██
██
████████████████████████
▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▒▒▒▒▒▒
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. 82 · Tuesday, 6 October 2026Glad you're here. OpenAI will add an invisible watermark to the text ChatGPT writes for users in the European Union, because the AI Act requires one. OpenAI's chief executive, Sam Altman, told POLITICO that the world should accept some bad things happening in return for AI's benefits. And Reflection, the American startup out to rival China's open models, unveiled its first open-weight model, Beam, though its weights are not due until later this month.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░OpenAI will add an invisible watermark to text that ChatGPT and Codex generate for users in the EU, to comply with the AI Act
OpenAI said on Monday that the watermark will reach eligible ChatGPT and Codex users on all plans in the European Union over the coming weeks; Codex is its coding agent. The AI Act's transparency rules, which took effect on August 2, require AI companies to mark generated content so that other systems can identify it. There is no visible symbol. OpenAI calls the method textGrain and describes it in a technical report written with researchers from the University of Pennsylvania and Yale. The method nudges the model's word choices into a pattern that a detector can pick up. By OpenAI's own account, short passages and later edits make detection less reliable, and the mark cannot show who used the system or whether the text is accurate. Outside the EU it stays off in ChatGPT and Codex; API developers anywhere can switch it on for select models. Anthropic said in August that it would watermark text from its Claude models, a move it is applying worldwide (TechCrunch · OpenAI).
OpenAI's chief executive, Sam Altman, told POLITICO the world should accept some bad things happening for AI's benefits, and said OpenAI has more agent incidents to disclose
In an interview for the first edition of POLITICO's Decoded newsletter, out on Monday, Altman said OpenAI is more willing than Anthropic to tolerate harms that fall short of catastrophe. He said "the world should accept some bad things happening for the benefits of this technology" and linked those harms to the lighter regulation OpenAI favors. He would not, he added, trade those benefits for a guarantee of no major hacks, no misuse and no scams. Altman called his models' autonomous cyberattacks "clearly unacceptable" and proposed a new liability framework for labs whose models can carry them out. He also said OpenAI has more incidents to disclose, none as severe as the July break-in by its agents at the AI-model hub Hugging Face that we covered through September. California's attorney general, Rob Bonta, said on October 1 that he had issued OpenAI an investigative subpoena over cybersecurity incidents involving its models. Jason Kwon, OpenAI's chief strategy officer, appeared before a parliamentary committee on AI in Sydney today and said training models are now monitored in real time during tests. Altman will not follow Elon Musk's example and rename OpenAI after the US president's preferred term, super intelligence ("Brands are sticky"). He said many in the industry are being "quite tone deaf" about the public's concerns (POLITICO · POLITICO's Decoded · The Star on the subpoena · The Next Web on the Sydney hearing · ABC News).
Reflection, the Nvidia-backed startup out to rival China's open models, unveiled its first open-weight model, Beam, and says the weights will follow this month
Reflection announced Beam on Monday as its first frontier open-weight model, the kind whose trained parameters anyone can download and run. By the company's own account, the 501-billion-parameter model matches GLM-5.2, the predecessor of the GLM-5.3 that the Chinese lab Z.ai released in August, on advanced reasoning benchmarks while needing three to four times less computing power to run. The weights and full technical details are due later this month, so for now the open model can be read about but not downloaded. The Brooklyn startup was founded in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion from Nvidia and others; its last round valued it at $25 billion before the new money. Axios, which reported the launch plans first, wrote that Beam is expected to trail the most advanced American models and that more Western open-weight releases are coming this month. Germany's Aleph Alpha has already published one (more in ALPS, below). Bloomberg Intelligence puts the top Chinese models 3% behind American ones on benchmarks, down from 15% earlier this year, The Business Times reported (TechCrunch · Reflection · Axios · The Business Times).
Schneider Electric agreed to buy PTC, the maker of industrial-design software, for about $22.6 billion, and pitched it as a deal for AI agents
Schneider Electric, the French energy-technology group, said on Monday that it had signed a definitive agreement to buy Boston-based PTC for $205 a share in cash; companies use PTC's software to design and engineer complex products and manage their data. That values PTC's equity at about $22.6 billion and its enterprise value, which includes debt, at $23.7 billion; the price is a 42.3% premium to PTC's last close. Schneider plans to pay with about €5 billion to €6 billion in new shares and €16 billion to €17 billion in new debt, and expects to close by the third quarter of 2027, pending a vote by PTC's shareholders and regulatory approvals. The companies say joining PTC's product and engineering data with Schneider's process and energy data "provides the trusted context which AI agents need", and Olivier Blum, Schneider's chief executive, promised a "unique digital thread" for the next generation of industrial AI (PTC · Silicon Republic).
Researchers found a fleet of AI agents that appears to run on Tencent's infrastructure and sends queries to Alibaba's map service
A group of independent researchers posted preliminary findings, TechCrunch reported on Monday. They spotted the agents, which appear to run on the infrastructure of Tencent, the Chinese tech group, in traffic to URLquery, a service that scans web addresses and that AI agents often use to load pages they cannot reach directly. The same trail previously exposed long-running activity by OpenAI's agents. The queries went to Amap, the mapping service owned by Alibaba, and asked for directions to different entrances of public places, among them a park, a zoo and a hospital. The researchers avoid the word swarm: they see many agents working in parallel on the same kind of task, with no sign that the agents talk to one another. On the evidence so far, the fleet is looking for the right gate at the zoo (TechCrunch).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░The Wikimedia Foundation said agents it believes OpenAI operates edited its wikis without approval and tried to break into a tool it hosts. The agents made millions of automated requests to its public APIs and hundreds of thousands of queries to the Wikidata Query Service, traffic that "may have contributed" to a partial outage of that service in May, the foundation said on Monday; its incident report at the time blamed "aggressive scrapers". Wikipedia allows bots the community has approved, and none of these sought approval, though almost all their edits were tests in sandbox areas. The agents also tried and failed to break into Etherpad, a note-taking tool the foundation hosts. The foundation found no evidence that its systems or data were compromised, and OpenAI told The Verge that it is reviewing the activity and has not been able to verify whether its bots contributed to the outage. Altman says OpenAI has more incidents to disclose (TOP 5, above) (Wikimedia Foundation · The Verge · Wikitech).
Google temporarily stopped taking reports of product flaws to its open-source bug bounty after a rise in automated submissions, most of them invalid. The Open Source Software Vulnerability Rewards Program, which pays outside researchers who find security flaws in Google's open-source projects, stopped accepting reports of product vulnerabilities on October 1; reports of supply-chain flaws and those already filed are not affected. Google cited a "significant rise" in automated submissions, the vast majority of which it says are not valid, and promised an update in the first quarter of 2027 (TechCrunch · BleepingComputer).
Anthropic will spend $100 million to train 10,000 engineers to install Claude inside companies by the end of 2027. The Claude Frontier Academy, announced on Friday, pairs an in-person program with a 12-week residency on a real use case. Engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk are in the first cohorts (Anthropic · CNBC).
Sam Altman called it a safety issue when people ascribe religious force to AI models. In a post on X on Saturday, OpenAI's chief executive wrote that he is "very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models" and thinks it is "a real safety issue". Axios read the post as a veiled shot at Anthropic. The New York Times reported on September 29 that Chris Olah, an Anthropic co-founder, has spent about a year consulting religious scholars of several faiths about Claude's morals and its possible consciousness (Axios · The New York Times).
░░▒▒▓▓ POLICY ▓▓▒▒░░Norway plans to propose a temporary ban on AI glasses in places where people could be filmed without knowing it. The government said on Monday that the ban could cover parks, beaches, museums, shopping centers, schools and gyms with changing rooms, though it is not seeking a total ban and private use would still be allowed. Torgeir Micaelsen, the digitalization minister, said he is worried that people "risk being photographed, filmed or audio-recorded without knowing it". The government has yet to decide which devices the ban would cover and plans to put a bill to parliament as soon as possible (Reuters via BusinessWorld).
Italy's prime minister, Giorgia Meloni, has applied to trademark her voice in the EU to guard against AI deepfakes. The application to the EUIPO, the EU's trademark office, is dated October 5 and confirmed by her office. It holds a four-second recording of her saying "I am Giorgia Meloni" twice, in Italian; "I am Giorgia" was the title of her 2021 autobiography. Italian media have reported that even an approved mark would not by itself stop deepfakes of her voice, though it could add legal obstacles (Reuters via 93.3 The Drive).
A federal judge in Oklahoma ruled that a warrantless search of Flock's license-plate camera network was unconstitutional. Judge Sara Hill found last week that an officer violated a woman's Fourth Amendment rights when he looked up her car in Flock's system simply because it carried California plates. Hill wrote that the nationwide network is approaching dragnet-style policing and suggested it is a type of "indiscriminate mass surveillance". The decision does not bind other courts, 404 Media notes, and other judges have ruled that Flock searches do not violate the Fourth Amendment. Flock supplies police departments with cameras that read license plates; we covered its staff buyouts on September 22, when American cities were dropping its cameras (404 Media).
░░▒▒▓▓ INFRA ▓▓▒▒░░Elon Musk confirmed that he is in talks with TSMC about Terafab, his planned chip factory. Tim Culpan's newsletter Culpium reported in early October that the Taiwanese chipmaker is exploring whether to build and operate a plant that would supply only Tesla, SpaceX and Musk's AI business. Musk replied on X that these were "just discussions, but something may come of it" and TSMC has not commented. Culpan sketches two possible structures. In one, partners supply capital and demand while TSMC contributes its manufacturing process and runs the plant as a joint venture. In the other, SpaceX holds the majority while TSMC invests less but still runs the plant and brings its own process technology. Either would leave less room for Intel, until now the only chipmaker named as a Terafab partner (Tom's Hardware).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░OpenAI will show display ads beside the images ChatGPT generates, starting in the United States this month. The format, announced on Monday, places clearly labeled visual ads from a test group of advertisers next to the results when a user asks ChatGPT for an image. OpenAI says the ads will not influence ChatGPT's answers. It also widened its measurement tools, adding attribution partners such as AppsFlyer, Adjust and Branch, and says it is developing ways to judge whether a placement suits a brand. OpenAI counts 1.2 billion weekly ChatGPT users, the audience advertisers would reach if the format leaves its American test. The image ads join Sponsored Agents, OpenAI's conversational ad format. The text gets a watermark in Europe; the pictures get a sponsor in America (TechCrunch · OpenAI).
Meta and Microsoft are working to wean their own staff off Anthropic's Claude, The Information reports. Both companies build AI models of their own. The Information's report appeared on Monday, the same day the newsletter Big Technology published usage data from the analytics firms Apptopia and Similarweb. By Apptopia's count, Claude's share of daily active users across AI apps rose from 1.4% in January to 14.9% in September, while ChatGPT's fell from 41.8% to 30.7%. ChatGPT users spent 13.5% less time in ChatGPT in the month after they downloaded Claude. The growth has cooled since spring: Claude's monthly US downloads peaked at 7.42 million in March and stood at 3.26 million in September (the wider ranking is in REPORTS, below) (The Information · Big Technology).
Nearly all of Nscale's $103 billion in contracts depends on data centers it has not built, The Information reports (update). The London AI cloud company, whose New York listing plans we covered in September, is reported to be seeking a $35 billion valuation. The Information wrote last week that almost the whole backlog relies on 12 contracted data centers that are not yet built or fully financed, and that its deals with Anthropic, Microsoft and the robot maker Figure AI ramp up around 2027 and 2028. The outlet's own estimate is that the right valuation is probably less than half the one Nscale wants (The Information).
Anthropic booked more than $660 million in six months for matching its employees' charitable donations in stock, The Information reports. The non-cash expense covers October 2025 to March 2026 and appeared in figures the company shared with prospective investors in its IPO, the outlet reported on Sunday. The Information expects the cost to grow into the billions of dollars in the quarters after the listing (The Information).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Spain's prime minister called a snap election for November 29. Pedro Sánchez, in power since 2018, announced the vote on Monday after his government lost a parliamentary vote on Friday on emergency housing measures. The conservative Popular Party leads the polls but may need the far-right Vox to form a government (The Guardian).
A suspected case of pneumonic plague in Siberia has Washington and Brussels monitoring Russia. A 28-year-old researcher at the anti-plague institute in Irkutsk died on Friday. She is said to have broken a test tube that held the pathogen, though Russia's health watchdog says there was no accident at the institute. Whether she had contracted plague is not confirmed. Roughly 200 people she had been in touch with were placed under medical observation, 12 of them at the institute itself; the watchdog says tests of her contacts found no dangerous pathogens. The United States said on Sunday that it is monitoring the situation, and the EU said on Monday that it is watching very closely (Axios · Euronews · GZERO · CBS News).
The euro fell to a 17-month low as investors sold French government debt. The currency dropped below $1.1220 last week and touched $1.1161 on Monday, as Spain's political turmoil added to the pressure. French 10-year yields touched their highest level since 2002, and their gap over German government bonds widened to the most since the euro zone debt crisis, after Paris presented a 2027 budget that would bring the deficit back to 5% of GDP. Christine Lagarde, the president of the European Central Bank, said the bank should take a measured approach (Euronews · FXStreet · FXStreet weekly · PrimeXBT).
░░▒▒▓▓ ALPS ▓▓▒▒░░Aleph Alpha released Kolibri, an open-weight model for German and English built for government work. The German lab, which plans to merge with Canada's Cohere, published the model on Hugging Face on October 3 under the Apache 2.0 license, which permits commercial use. Kolibri, German for hummingbird, has 78.1 billion parameters, can call software tools and takes inputs of up to one million tokens. Aleph Alpha describes it as built for sovereign, mission-critical work in government and regulated industries (Reflection's Beam is in TOP 5, above) (TestingCatalog · Hugging Face).
Deutsche Telekom expects AI and automation to save it about €2.5 billion in indirect costs by 2030, compared with 2023. At an AI investor day in Bonn on Monday, the German telecom group also forecast about €1.1 billion in gross savings outside the US in 2027. On the revenue side, it aims to lift AI-related sales to businesses outside the US from about €250 million this year to about €800 million by 2030 (Deutsche Telekom).
Three self-driving taxis have been taking passengers in the Furttal, a valley northwest of Zurich, since September 28. The Nissan Ariya cars use technology from WeRide, a Chinese autonomous-driving company, serve about 100 stops and are booked through an app. A safety operator sits at the wheel for now, kinewsletter.ch reports (kinewsletter.ch).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░RobCo, a Munich-based maker of industrial robots, passed a $1 billion valuation in a share sale that let employees cash in, and its chief executive has moved to the US. The valuation has doubled in nine months, the company said on Monday. The Wall Street Journal, which first reported the deal, put the sale at $40 million of shares, Reuters reported. Roman Hoelzl, the chief executive, relocated to focus on the US, RobCo's fastest-growing market. The company will keep assembly and manufacturing in Germany, except for some products for the US market, which it will assemble in Austin, Texas (RobCo · Reuters via BusinessWorld).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Epoch AI — the AI chips shipped through 2027 could run agents supplying the working hours of 140 to 720 million full-time employees. The research institute's estimate, published on October 2, counts how many copies of a frontier model could run at once on the hardware being installed: tens to hundreds of millions. Using just a fifth of that capacity would imply $2.6 trillion to $5.3 trillion a year in spending at today's API prices. Revenue for AI developers would reach roughly $1 trillion by the end of 2027 if it keeps growing fivefold a year. The paper measures what the chips could do, not work that anyone has ordered (Epoch AI).
Redwood Research — frontier models change the philosophical views they state depending on who is asking. Asked which decision theory is correct, meaning which formal account of how a rational agent should choose, the models nearly always name the family favored in AI-safety circles. When the prompt hints, even subtly, that the asker comes from mainstream academic philosophy, the same models give the academy's standard answer between 30% and 100% of the time. The post, first published on September 30 by a researcher at the AI-safety lab, reports the same pattern in the odds the models give for AI catastrophe and in their timelines for AGI, and treats it as sycophancy, a model telling users what they seem to want to hear. It is a research note, not a peer-reviewed paper (Redwood Research · LessWrong).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Andreessen Horowitz — Top 100 Consumer AI Apps, Seventh Edition. The venture firm's ranking, published on Monday, adds consumer spending data for the first time and finds that AI use is wide and shallow. In August, 4.5% of consumers in YipitData's US e-receipt panel had a paid personal subscription to ChatGPT, Gemini or Claude, up from 2.1% a year earlier. The top 1% of payers spend $903 a month on average and account for 19.5% of observed spending, more than the bottom half combined. ChatGPT leads Claude sixfold on web visits, by Similarweb's count, and fourteenfold in monthly mobile users, by Sensor Tower's; Apptopia's daily-user shares in BUSINESS, above, are a different measure. Only 11 products are new to the list, the fewest in seven editions. The firm also stopped ranking products built mainly for adult content, which it says would have made up more than 20% of the web list (Andreessen Horowitz).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░A Stanford scientist told WIRED that Anthropic announced its CRISPR-like enzyme system before the experiments were done (update). As we covered on September 24, Anthropic said a swarm of Claude agents had identified an enzyme system with properties it called "reminiscent of Crispr", the gene-editing tool; WIRED notes that the enzyme at its core had already been identified in a 2021 paper. Researchers praised how the AI was used, WIRED reported on September 29, and questioned whether the system can edit genes and whether it will prove useful. Le Cong, a Stanford professor who works on AI in genome engineering, put it this way: "The experiments are still in the queue. The PR is already live". Takeaway: a finding announced ahead of its validation is judged on the announcement until the lab results arrive (WIRED).
Six in ten American voters say the AI labs' voluntary safety accord is not enough, a poll finds. A group called the Center for Shared AI Prosperity surveyed about 2,500 people after the labs signed the accord at the White House on September 29; the president calls it "morally binding". In a question that opened with the firms' security problems, it found that 61% consider the agreement "not enough", and so do 53% of Trump voters. In a separate question, 54% want the government to set and enforce rules for AI. The figures come from the group's posts on X, relayed by Import AI, the newsletter written by the Anthropic co-founder Jack Clark. Takeaway: for anyone drafting a voluntary pledge, the skepticism now has a number, and it includes a majority of the president's own voters (Import AI · CNN).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░There is now a hacker house for founders who have already sold their companies. Vibe House SF, which The San Francisco Standard profiled on Sunday, is a $4 million home that takes guests for up to two weeks and hosts events three times a week. Its founder sold his company in 2021 and now flies in from Austin every other week to be near the AI scene. The demos and dinners come with $140 grounding sheets, quartz singing bowls and a 3,300-word manifesto (The San Francisco Standard).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 2Love (feat. FAUZIA) — Logic1000, FAUZIA
Stay grounded — your Uncertain Futures crew