It's Monday, August 31st: OpenAI told SpaceX it is winding down the contract that puts OpenAI models inside Cursor, two weeks after SpaceX closed its $60 billion purchase of the company, and Anthropic opened a research preview of a standard that lets AI agents discover and operate physical lab and factory machines.
Covering what’s happening on the ground in AI, every Monday.
1️⃣ SUPPLY CUT: OpenAI Is Pulling Its Models Out Of Cursor, And The Reason Is Elon Musk
Image from Tipranks
OpenAI notified SpaceX on Friday that it intends to wind down the contract providing OpenAI models to Cursor, with a proposed shutoff date of November 12, 2026, two weeks after SpaceX completed its $60 billion acquisition of Cursor's parent company.
OpenAI published the decision itself. It says it is giving the maximum notice its contract allows, and that it will not provide future models to Cursor during the wind-down. The stated reason: "We are making this choice because we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts."
Under oath earlier this year, Musk admitted that xAI, now also part of SpaceX, had distilled OpenAI data in violation of OpenAI's terms. A change-of-control clause in the Cursor agreement gave OpenAI a limited window to cancel, and it used it.
Cursor CEO Michael Truell put the exposure at about 5% of user traffic and said the two companies are talking. SpaceX completed the $60 billion acquisition on August 14 according to its filings, having already absorbed X and xAI in February. Musk's public response on Saturday was that he "couldn't care less."
Anthropic Co-founder Tom Brown posted that Anthropic will keep increasing compute for Claude models in Cursor, which drew immediate reminders that Anthropic itself limited Windsurf's direct access to Claude in June 2025.
If you build on someone else's model through someone else's product, you now have two counterparties who can each end your access for reasons that have nothing to do with you. Five percent of traffic means most Cursor users will not notice, and the notice period runs to mid-November. Treat it as the reminder to keep a small eval you can rerun the day a model disappears from your picker, so switching is a Tuesday and not a quarter.
2️⃣ MACHINE CONTROL: Anthropic Gave AI Agents A Standard Way To Run Physical Equipment
Image from Anthropic
Anthropic opened a research preview of the Model Hardware Standard, a shared specification that lets AI agents discover and operate microscopes, liquid handlers, robotic arms and lasers in parallel, cutting integration work that takes labs weeks down to hours.
Anthropic built it with HHMI Janelia, where researcher Arco Bast wrote a shared memory dictionary to get mismatched microscopes talking to each other. Agents reach devices through MCP, the command line, or code files, with auto-discovery across a network so nobody has to write a custom translator.
Carnegie Mellon connected an instrument in eight hours that vendor setup had taken several weeks to do, and the University of Washington's Baker and Pinglay labs wired up six instruments in under a week. Anthropic puts typical integration at weeks, if not months.
Operators tag machines in plain language, writing down physical specs and safety rules, or answering an AI interview, so an agent can read the file before touching hardware it has never seen.
Genentech, QuEra and Tetsuwan Scientific are testing it, while AWS, Tecan, Automata, Doosan, Danaher, QIAGEN, Universal Robots, Hugging Face (LeRobot) and Raspberry Pi build driver support. It does not yet work with hardware that has no programming interface, and Anthropic says Claude still misses physical constraints, including the physics behind bubble formation in a failed run. Waitlist preview is open at modelhardwarestandard.com .
MCP went from an Anthropic spec to a thing most of the industry supports in about two years, and this is the same play aimed at machines instead of software. The detail worth acting on is not the microscopes, it is the tags. The standard wants the knowledge that currently lives in a paper manual and in the head of the one person who knows the rig, written down in plain language. If you run equipment, that write-up is useful the day you do it, and it is the input this needs later.
📰 Other Headlines
OPENS TOMORROW: The AI Collective's global civic hackathon Hack for Humanity starts September 1 across 120+ chapters in 50+ countries and runs through October 31, building on job displacement, technical literacy, and accessibility.
LIMITED WINDOW: More than 100 companies and organizations, including OpenAI, Anthropic, Google, Microsoft, Amazon, Cloudflare, CrowdStrike, Visa, Mastercard and General Motors, signed an open letter warning that AI-enabled cyberattacks will become far more widespread in the coming months, and naming hospitals, water treatment plants and the infrastructure behind the internet as the exposed edge.
MYSTERY SOLVED: Z.ai confirmed that the anonymous Ox Alpha we covered last week is GLM-5.3-Flash, a 320B mixture-of-experts model with 18B active parameters, a 1M-token context and an MIT license, which Z.ai says was served entirely on Chinese AI chips.
FORTY SECONDS: Google shipped Gemini Omni 1.1 Flash, which reads up to 10 seconds of prior footage when you extend a shot and stacks extensions to a 40-second cap, adds first-and-last-frame control, and renders 360p drafts up to 60% faster at a third the cost. The 4K is upscaled, not native.
NEVER SLEEPS: WIRED found a "persistent mode" in Codex's public code that keeps an agent working until it is put to sleep, alongside a proactivity feature that has it generate its own follow-up tasks and message you unprompted. OpenAI confirmed it is testing and said there are no immediate launch plans.
LABELS BUY IN: Universal, Sony and Warner all took equity in Stability AI's $76M Series B, alongside Electronic Arts and AMD Ventures, in what Music Business Worldwide reports is the first time all three majors have invested in the same AI company. That brings the total raised under CEO Prem Akkaraju to $232M.
770 BILLION: Tencent open-sourced Hy4 Preview, a mixture-of-experts model with 770B total parameters, 49B active per token and a context window past 1M. It is free for two weeks inside Tencent's WorkBuddy and CodeBuddy, and reachable by API through Tencent Cloud TokenHub and OpenRouter.
TWO INBOXES: ChatGPT and Codex now support multiple Gmail, Google Calendar and Google Contacts accounts, so one conversation can read a personal inbox while another handles a work calendar. Availability depends on your plan and your organization's settings.
🎓AIC AUTOMATIONS
🪞 Ask It What It Knows About You
You have spent months telling an AI what to do. You have never asked what it noticed about you while you did it.
Paste this into an AI that can actually use your saved memory or your past chats. (A sidebar full of old conversations does not mean it can read them, so check that memory is on before you run this.)
Based on what you actually know about me from our conversations, answer three questions. Be specific rather than flattering, and label anything you are guessing at.
1. What patterns have you noticed in how I work that I have never explicitly told you?
2. What is the oldest thing you remember about me, and what have you probably lost track of?
3. If you could reorganize what you remember about me, what would you keep, what would you compress, and what would you throw away?
"Label anything you are guessing at" is the whole trick. Without it you get a horoscope. And the answer is only as good as what it can reach: with memory off, it sees only the chat in front of it, and it will be confident and empty.
What came back for me. Fair warning: mine was not a chat window. My setup reads my own written work notes, so it had more to go on than most accounts will. Three answers landed anyway.
"You approve by moving on." True, and I had never said it. When I go quiet and start the next thing, that is a yes.
"You do not start sessions, you resume spirals." It had counted: 46 working sessions in under three weeks, seven of them on one day.
"I have no way to represent what you were excited about." A confession. It knows what I built and what broke, not what felt right.
Then the part that makes this more than an interesting afternoon. Rewrite what is true as an instruction. "You approve by moving on" is an observation. "When I go quiet and move on, treat that as approval and keep going" is something it can follow. Save the two or three that survive into wherever your AI keeps standing instructions.
Image from Anthropic
There is also a button. In July, Anthropic shipped a reflection dashboard: Settings in Claude on web or desktop, in beta, on Free, Pro and Max, and it needs memory turned on. Pick a window from the last 1, 3, 6 or 12 months and it summarizes patterns in the Claude chats it includes: what you tend to work on, and how you tend to collaborate. It leaves out incognito chats, the underlying files from your connected tools, and anything touching a health integration. The dashboard tells you what you did. The prompt tells you who you have been to it.
Worth stealing. Same move, different angle:
"What do I ask you to redo most often, and what does that say about how I brief you?"
"Where have I been inconsistent with you?"
"What would you tell the next assistant I hire about working with me?"
"What do I keep saying I will do and never do?"
Run whichever one stings. Run it again in three months and compare.
Pro tip: a citation is not proof. Ask it to point at the conversation behind each claim. If it cannot, treat that line as a guess, which is not the same as a lie. If it can, that only proves the conversation happened, not that the conclusion was right. You are still the one who decides what is actually true about you.
One Agentic Tip a Week, from Joy Dong
🫵 Want your message in front of 200,000 AI builders?
Our partners and sponsors get exclusive placements across the newsletter and access to AIC's in-person network — demo nights, dinners, hackathons, and forums across 180+ chapters.
For all inquiries, send us a note at [email protected].
The AI Collective is built by volunteers across 180+ chapters in 40 countries.
Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.
🧑💻 About the Editors
About Noah Frank
Noah is a researcher, innovation strategist, and ex-founder thinking and writing about the future of AI and the workforce. His work and body of research explores the economics of emerging technology and organizational strategy. Outside of AIC, Noah heads research for Centaurian AI.
About Lindsay Gross
Lindsay is an AI engineer, researcher, and writer focused on how AI systems behave in practice and what it takes to make them safe. Her work sits at the intersection of AI safety, governance, and product design, and at AIC she writes about the questions that matter most as these systems scale.