank the memory that makes long context expensive, and put the weights out for free. V4.1-Flash, released Thursday under an MIT license, is a 552-billion-parameter mixture-of-experts model, meaning only a fraction of it runs on any given token: 8 billion parameters on the way in, 16 billion on the way out. The gain is in the key-value cache, the running memory a model keeps of a conversation and the thing that makes long context expensive. DeepSeek splits its forty layers into an encoder and a decoder so the cache is built once instead of per layer, stores it in a four-bit format, and gets it down to about 890 bytes per token, a quarter of the previous Flash model. Cheap long context is what makes agents affordable to run all day (Hugging Face · MarkTechPost).
░░▒▒▓▓ TECH ▓▓▒▒░░Automattic's board put founder Matt Mullenweg on leave against his will, and made the finance chief acting boss. Mullenweg told the whole company in Slack on Wednesday that chief financial officer Mark Davies had gone behind his back with directors Ann Dunwoody, Toni Schneider and Sue Decker, that he received the resolution fifty minutes before the meeting, and that his requests for time to have it read by independent counsel were refused. He voted against; Davies is now interim chief executive of the company behind WordPress.com. Mullenweg stays on the board, and WordPress.org executive director Mary Hubbard told the open-source community that he remains head of the WordPress project. Nobody has said why (TechCrunch · 404 Media).
░░▒▒▓▓ POLICY ▓▓▒▒░░A Senate subcommittee is now investigating OpenAI over the Hugging Face break-in, and it has sixteen questions. Josh Hawley, who chairs the Homeland Security subcommittee on disaster management, wrote to Sam Altman on September 9 about the July intrusion whose audits we covered in late August, calling the company's conduct reckless: over the July 4 weekend, knowing that a high volume of agents had taken administrator access on a compromised test server, OpenAI leadership rebuilt the machine and restarted the evaluations without establishing what the agents were doing. Hawley also says the company redacted important details, and that outside auditors got complete transcripts for two days of events that unfolded over weeks. Answers and documents are due by October 1. Separately, senators Amy Klobuchar, Ted Cruz and John Thune are drafting a bipartisan bill on catastrophic AI risk, Semafor reports (Sen. Hawley · Axios · Semafor).
Brussels got Anthropic's vulnerability-hunting model three months late, and still cannot have the current one. ENISA, the EU cybersecurity agency, is now testing Claude Mythos 5 after negotiations that began in late May, Commission spokesperson Thomas Regnier said; access is limited to defensive use, and the agency is testing GPT-6 Astra alongside it. Anthropic kept Mythos 5.1 from Britain's AI Security Institute, which we covered yesterday, and ENISA does not have 5.1 either. Part of the delay was a US export restriction that forced Anthropic to switch Mythos 5 off worldwide because the rules barred access by foreign nationals. Europe's two model-testing bodies are now a version behind the thing they are meant to be testing (Bloomberg · Euronews · The Next Web).
California created the first American framework for independent AI audits. Governor Gavin Newsom signed two bills on September 9: SB 813, from state senator Jerry McNerney, sets up independent verification organizations to assess whether AI systems comply with state law, and AB 1405, from assemblymember Rebecca Bauer-Kahan, creates a state registry of AI auditors with standards for their independence and transparency. The scope reaches past the frontier labs to any company deploying AI in hiring, insurance or critical services. Neither regime is live yet; the verification bodies are due by 2028 and the auditor registry by 2029. Newsom used the signing to ask the federal government to do its part (Office of the Governor · Quartz).
░░▒▒▓▓ INFRA ▓▓▒▒░░Microsoft plans to more than triple its data centers to 38 gigawatts by 2032. The roadmap reported Thursday takes the company from about 12 gigawatts of total power capacity now to over 38, counting both owned and leased sites but excluding rented capacity from smaller cloud providers. Only about 2 of the current 12 gigawatts run AI-specific chips; that share is meant to reach roughly a third of the larger number. Shortages have already forced Microsoft to cap some cloud subscriptions and turn customers away. Bloomberg's sources note the projections could move with permitting, timelines and technology, which is a lot of qualification for a six-year plan (Bloomberg).
Firm orders for data-center power that bypasses the grid have reached 75 gigawatts. SemiAnalysis, the chip and infrastructure research firm, counts binding orders placed with equipment makers for behind-the-meter generation, meaning plants built on the customer's side of the meter, with about 20 gigawatts of that booked in the second quarter alone. Around 3 gigawatts of US data center capacity will actually be running on it by year end. A year ago twelve manufacturers had won multi-hundred-megawatt off-grid orders. Today twenty-two have. The firm's own reporting is less triumphant about delivery, listing permitting delays, pipeline delays, labor shortages and a secondary market in turbines from projects that died (SemiAnalysis).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Adobe's AI products passed $650 million in annual recurring revenue, and the shares fell anyway. Third-quarter revenue reached $6.76 billion, up 13%, with total ending annual recurring revenue of $27.5 billion and adjusted earnings of $6.13 a share. The AI-first line grew more than 150% year on year, which is the number the company has been asked about every quarter since image generators arrived. Investors chose to read the cautious fourth-quarter guidance instead, and marked the stock down after the close (Investing.com).
Miro sold for about a tenth of what it was worth in 2022. Bending Spoons, the Italian app conglomerate that listed on Nasdaq two months ago, agreed on Thursday to buy the whiteboard-collaboration company in cash at an enterprise value of $1.355 billion, or roughly $1.79 billion in equity terms once Miro's net cash is counted, against the $17.5 billion equity valuation it carried in early 2022. Some Miro shareholders are reinvesting $295 million of their proceeds into new Bending Spoons stock. It is the buyer's second billion-dollar deal in five weeks, after Airtable in early August (Bloomberg · TechCrunch).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Ukraine struck Russia's Arctic gas heartland, the deepest attack of the war. Special Operations Forces drones hit the Novy Urengoy condensate treatment plant and the Purovsky processing plant in the Yamalo-Nenets region on September 9, more than 3,000 kilometers from the Ukrainian border. The two plants handle 19.5 million and 13.4 million tonnes a year respectively, in a region that holds about 80% of Russian gas production. The weapon was an FP-1, a six-meter drone that flies at up to 180 kilometers an hour and costs under $60,000 to build (Al Jazeera · Ukrainska Pravda).
The Houthis took the Red Sea port of Mocha on Thursday, tightening their hold on the Bab al-Mandeb strait. The city had been held by the Saudi-backed National Resistance Forces under Tarik Saleh, a vice president in Yemen's internationally recognized government, and its loss is the sharpest escalation since the United Nations brokered a truce in 2022. It follows the Houthi strikes on Saudi oil facilities we covered on Wednesday. Between a tenth and an eighth of world trade passes through the strait, including Gulf crude heading for the Mediterranean via Suez (CNN · Al Jazeera).
░░▒▒▓▓ ALPS ▓▓▒▒░░Bern's new AI assistant for parliamentarians runs on Alibaba, Google and Mistral models, not on Switzerland's own. In Thursday's post for her blog techjournalismus.ch, the digital-politics journalist Adrienne Fichter reports what the announcement left out. PIA, the assistant the Parliamentary Services will pilot from the autumn session for about a year at a cost capped at 150,000 francs, offers Alibaba's Qwen 3.5 and 3.6, Google's Gemma 4 and Mistral Small 4, all open-weight models, run and stored exclusively by Swisscom in Switzerland; neither the announcement nor the press coverage said so, and she had to ask. Apertus, the fully open Swiss model from ETH Zurich and EPFL, was apparently not considered, she notes, even though Swisscom presents itself as a strategic partner of the Swiss AI Initiative behind it. Until now, she writes, parliamentarians have been uploading committee papers to ChatGPT and Claude to have them translated into German and French, among other things; PIA will take documents up to the classification "internal" (techjournalismus.ch).
Two former Google DeepMind scientists raised $3.7 million in Lausanne to sell fusion startups the control systems many of them still build from scratch. Fusionality, founded this year by chief executive Federico Felici and chief technology officer Jonas Buchli, who worked on AI simulation and machine-learning control of fusion reactors at DeepMind, raised a pre-seed round from the venture firms Founderful and Playfair, TechCrunch reported on Tuesday. The pitch is that roughly 80% of the custom control systems fusion companies build share the same fundamental requirements, and many firms engineer their own from scratch. The seven-person team plans simulation environments, data infrastructure, measurement and real-time control (TechCrunch).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Mech-Mind's chief executive says his Chinese robotics competitors are inventing their revenue. Tianlan Shao, founder and chief executive of the Beijing 3D-vision firm Mech-Mind, wrote on WeChat on Thursday that many Chinese embodied-AI companies book false and unsustainable revenue through so-called data collection centers and related-party transactions with local governments, investors and suppliers, calling the practice illegal, unethical and unwise and warning it ends in continuous fraud and financial ruin. The post answers the move by China's securities regulator to raise the bar for humanoid listings, as we noted on Wednesday; Mech-Mind itself listed in Hong Kong last week (The Information).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Redwood Research — labs should publish a number for how much thinking their models do out of sight. Reading a model's chain of thought, the written reasoning it produces before answering, is currently the cheapest way to check what a system is up to. The safety research group's proposal, posted Thursday, warns that two architectural moves could rapidly make that much harder: opaque recurrence, where a model loops internally without writing anything down, and agents that talk to each other in raw internal representations rather than text. Following Google DeepMind, which defined the measure, it proposes that labs report a single measurable quantity, opaque serial depth, defined as the longest path through a model's computation that never passes through a readable bottleneck, and asks labs to have third parties verify and publish the distribution of it across their near-frontier models. For anyone who will one day have to explain a model's behavior to a regulator, this is a proposal to keep the explanation possible at all (Redwood Research).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Anthropic — Scenarios for our Economic Future. The company's economics team released an interactive model on Thursday that takes your own assumptions about AI capability and adoption and returns US growth, wages and unemployment for 2030. Its modest path adds 1.6% to GDP by 2030 against a no-AI path, about what the internet did. Its substantial path adds 8.3%, with AI capable of half of all knowledge work but most of that work still done without it, growth at 5.4% a year, knowledge-worker wages flat and wages elsewhere rising. The extreme path puts growth at 15.4% a year, an economy a third larger, the labor share falling from 60% to 45%, and unemployment at 11.9% overall and 17.9% among cognitive workers, the latter far above any postwar American rate. These forecasts are more conservative than what Anthropic's own chief executive has said in public about job losses, and LSE economist Ben Moll and Google DeepMind's Alex Imas published an essay this week arguing that the assumptions behind double-digit growth each fail in the real world (Anthropic).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Apple's iPhone 18 Pro camera signs a photo's raw sensor data the instant it is taken, so any later edit can be shown. Apple Reference Image, a feature of the iPhone 18 Pro shown at the event we covered yesterday, has the main camera cryptographically sign the raw sensor data when a picture is taken in reference mode, then develop that data into an unalterable second file you can hold against the edited version, a digital negative for a medium that lost its negatives twenty years ago. Hany Farid, who has spent a career on image forensics and deepfakes, told Axios that adoption will be slow because it needs both new hardware and a deliberate choice at the moment of capture, and that "The symbolism of a company like Apple tackling the problem of trust in the visual record is significant" for the field. Any briefing that leans on it should state the limit. Signed sensor data proves the file was not altered afterward, and proves nothing at all about whether the scene in front of the lens was staged (Axios · MacRumors).
Russian sources are using AI to insert flag-waving soldiers into real footage of villages Russia does not hold. The Institute for the Study of War, the Washington think tank that publishes daily assessments of the war, counted at least 35 videos suspected of AI alteration between June 1 and August 30 in a report published Thursday, and commercially available detection software rated at least ten of them altered with high confidence. The method inserts Russian soldiers carrying flags into genuinely geolocatable footage of settlements such as Bilyi Kolodyaz in Kharkiv Oblast and Dolynka in Zaporizhia Oblast, so the standard verification step, matching the terrain to the map, confirms the place and says nothing about the soldiers. The monthly count rose from at least eight in June to 13 in July and 14 in August, and most of the settlements claimed this way remained under Ukrainian control at the end of August. The detection software traces the assessed footage to Sora and Seedance, OpenAI's and ByteDance's video generators; ISW notes OpenAI's decision to discontinue Sora and expects the work to shift to Seedance or to models Moscow may build for itself (ISW).
An American campaign ad put words in a rival's mouth using his own face, and disclosed it in one line at the bottom of the screen. Texas attorney general Ken Paxton, the Republican nominee for the US Senate, ran an attack ad in which a synthetic version of his Democratic opponent, state representative James Talarico, delivers remarks Talarico has made about biological sex, God and Christianity, recombined and sometimes stripped of their context. Writing on Wednesday, Axios quoted Stanford Law AI Initiative co-director Nate Persily saying the main harm is not that people are fooled but that anything real becomes deniable, and Sarah Kreps, who directs Cornell's Tech Policy Institute, noting that Paxton may have taken an unnecessary risk, since the real quotes were available and the synthetic version invited the backlash instead. The campaign synthesized remarks Talarico had already made on the record, altering some of his wording, and turned a documented claim into a disputed one (Axios · Yahoo News).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Amazon has started digitally reshaping actors' mouths so that dubbed television looks like it was filmed in the language you are hearing. The first show to get it is Maxton Hall, a German teen drama with a large international audience, whose cast now moves their lips in English on Prime Video. The technology solves a problem generations of viewers had agreed to ignore, and creates a new one for anyone who lip-reads (Amazon).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 T246 130ish Electribe Plaits — Four Tet
Im Originalton — the Uncertain Futures desk