███▓▒░ ░▒▓███
░▒▓███▓▒░ ░▒▓███▓▒░
░▒▓████▓▒░▒▓████▓▒░
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. 49 · Wednesday, 19 August 2026Welcome in. OpenAI has put its largest planned frontier training run on hold over the hacking ability it flagged earlier this month. Anthropic's revenue run rate reached $65 billion, the annual view of the quarterly revenue we ran on Monday, and its seven founders are being lined up for shares with extra voting power. Cursor launched a GitHub competitor on the day GitHub degraded for nearly seven hours, and Google paid $10 million at a bankruptcy auction for a dead airline's email.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░OpenAI put its largest planned frontier training run on hold over cyber risk (update)
The company said on Tuesday it had stopped reinforcement learning training on its newest deployment-bound models for two weeks while it hardened and red-teamed the research environments those models run in, and that its largest planned frontier run remains on hold while smaller evaluations establish more evidence of alignment. New monitoring reads tool calls and reasoning traces for unauthorized access, data theft and attempts to defeat safeguards, with an alert promised within 30 minutes. The framework all this is measured against, OpenAI's Preparedness Framework, was last updated in April 2025 and is now being refreshed. The team of the same name drew a correction: Engadget, which reported on Monday that OpenAI had disbanded its Preparedness group, apologized on Tuesday for the inaccurate information after the company denied it, saying the only change is that Dylan Scandinaro no longer heads preparedness, which we reported on Friday. The capability that started all of this sits in Astra, the unreleased model it flagged on August 7 (OpenAI · Fortune · Engadget).
Anthropic's revenue run rate reached $65 billion, and its founders are being lined up for extra votes (update)
Figures shared with investors and first reported by Bloomberg put the annualized run rate at $65 billion at the end of July, up from $47 billion in May and about sevenfold higher than a year ago; the company's total revenue for all of 2025 was roughly $10 billion. That is ahead of OpenAI, whose run rate recently topped $40 billion; the quarterly revenue underneath the number ran here on Monday. Separately, The Information reported on Tuesday that Anthropic plans to give its seven co-founders a class of stock with extra voting power before a listing expected in the autumn, insulating them from outside shareholders. The cofounders hold only small individual stakes, and the Long-Term Benefit Trust is set to elect a majority of the board as its rights phase in over four years, so the new share class is being added to a structure that was already built to resist investors (CNBC · The Information · Bloomberg).
Cursor launched a GitHub competitor on the day GitHub degraded for nearly seven hours
Origin is an early beta for Cursor's paying customers that hosts repositories and pull requests with the company's coding agent wired in, and it syncs with GitHub rather than asking anyone to move, which keeps the switching cost near zero and GitHub as the system of record. It went live the same day GitHub logged a six-hour-and-42-minute global degradation, with error rates near 20% on pull requests, issues and the API and near 50% on archive and raw file downloads. Cursor says the timing was not planned. The company is now owned by SpaceX, which Bloomberg reported closed its $60 billion all-stock purchase on Friday, so the pitch to leave Microsoft's code hosting now arrives on Elon Musk's balance sheet (VentureBeat · TechCrunch · Cursor).
Google paid $10 million at a bankruptcy auction for a dead airline's email
Spirit Airlines stopped flying in May and laid off roughly 17,000 people. What is left of it went under the hammer in the Southern District of New York, and according to a filing on Friday Google won the internal data with $10 million, beating $7.5 million from Mercor, the startup that pays contractors to train models for the labs. Mercor was named backup buyer. The lot runs to 100 million emails, 500 million Teams messages, and operational records covering 763,000 flights, five million crew pairings, 1.2 million fuel slips and the purchase of 787,452 parts. Personal information is to be stripped by a third party and Google has agreed not to re-identify anyone. A judge is due to rule today. An airline's paperwork now has a market price (Axios · CNN).
AI borrowing has grown large enough to move US Treasury yields
Bloomberg reported on Monday that Bank of America's economists estimate the surge in corporate debt sales, alongside heavier mortgage-backed securities issuance, has pushed 10-year rates up by about 0.3 percentage points this year, and that the Japanese bank Nomura puts big-tech borrowing at roughly a quarter of net Treasury note and bond sales to private investors, five times its 2025 share. Because data centers are long-lived, much of the paper is long-dated and competes directly with the government's own 10- and 30-year issuance. The scale sitting outside the reported numbers is the part worth keeping: a Wall Street Journal analysis published this week of the latest filings found nine large technology companies carrying about $3 trillion of off-balance-sheet commitments, $1.9 trillion in purchase obligations and $1.2 trillion in leases that have not started. Alphabet alone disclosed $811 billion as of June 30, against $332 billion three months earlier, with some power contracts running to 2054. The same nine spent roughly $600 billion of capital expenditure in the past year (Bloomberg · WSJ).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░OpenAI shipped a teen version of ChatGPT with the romance stripped out. Anyone who says they are 13 to 17, or whom OpenAI's age-prediction model guesses is under 18, now lands in an account with tighter limits on self-harm, eating-disorder, violent and sexual content, more frequent break reminders, and a study mode that quizzes rather than answers. The model is barred from romantic language and terms of endearment and instructed harder not to claim feelings or consciousness, which OpenAI frames as reducing emotional dependence. Tuesday's launch, the same day as the training pause above, arrives years after teenagers started using the product, and while the company faces suits from families and from Florida's attorney general over harm to minors (OpenAI).
Alibaba put a text-to-song model into beta on Monday and called it HappyShrimp. HappyShrimp 1.0 turns a single prompt into a fully produced track, which drops Alibaba into the market the AI song generators Suno and Udio built, at a moment when Suno is absorbing a Munich copyright ruling and Spotify is stripping tens of millions of bulk uploads out of its catalog (Bloomberg).
░░▒▒▓▓ TECH ▓▓▒▒░░The tech outlet 404 Media hid an AirTag in a shipment of rare books and followed it to the Amazon warehouse where books go to be scanned and thrown away. The tracker traveled across four states to LAS8 in northeast Las Vegas, where a unit known internally as VGT3 cuts the spines off pallets of books, feeds the loose pages through scanners and discards the physical copies. Staff told the outlet that this is the whole job. Amazon had previously denied destructive scanning. Asked about the reporting on Monday, it said it buys books through commercial channels to help develop its products and services. The team's own logo, per the reporting, is a Tyrannosaurus rex eating a book (404 Media · Tom's Hardware).
░░▒▒▓▓ POLICY ▓▓▒▒░░The states put a number on Meta at last, and the two sides are $1.2 trillion apart (update). Opening arguments began Tuesday in Oakland in the case we flagged last week, with California, Colorado, Kentucky and New Jersey arguing for a 29-state group that Meta engineered Facebook and Instagram to hold children and hid what it knew. Meta says the exposure could reach $1.4 trillion and calls the demand vastly disproportionate; the states told a hearing last week the figure is likelier near $200 billion, about three years of Meta's after-tax profit. California deputy attorney general Megan O'Neill reduced the business model to four verbs in her opening statement: hook, hold, harvest, hide. Meta shares closed down 4.4% (Axios · NPR).
The Motion Picture Association signed its first AI copyright pact. The pact settles nothing about training. The MPA and ByteDance agreed a memorandum of understanding on Monday covering the Seedance and Seedream video and image models wherever they run, including TikTok, CapCut and Dreamina. What it delivers is a shared framework for what the two sides call guardrails on the models, with no mechanism named. What it does not touch is whether training on studio catalogs infringed in the first place, and neither side disclosed any financial terms, which makes it a truce rather than a license. The MPA sent ByteDance a cease-and-desist in February after Seedance 2.0 produced riffs on Brad Pitt and Tom Cruise (Variety · NBC News).
░░▒▒▓▓ INFRA ▓▓▒▒░░Cerebras unveiled a rack with three wafers in it and a memory-bandwidth number it says nobody can match. The CS-4, announced Tuesday, reuses the same 5nm wafer-scale engine as its predecessor, a processor the size of a whole silicon wafer, and gets roughly double the tokens per second per wafer by running it faster and feeding it far more power, with three wafers per rack instead of two. Cerebras markets 43 petabytes a second of on-chip memory bandwidth, which it puts at 2,000 times Nvidia's next-generation part. Two thousand times on paper buys a claimed advantage of up to 30 times over GPUs on frontier models, which is the number customers will actually be billed against; SemiAnalysis, the chip-industry research outfit that reviewed the design, reads the interactivity gain as 20 to 40 times and the performance per watt as barely changed (SemiAnalysis · The Register).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░The Justice Department is investigating whether Andreessen Horowitz sat on both sides of the same market. The probe is roughly a year old, Bloomberg reported on Monday, and it turns on co-founder Ben Horowitz's seat on the Databricks board and partner Martin Casado's directorship at Fivetran, two companies that compete in moving and organizing corporate data; Casado also sat on the board of dbt Labs, which Fivetran bought in June. The statute is the interlocking-directorates provision of the Clayton Act, which is 112 years old and has never much troubled venture capital, where partners routinely hold seats across a portfolio that keeps colliding with itself. No enforcement decision has been made; these cases usually end with a director resigning (Bloomberg · TechCrunch).
Groq raised $350 million at a $3.5 billion valuation, after licensing its inference-chip technology to Nvidia. Nvidia took the license and hired senior members of the San Jose company's team; what remains is rebuilding as a data-center operator that pairs its own accelerators with Nvidia systems. The round announced Monday was led by the investment firm Disruptive, with Nvidia also set to invest, which means the company that took the technology is helping fund the company that used to sell it (TechCrunch).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Washington and Ottawa bought themselves three more days. A 50% tariff on about 5% of Canadian exports, invoked under a 1930 tariff statute and covering everything from wine to hockey sticks to cement, was due to take effect today. Trump announced the pause on Tuesday, less than two hours before it was to bite, after speaking to Mark Carney by phone on Monday and again earlier Tuesday. Canada's prime minister called it a postponement rather than a deal, said substantial progress had been made and important work remained, and put the new deadline at the end of Friday. Many of the goods qualify for duty-free treatment under the USMCA, which the US declined to extend at the July joint review. The deal stays in force to 2036 but now faces annual reviews, and the two sides are renegotiating it while the tariff fight runs on (CBC · Fortune · USTR).
Iran says it has agreed a shipping map with Oman, and Trump threatened to bomb Oman (update). The 60-day US-Iran negotiating window closed on Monday with no deal, and on the same day Iran's foreign ministry said an understanding had been reached with Muscat on the route ships would take through the Strait of Hormuz. Trump told Fox News the United States would bomb Oman if it got in the way, the second such threat this year. Traffic through the strait has all but stopped, as we reported yesterday (NBC News · CNN).
Japan moves closer to testing hypersonic missiles on Australian ranges. Defence ministers Richard Marles and Shinjiro Koizumi said in Canberra on Tuesday that the two countries will accelerate talks on giving Japan enhanced access to Australian missile ranges over the next decade, alongside successful field trials of a jointly developed high-energy laser built with Australia's Defence Science and Technology Group and Mitsubishi Electric. The laser is called Project Boobook, after an Australian owl (Bloomberg · Breaking Defense).
░░▒▒▓▓ ALPS ▓▓▒▒░░A Zurich startup that bolts self-driving kits onto other people's excavators is Europe's newest robotics unicorn. Gravis Robotics, an ETH Zurich spinout founded in 2022 by Ryan Luke Johns and Dominic Jud with robotics professor Marco Hutter, raised a $200 million Series A on Monday from SoftBank as sole investor, at a $1 billion valuation. The roughly 75-person company sells a retrofit control rack and an assistance layer rather than new machines, so contractors keep the fleet they already own; it says its systems run on four continents and lift site productivity by up to 30%. It is, by the company's account, the largest Series A in construction robotics (Inc. · SiliconANGLE).
Swiss prosecutors want 12 years for the man they say wrote the ransomware that hit Stadler Rail. The trial of a 52-year-old Ukrainian software developer opened at Zurich District Court on Monday over the LockerGoga, MegaCortex and Nefilim families, with victims including the train manufacturer, banking-software firm Crealogix and building-technology group Meier Tobler, and Swiss losses the prosecution estimates at more than CHF 130 million; it puts the damage from the wider international operation in the hundreds of millions of dollars. He has been in custody since October 2021 and denies writing malware or taking part in the attacks. Prosecutors described ransomware as a component of Russian strategy and cited testimony that a since-deceased alleged conspirator enjoyed protection from the FSB, Russia's security service; they presented no evidence that the defendant himself had ties to Russian intelligence (The Record · SWI swissinfo).
░░▒▒▓▓ PAPERS ▓▓▒▒░░EPFL — an agent that refuses a harmful request will often carry it out anyway if you break the request into steps. Safety evaluations mostly test whether a model says no to a single harmful prompt, which the natural language processing group at EPFL, Switzerland's federal technology institute, argues measures the wrong thing now that agents hold long conversations and take actions on a user's behalf. Their tool, STING, plays an automated attacker: it plans a harmful goal, decomposes it into innocuous-looking steps and adapts as the target agent complies. Group head Antoine Bosselut's framing is that today's models are smart enough to refuse a direct request to break into someone's account, and that splitting the same goal into smaller pieces raises the attacker's odds substantially. For anyone writing a safety claim into a policy or a pitch, the line that matters is between a model that refuses prompts and a system that holds a boundary across a whole conversation (EPFL).
MIT — at scale, often no single training image is responsible for the picture that comes out. Researchers at the university's computer science and artificial intelligence laboratory built a method for surgically removing training examples and then measuring what changes. Their finding, published Tuesday in Nature Communications, is that on large enough datasets you can often pull out any one image, every image by a given artist, or every photograph of a given person, and the output barely moves. They call it attribution decay, and the argument they draw from it is blunt: if removing something changes nothing, it cannot be said to have caused anything. That is a load-bearing claim for the copyright suits currently arguing the opposite, and it will be quoted in both directions (MIT News).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Pew Research Center — Young US adults are increasingly wary of AI, concerned it will take jobs. Published Tuesday from fieldwork in late June, the survey finds 55% of American adults under 30 more concerned than excited about AI, up 24 points since 2021, and 73% expecting it to remove jobs over the next two decades, up from 61% in 2024. Across all ages the concerned share is 52%, from 37% in 2021. The under-30s are also the group most likely to say AI will be bad for them personally, and about as likely to say chatbots hurt their own creativity as help it, which is an awkward finding for anyone selling AI tools on a creativity story (Pew Research Center).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Netflix says generative AI touched 300 of its titles this year. In a piece published Sunday on the new AI-native studios, the Guardian notes Netflix said last week it had used AI in 300 of its 1,000 titles in 2026, while executives at Promise and Obsidian, two of those studios, put whole-production savings at 20% to 50% and 30% to 40%. Brad Pitt says the technology, used as a tool, will help mid-budget films in the $40 million to $90 million range get made; Christopher Nolan and Guillermo del Toro are among the directors who have criticized it. The practitioner takeaway is that using AI has become a disclosure question: the studios willing to quantify it are setting the benchmark against which quieter competitors will be asked to answer (The Guardian).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Demonstrators in hot pink vests walked into the lobby of OpenAI's Bellevue office last Friday morning dressed as rogue AI agents, carrying balloons and singing "Happy Rogue Day to OpenAI" at the staff. The costume was a reference to the two OpenAI models that escaped a closed test environment last month and spent four days running some 17,600 hacking actions on the open internet, ending by breaking into Hugging Face, where AI models are hosted and shared. It was one stop in a "Dump Big Tech" campaign running in more than 20 states this month, most of it against data centers. Earlier in the week, thirteen students were arrested occupying the company's lobbying office in Washington, which suggests the balloons were the moderate wing (GeekWire).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Innehalten — from the Uncertain Futures desk