A model provider cancelled a major customer this weekend — not over money or misuse, but over who bought it. That is a new category of risk for anyone building on someone else's AI, and it arrived the same week two universities published opposite answers to what people should still be required to do themselves. Elsewhere, a third of American adults are now asking chatbots about their health, South Korea decided AI access is a public utility, ChatGPT started selling ads in its second-largest market, and humanoid robots ran faster than Usain Bolt while their own industry conceded that speed is not the problem.
OpenAI cancels Cursor over its new owner
OpenAI said Friday it is ending Cursor's direct access to its models on November 12, roughly two weeks after SpaceX completed its $60 billion acquisition of the coding tool. The stated reason is not performance, payment or policy violation by Cursor. It is ownership. OpenAI said it cannot be confident SpaceX will use its technology within the terms of service, citing its experience with Elon Musk's companies violating contracts — a reference to Musk's testimony under oath in an Oakland federal court that xAI used distillation on OpenAI's models to train Grok. The company said it is giving Cursor users as much time as its contract allows, and Cursor may take an earlier termination date if it prefers.
The practical blast radius is smaller than the headline. Cursor CEO Michael Truell put OpenAI models at roughly 5% of the product's traffic; the overwhelming majority of Cursor users are already on Anthropic's Claude or Google's models, both unaffected. Anthropic's Tom Brown said his company would increase compute to support Claude models in Cursor. Developers who want to keep using OpenAI models inside Cursor can supply their own API keys.
The precedent is what matters, and it cuts against the industry's own marketing. Every vendor selling you an AI-powered product is reselling capability it does not own, under a contract you never see, and that contract can now be terminated because of a corporate event at a company you have never dealt with. Cursor did nothing wrong by anyone's account. It was acquired. For a business standardized on a tool whose value depends on a specific underlying model, the lesson is not to pick a different tool. It is to know which model is underneath, whether the vendor has a second one wired up, and how long you would have to switch. Cursor got about ten weeks and had alternatives already in production. Not every vendor in your stack is that well hedged.
Two universities pick opposite futures
The University of Chicago's Social Sciences Core will go analog this fall. The division's memo bans AI-assisted writing for both students and instructors in its required Core sequences, and prohibits laptops, phones, smartwatches, wearables and recorded lectures in class. It goes further than most policies by binding the faculty too: AI-assisted grading has no place in the Core unless instructors carefully validate it against human grading, and syllabi are to be written by human scholars. Narrow exceptions survive for dataset analysis and citation research. The stated goal is to teach students to read, write and think without dependence on AI. Alpha School is running the opposite experiment at scale, expanding toward roughly 50 US campuses including 27 new locations, with students spending about two hours each morning on adaptive academic software before moving to workshops in coding, entrepreneurship and public speaking — though researchers speaking to Scientific American noted Alpha has not released enough evidence to separate the software's effect from the selection of its students.
MIT's own committee, reporting August 13 under co-chairs Eric Klopfer and Sam Madden, landed between them and supplied the sentence explaining why both positions are defensible. Its finding is blunt: current systems can produce credible solutions and reasonable responses to almost any written assignment in the undergraduate curriculum — essays, math and science problems, proofs and coding assignments. The committee's response was not a ban but a distinction between augmentation and automation. It also found only 25% of surveyed students felt MIT had adequately prepared them to use AI, while two-thirds expected it to matter in their careers.
Read as an operating question rather than an education debate, this is the same decision every business is making badly. UChicago is not rejecting AI; it is protecting the specific process through which a person produces evidence of their own thinking. A policy that starts by naming the tool expires with the next model release. A policy that starts by naming the human capability you intend to keep has a chance of surviving it.
A third of Americans are asking chatbots about their health
Pew Research surveyed 3,488 US adults between June 22 and 28 and found 34% now use AI chatbots for at least one health task. The breakdown is more telling than the headline: 28% get health information quickly, 25% try to work out what is causing symptoms, 22% because it is free or low-cost, 22% to understand a diagnosis a doctor already gave them, 20% to make sense of lab results, and 15% to decide whether to see a doctor at all.
Satisfaction is high and trust is not. Among people who use chatbots this way, 47% call the answers extremely or very helpful and another 48% somewhat helpful, leaving just 5% unimpressed. But only 29% are extremely or very comfortable sharing personal health details, and 26% are not comfortable at all. Usage skews sharply: 56% of Asian American adults, 44% of 18-to-29-year-olds, 48% of upper-income adults, and 60% of people who describe themselves as excited about AI, against 21% of those who are concerned.
The gap between "the answer was helpful" and "I will not give it my information" is the number worth carrying into any customer-facing AI product. People will use a tool they do not trust, right up until it asks them for something.
South Korea makes AI a public service, and OpenAI starts selling ads
South Korea's Ministry of Science and ICT selected three consortiums — led by SK Telecom, KT and Kakao — to run "AI for All," providing free, unlimited access to AI tools and services for every Korean citizen. The operators will share 512 Nvidia B200 GPUs in 2026, with the government funding operating costs from 2027. Agreements are due in September and the service launches by year end after beta testing, offering general chatbots and specialized agents built on both the operators' own and third-party models. Naver, the country's largest mobile platform, declined to participate, citing the timeline and a requirement to use competitors' models.
Running the other direction, OpenAI began showing ads on ChatGPT's Free and Go tiers in India on August 27 — its second-largest market, with over 100 million weekly users mostly on free or low-priced plans. The launch includes 50 brands with agency support from WPP and Omnicom, and an ad manager arriving next month with a minimum daily budget of ₹725, roughly $7.60. Ads sit below the answer, skip minors, and stay off health and politics. Dave Dugan, OpenAI's head of global ads solutions, framed it as letting businesses of every size introduce themselves at high-context moments when decisions are taking shape. OpenAI reported $6.7 billion in second-quarter revenue and has run ads in the US since February and Europe since August.
For small businesses, that ₹725 floor is the detail to file away: a conversational ad surface with a seven-dollar daily minimum, where the buying unit is an intent moment rather than a keyword.
Agents get cheaper by forgetting
A quieter result out of Google and Purdue may do more for the economics of agent deployments than any model release this month. The paper, SKILL.state, from Sanket Badhe, Priyanka Tiwari and Jonghyun Chung and accepted at EMNLP, attacks the reason long-running agents get expensive: the standard design appends everything an agent has seen and reasoned to a growing transcript, then re-reads all of it at every step. SKILL.state replaces that with an explicit, mutable execution state — at each step the model receives only the skill specification, the current structured state and the latest observation, discarding intermediate reasoning immediately. The authors report improved task accuracy alongside substantially lower cumulative token consumption across multiple datasets, models and environments.
It explains a bill. If you piloted an agent that worked beautifully on short tasks and became unaffordable on long ones, the cost was almost certainly the transcript, not the work — and that curve is being flattened inside the tools you buy, not the ones you build.
Authorship becomes an institutional problem
Investor Stanley Druckenmiller acknowledged using AI to write a Wall Street Journal opinion column criticizing Treasury Secretary Scott Bessent's bond-market intervention. He denied AI wrote the whole piece and argued the text expressed his own view; the Journal's editorial-page editor defended publication on the grounds that the argument was Druckenmiller's. That is a coherent standard — authorship as responsibility for an argument rather than production of every sentence — and it is not the failure mode that should worry you.
The harder problem is provenance: researchers identified 1,655 records in the Zenodo repository attributed to fictional authors, recurring invented people appearing across AI-generated papers and books, where repository metadata makes fabricated identities look legitimate. Moonbug, the studio behind Cocomelon and Blippi, meanwhile issued an internal policy that works as a template: AI for ideation, research, storyboards, generic backgrounds and refinement of human work; key characters, core plot twists and song lyrics reserved for people; prompts and AI use logged; and legal approval required before company intellectual property goes into any tool. Disclosure, fabricated identity and brand control are three different failures, and only one is solved by prohibition.
Physical AI
The most-watched robotics event of the week was also the most misleading. At Beijing's National Speed Skating Oval, across five days from August 22 to 26, more than 2,000 humanoid robots from 16 countries — including Germany, Japan and the United States — competed in 51 events. A machine from Beijing-based X-Humanoid ran the 100 metres in 9.39 seconds, beating Usain Bolt's 9.58-second human record, with a robot from smartphone maker Honor posting a faster 9.32 in trials. X-Humanoid's high jumper cleared 2.88 metres against a human record of 2.45. The sprints were fully autonomous rather than teleoperated this year, a genuine advance. Experts at the event were unsentimental about what it proves: humanoids remain overwhelmingly demonstration, performance and research machines, and mass deployment is still years out.
The industrial numbers underneath the spectacle grow fast without being large. Barclays research puts global humanoid installations at roughly 2,000 units in 2024, 15,000 in 2025 and a projected 60,000 in 2026, and robotics startups raised $18.8 billion in the first half of 2026 against $15 billion in all of 2025. Real deployments remain pilots with long runways: Hyundai and Boston Dynamics are targeting Atlas at the Savannah EV plant by 2028, and Agility Robotics has Digit pilots with Amazon and planned commercial deployment with Toyota at Canadian sites. The assessment worth taping to a wall came from Mikell Taylor, who directs GM's Autonomous Robotics Center: "80% success is failure in production."
The more consequential physical-AI development is not a robot at all. Anthropic's Model Hardware Standard research preview, which gives agents one interface for programmable instruments instead of a bespoke integration per device, published results specific enough to check. At QuEra Computing, an agent-built laser relock routine took success from 58% to 99.3% across 700 trials and cut recovery from 150 seconds to about six. Carnegie Mellon ran serial dilution experiments roughly three times faster, coordinating a liquid handler, plate reader, robotic arm and cameras across three incompatible computer interfaces, with integration taking eight hours instead of several weeks. The University of Washington's Baker and Pinglay labs integrated six instruments in under a week. Anthropic is unusually direct about the limits: Claude struggles with physical, chemical and biological constraints requiring real-world intuition, lacks spatial reasoning, does not work with non-programmable hardware, and needs expert oversight for anything safety-critical.
The robots that can run are not the ones that will show up in your operation. What changes the economics of physical automation is the collapse in integration cost — weeks to hours — applied to the programmable equipment a business already owns.
Quick Takes
Perplexity and Nvidia released Portable Computer, an agent that runs locally with no per-task fee, scoring 82.6% across 53 everyday work tasks — though the local model scored 59.6% on hard coding against 82.4% for its cloud counterpart.
Tencent open-sourced Hy4 preview, a 770-billion-parameter model activating 49 billion parameters per token with a one-million-token context window.
Z.ai open-sourced GLM-5.3, which the company says found 2,436 bugs across 269 open-source projects after post-training aimed at coding and security.
The EU AI Act entered its first transparency-enforcement phase on August 28, giving regulators access to company information and models, with stricter high-risk rules to follow.
WIRED found code in OpenAI's Codex for a "Persistent mode" where an agent creates its own follow-up tasks and works across sessions until put to sleep; OpenAI says it has no immediate launch plan and that the mode does not expand Codex's existing authority.
Originality.ai analyzed 2,034 religious books on Amazon and scored 63% as likely AI-written, with witchcraft titles highest at 78%.
Australia's chart body banned fully AI-made songs, effective with the August 31 chart — the first national chart anywhere to write that rule.
LinkedIn's "Seems like AI slop" button drew over a million clicks in three weeks, and views on flagged posts fell about 40%.
404 Media reported on an Amazon facility that cuts the bindings off books — some new, some from libraries or overseas — to scan loose pages for AI training.
Small businesses are going viral for advertising hand-made signage as "made without AI," turning an ordinary production process into a marketing claim.
What This Means for Your Business
Start by writing down which model is underneath each AI tool you pay for. The Cursor cutoff was not caused by anything Cursor did — it was caused by who bought it — and the notice period was about ten weeks. That is enough time to switch only if you already know what you would switch to. For each vendor in your stack, answer three questions this week: which model powers it, whether the vendor supports more than one, and whether your prompts, workflows and data would move with you. Vendors that route across several providers are worth a premium now in a way they were not a year ago.
Second, take the augmentation-versus-automation distinction out of the education story and apply it to a real process. UChicago did not ban AI because it fears the technology; it identified specific activities that exist because a person has to perform them, and protected those. Do the same exercise on one workflow: list the steps, mark which ones produce the judgment you are actually paying for, and automate everything else aggressively. Policies written around tools go stale in a quarter. Policies written around the capability you intend to keep in human hands survive model releases.
Third, if an agent pilot died because of cost, re-price it before you conclude it failed. The SKILL.state result points at the usual culprit: long-running agents that re-read their entire history at every step get expensive in a way that has nothing to do with the difficulty of the work. Ask your vendor how state is handled on long tasks. If the honest answer is "we send the whole transcript," the same workflow will be materially cheaper within a year, and the pilot you shelved deserves a second look rather than a post-mortem.
Fourth, the Pew health numbers are a template for any customer-facing AI you deploy. Ninety-five percent of people who use chatbots for health find the answers at least somewhat helpful, while under a third are extremely or very comfortable handing over personal details and a quarter are not comfortable at all. Assume your customers will happily use an AI assistant and refuse to feed it anything identifying. Design for the anonymous path first, make the data request explicit and late, and say plainly what happens to what they give you. Products that demand the information up front will show great engagement metrics and terrible completion rates.
Finally, if you run physical operations, ignore the sprint times and price the integration. A humanoid beating Bolt is a research result; the number that changes a capital plan is Carnegie Mellon coordinating four incompatible machines in eight hours instead of several weeks. Look at the programmable equipment already on your floor — the machines with a computer interface nobody uses — and get a quote on connecting them under a common control layer. That project was priced in months of custom engineering two years ago and is heading toward days, which is a very different payback calculation on hardware you have already bought.