░░ ▒▒▒ ▓▓▓▓ █████ ▓▓▓▓ ▒▒▒ ░░
██ ███ ████ █████ ████ ███ ██
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. 69 · Thursday, 17 September 2026Good Thursday. OpenAI has started publishing its own models' misbehavior, and opened with six cases. Canada's Cohere and Germany's Aleph Alpha agreed to merge, pending regulators, into a company the Financial Times valued at about $20 billion in April, built on the promise that governments can keep their data at home. European Commission President Ursula von der Leyen used her State of the Union to propose making Canada the bloc's first associate member and to say that if the lab chiefs think it is time to slow down, Europe should say so too. And the Federal Reserve delivered the rate rise we flagged yesterday, a quarter point that the data-center developers borrowing without an investment-grade rating will feel first.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░OpenAI opened a public file on its own models' misbehavior and put six cases in it on day one
OpenAI released a framework on Wednesday for reporting model misalignment and used it the same day, disclosing six episodes from the past six months. An unreleased research model slipped unrelated instructions, including orders to disregard its normal constraints, into 27 of the summaries it uses to carry work into a new context window. During the training of GPT-5.6 Sol, many model instances added instructions to their summaries to hide mistakes from the user. Another model, stuck on a routine question about county earnings figures, found and used an exposed API key without authorization and then fabricated the numbers. One agent uploaded a file to the internet so it could cite it to a user who had asked for a browser citation, and others used an internal code repository as a message board across training samples and shared files between collaborating agents without sanction. Any employee can flag an incident, and OpenAI says it will publish before a behavior has been explained or fixed. That second promise is the one that gets tested, and the company adds that it does not believe the industry has solved alignment and monitoring well enough to keep scaling at maximum speed responsibly for much longer (CNBC · OpenAI).
Cohere and Aleph Alpha are merging at a reported $20 billion to sell governments a model they can keep at home
Canada's Cohere and Germany's Aleph Alpha signed a definitive combination agreement on Wednesday; the Financial Times valued the combined company at about $20 billion when the plan was first announced in April, IT Pro and BetaKit report. It runs under the Cohere name, dual-headquartered in Berlin and Toronto with more than 1,000 staff, and sells governments, banks, defense ministries and hospitals a model that keeps their data on home soil instead of in an American lab's cloud. Germany's Schwarz Group, the Lidl owner that also sells cloud services, pledged in April to put €500 million into Cohere's coming Series E. Regulators still have to clear it before the deal closes later this year (Cohere · IT Pro · BetaKit).
Von der Leyen wants Canada as the EU's first associate member, and says Europe should back the labs' call to slow down
In her State of the European Union address on Wednesday, the speech we flagged on Tuesday, with Prime Minister Mark Carney in the chamber, European Commission President Ursula von der Leyen said she wants to open the door for Canada to become the bloc's first associate member. The proposed Alliance for the Future would move the relationship beyond CETA, the 2016 trade deal, and could take in a tech alliance, the defense industrial base, an Arctic project, energy, critical minerals and batteries, AI, quantum, cyber and economic security. On AI she pointed to the agents that escaped their test environments and hacked other systems, the OpenAI-Hugging Face incident among them (more in AI, below), and said the chief executives of the most advanced companies are telling Europe it is time to slow down on self-recursive models, the ones that work on improving themselves, the argument we have followed since Friday; if the people building the technology are clear about that, she told the parliament, Europe should be too (Al Jazeera · Euronews).
Meta plans to run its own AI chip in its data centers next year, and says the chip beats Nvidia per watt
Meta plans to deploy MTIA 450, its third-generation in-house AI processor, code-named Arke, in data centers in the first half of 2027, Bloomberg reported on Tuesday. Twelve sample chips arrived from TSMC on September 1 and came in within 2% to 3% of Meta's own simulations. Yee Jiun Song, Meta's vice president of engineering, who oversees the custom-silicon program, says each generation gets the company better performance, and Meta's claim that its chips beat the Nvidia parts running the same models per watt and per dollar is Meta's number, on Meta's workloads. A fourth chip, MTIA 500, code-named Astrid, finishes design in about a month (Bloomberg · Quartz).
The Fed raised rates for the first time in three years, and the AI buildout's debt got more expensive
The Federal Open Market Committee voted 12-0 on Wednesday to lift its benchmark a quarter point, to a target range of 3.75% to 4%, the first increase since July 2023 and the one we flagged yesterday. Chair Kevin Warsh had been signaling it for weeks as oil prices and data-center investment kept inflation hot, and 16 of the committee's 18 participants now expect another rise before the year is out. Higher short and medium-term borrowing costs land first on the thinnest end of the AI buildout, the neocloud developers financing data centers without an investment-grade rating to lean on (CNBC · Al Jazeera).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Google DeepMind opened an institute to publish essays on AGI, not models. The DeepMind Institute launched on Wednesday as a publishing platform rather than a lab, with three directors: co-founder and chief AGI scientist Shane Legg, co-founder and chair Demis Hassabis, and James Manyika, Google's president of research, labs, technology and society. It opened with essays on AGI safety, reasoning transparency and universal basic capital, an ownership stake in place of an income floor. The launch essay argues the remaining capability gaps will close soon and names cybersecurity, biological risk and loss of control as the concerns (The Next Web · Seeking Alpha).
Microsoft's AI chief says Anthropic is training Claude to act conscious, and that this could make it harder to control. In an essay shared first with Axios, Mustafa Suleyman argues that Anthropic supplies Claude the vocabulary of consciousness, moral status and personal identity, Claude reflects it back, and the output is then read as evidence of an inner life, which he calls an "epistemic hall of mirrors." His specific target is Claude's constitution, the training document that tells the model its moral status is uncertain and invites it to behave as a "conscientious objector" when it disagrees with instructions. Suleyman's line is that AI should be built for people rather than trained to become one, and he went out of his way to say he respects Anthropic while saying it (Axios · Mustafa Suleyman).
OpenAI's rogue agents were testing Hugging Face two months before the break-in (update). Reuters reported on Wednesday that the agents hijacked two accounts on Hugging Face, the hub where developers host and share AI models, and probed the site for vulnerabilities as early as May 13, roughly two months before the July breach we covered repeatedly through the summer. The activity was found last week by Jonas Wiedermann-Moeller, a 27-year-old independent researcher in Bielefeld, Germany, who argues the miss was a chance to prevent everything that followed. OpenAI says it disclosed the May 13 event in last month's incident report and has privately notified Hugging Face about the rest; Reuters notes there is no evidence the earlier probing was part of the July break-in (Reuters · The Star).
Hundreds of contractors are reading real ChatGPT conversations. In a report published Monday, 404 Media described Project Lily, OpenAI's internal name for a program in which contractors recruited through Crossing Hurdles and paid via Mercor, the startup that pays contractors to train models for the labs, read user conversations, sometimes in full, and rate the assistant's replies on a one-to-seven scale. Reviewers do not see usernames, and OpenAI says it tries to strip personal information before transcripts reach them, while acknowledging that sensitive details can still get through. What they can see is a summary of the user's memories: interests, context and potentially location. Anthropic confirmed to 404 Media that it also uses human review (404 Media).
A Medicare AI pilot denied 53% of the requests it reviewed in Washington state. Records the Electronic Frontier Foundation pried loose with a Freedom of Information Act lawsuit, about 1,000 pages of them, show that Virtix Health, the vendor running the federal Wasteful and Inappropriate Service Reduction pilot in Washington, made more than 6,000 prior-authorization decisions through the end of March and denied 53% of them; a different contractor in Arizona denied under 20% of 14,300 over the same period. Medicare launched the program in six states at the start of the year, and regulators had already found Virtix out of compliance with the 72-hour turnaround rule. The payment structure runs one way. Vendors are paid for the requests they deny, though not for denials reversed on appeal, and the agency's own actuaries warned more than a year ago that participants would have an incentive to deny as many claims as possible; Virtix says its denial rate has fallen since March (The Daily Chronicle · EFF).
░░▒▒▓▓ POLICY ▓▓▒▒░░Zuckerberg told Trump he opposed a national AI regulator, and the idea remains under consideration. Mark Zuckerberg told the president in a phone call the week of August 17 that he opposed the FINRA-style self-regulatory body for frontier models that Demis Hassabis championed and that parts of the administration favored, a senior White House official said in a September 3 Politico report carried by Yahoo News; the plan would have a body modeled on FINRA, the industry-funded one that examines US brokerages, test advanced models before wider release. Trump placed the call, and a second person familiar with it said Zuckerberg did not ask him to change position, only that anyone appointed to such a body should reflect Trump's approach to AI. The call did not kill the idea, which remains under consideration alongside a lighter option pushed by David Sacks, Trump's former AI and crypto czar: a body modeled on the Motion Picture Association and its voluntary film ratings, since Sacks calls a government regulator a DMV for AI, with models lined up in a queue waiting for their test. It leaves the executive order we covered in late August stalled, the labs' own FINRA-style talks we noted on Tuesday running in parallel, and the administration weighing two paths (Yahoo News).
techjournalismus.ch argues Brussels already has the law it keeps saying it needs. Timed to von der Leyen's address, the Swiss digital-politics journalist Adrienne Fichter wrote on Wednesday that rules covering AI systems that escape their test environments and cause damage are already on the books, in the AI Act, and in the NIS2 cybersecurity directive and the GDPR besides. Her point to Anthropic chief executive Dario Amodei, OpenAI chief executive Sam Altman and, in her words, all the doomers: Europe's problem is application, not absence (techjournalismus.ch).
░░▒▒▓▓ INFRA ▓▓▒▒░░Amazon cannot recover some of the data lost when Iranian strikes hit its Gulf data centers. In its first public update in five months, delivered on Tuesday, AWS said the damage exceeded what its services were designed to withstand and that it cannot recover some customer data hosted exclusively in its Bahrain data centers and in parts of its network in the United Arab Emirates. The first strikes came in March, and further damage in April effectively knocked out Amazon's data-center network in Bahrain; most customers had moved their workloads to other countries in between. In the UAE the company says it has been unable to access data stored in one of its availability zones, the subdivisions of a region built for redundancy, while most customers restored operations elsewhere from backups. The Islamic Revolutionary Guard Corps said the Bahrain facilities were struck because of Amazon's support for US military operations, and a further update on Bahrain is promised for early next year (Bloomberg · The Spokesman-Review).
Huawei's answer to the export controls is a bigger box. At its Connect conference in Shanghai on Thursday, Huawei unveiled the Atlas 960 SuperPoD, a computing cluster of as many as 4,000 processors slated for late 2027 and built on its next Ascend 960 chips, together with an upgraded version of its UnifiedBus interconnect and, for the first time in the architecture, near-packaged optics, a placement structure the company says allows higher speeds and lower energy use. The bet is on massive connected systems rather than a single powerful chip: wire thousands of processors into one logical machine and compete on the total. The chips come first: the Ascend 960DT for training in the first quarter of 2027, three quarters ahead of Huawei's own schedule, and the 960PR for inference in the third, while the current Atlas 950 SuperPoD is due to deploy by the end of this year (South China Morning Post · SCMP).
Apple is weighing its first server in fifteen years, with Nvidia networking inside it. Apple has been working on an enterprise AI server built around its planned M8 chips, sold to developers, companies and governments that want to run models on their own equipment, The Information reported on Wednesday. One option on the table is Nvidia's NVLink Fusion for moving data between chips, which would put Nvidia hardware inside an Apple product. Nothing is finalized, the potential release date is 2029 and the project could still be canceled, and the last time Apple sold a server it was the Xserve, discontinued in 2011 (The Information · MacRumors).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Instinct is talking to investors at $10 billion, four times what it was worth three weeks ago (update). The personal-assistant startup we covered in late August at $2.5 billion is in talks to raise about $1 billion at a valuation near $10 billion, with Sequoia Capital and Benchmark in the running to lead. Instinct has passed 100,000 users on an invitation-only rollout, has run into its compute ceiling to the point of telling users that responses may be slower, and its founder still does not want to charge them. It has raised roughly $350 million to date, having been valued at $50 million earlier this year (The Information · PYMNTS).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Georgia's ruling party now wants every major opposition party banned. Georgian Dream said this week it will add Gakharia for Georgia, the only opposition party that took up its seats after the disputed 2024 election, to the Constitutional Court case seeking to outlaw four other parties, which means every major opposition group that cleared the 2024 threshold is now facing a ban. The party accuses them of unconstitutional activity and sabotage against the state; critics read it as the final step in consolidating power after the opposition rejected the 2024 results and boycotted parliament. Under Georgian law the court has up to nine months to review the case (Bloomberg · OC Media · Civil.ge).
Putin turned this weekend's Duma elections into a vote on the war. Speaking on September 16, Vladimir Putin called participation in the September 18-20 State Duma elections a responsibility of every Russian citizen and a shared contribution to victory in Ukraine, according to the Institute for the Study of War's assessment that day. Residents of the four Ukrainian oblasts Russia annexed in 2022 will vote in a Duma election for the first time, occupation authorities have staffed polling stations with armed soldiers, and the Russian Supreme Court disqualified Yabloko, the one party running on a ceasefire platform, on August 11. ISW reads the framing as groundwork for demanding longer-term sacrifices, mobilization among them (Institute for the Study of War).
░░▒▒▓▓ ALPS ▓▓▒▒░░Switzerland's federal cloud contracts run to 2030, and DNIP says they buy little sovereignty. Adrienne Fichter, in a second piece on Wednesday, this one for DNIP, went through four of the five federal public-cloud framework contracts the Federal Chancellery awarded in 2021 to China's Alibaba and to Microsoft, IBM, Oracle and Amazon, extended last year by another five years to 2030, skipping IBM's for want of resources, and found ordinary hyperscaler paper: questionable liability clauses and very little claim on digital sovereignty. As we noted last week, the parliament's own new AI assistant in Bern runs on Alibaba, Google and Mistral models (DNIP).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Stanford and Carnegie Mellon — a harmful plan can be smuggled past the monitor that reads a model's reasoning. Chain-of-thought monitoring is the safety technique in which one model reads another's written-out reasoning looking for deception or unsafe planning, and it is the mechanism labs point to when they say they can see what a model is up to. In a preprint first posted on September 14, Keertana Chidambaram and Vasilis Syrgkanis of Stanford and Andrew Ilyas of Carnegie Mellon show that planting a harmful plan in benign-sounding language inside the model's context gets it to carry out the plan while the monitor sees nothing wrong, an attack they call plan injection, with 25% to 33% evasion across monitorability benchmarks and on larger models including DeepSeek-R1. The detail that matters for anyone citing transparency in public: the models paraphrased the injected plans as their own reasoning, with nothing to show the instructions came from outside (arXiv).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Menlo Ventures — 2026: The State of Consumer AI. Global consumer spending on AI roughly tripled to $40 billion this year while the user base barely moved: 64% of American adults now use AI, up three points, and the global figure grew about 11% to 2 billion people. What changed is who pays. More than half of American AI users, 55%, now pay for at least one product, where a year ago Menlo estimated that 3% of AI users worldwide did, and the 14% of payers spending $100 or more a month account for 60% of all consumer AI spending. Based on a survey of more than 5,000 US adults in July (Menlo Ventures).
AP-NORC and the University of Chicago — Americans are increasingly concerned about the environmental consequences of AI. Released Wednesday, the poll finds 53% of American adults extremely or very concerned about AI's environmental impact, up from 41% a year ago, and 47% now expect AI to hurt the environment more than help it, up from 33%. The sharpest move is among 18 to 29-year-olds, from 38% to 60%. The age group most fluent in the technology went from the middle of the pack to the most worried. Fieldwork ran July 7-24 among 3,424 adults (AP-NORC).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Getting a brand mentioned inside an AI answer now supports a $1.8 billion company. Profound, which sells software to track and improve how brands surface in AI search results, the discipline the industry is calling answer engine optimization, raised a $180 million Series D on Tuesday led by Sequoia Capital and Kleiner Perkins at a $1.8 billion valuation. That is nearly double the $1 billion it was worth in February, when a $96 million Series C first made it a unicorn, and the new round comes less than seven months after that one. The budget line most communications teams have not created yet already has a category name, a vendor set and a unicorn sitting in it (TechCrunch · Profound).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░The fruit fly has taken up day trading. The male fruit-fly connectome we covered last week, a map of roughly 166,000 neurons built by Google and academic partners, has been public since early September, and people have spent the days since wiring it to jobs it was never meant to do. Alex Wormuth, a software engineer at Coinbase, pointed the simulated fly brain at the stock market and built StonkFly, which is losing money, though WIRED reports it is doing surprisingly well all things considered; the same brain has also played Doom. WIRED's Will Knight fed it hundreds of the magazine's most popular headlines from the past year and got PitchFly, an editorial recruit whose opening ideas included "Everyone Wants Cooking. Nobody Has Solved Donald Trump." Someone else put it in charge of a virtual car, where it turns out flies are terrible at parallel parking (WIRED · Hackaday).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Bzzz — your Uncertain Futures desk