Wednesday, 30 September 2026
Signed, Renamed, and Morally Binding
The White House Renamed the Technology, Then Asked It to Police Itself
Two separate documents came out of the same Tuesday event at the White House, and they are worth keeping apart. The first is an executive order directing every department and agency to stop writing “Artificial Intelligence” and “AI” in official correspondence, public communications and policy documents, and to write “Super Intelligence” and “SI” instead — with the Assistant to the President for Science and Technology tasked to go away and propose a federal definition of the new term. The second is a one-page accord, the Joint Commitment on Frontier Responsibilities, signed by Trump alongside Sundar Pichai, Mark Zuckerberg, Dario Amodei, Greg Brockman, Elon Musk, Jensen Huang, Satya Nadella, Lisa Su and Nikesh Arora. It commits them to four layers: internal controls monitoring their models’ capability and alignment during training and deployment, across cyber, bio and chemical risk; an internal team checking that those controls work as intended; an independent external auditor doing the same assessment from outside; and an independent board committee overseeing the whole chain and receiving its reports. Nothing in it is enforceable. Trump said he thought it was “morally binding” and praised the “tremendous self-policing” in front of him, and the text itself defers the question, allowing that “over time, it may make sense to codify these steps into laws or regulations.” Senator Mark Warner’s summary is the one worth writing down: “The president’s response? To rename it and tell the companies developing it to regulate themselves.” Four layers of oversight is a genuinely more detailed structure than anything the industry has signed before. All four of them report to the companies.
Source: CBS News, September 29, 2026
OpenAI Held Its Developer Day the Same Morning and Got Buried
Some genuine scheduling misfortune: OpenAI’s DevDay landed on the morning of the White House signing and got buried under it. The headline product is Dots — persistent, always-on agents inside ChatGPT with names and avatars you choose, running on GPT-6 Astra, reaching some four thousand apps through plugins, and going to Pro 200 and Business Premium subscribers first. That is the same shape as Meta’s Muse and xAI’s Grokbot, and the category now has enough entrants to need a name. The more interesting number is GPT-6.1 Sol, which Sam Altman pitched as “near-Astra level intelligence at a fifth of the price”: $2 per million input tokens against Astra’s $10, $10 output against $50, and cached input at ten cents against a dollar — a 90% discount on the part of the bill that agent work generates most of. Also launched: Ultrafast, a tier that runs up to eight times quicker in Codex at around 300 tokens per second and six times in the API, for six times the standard rate. And a small, telling adjustment underneath all of it — new Pro 200 subscribers now get ten times the usage rather than twenty. The price of the frontier fell by four-fifths this week. The price of unlimited did not.
AMD Bought a Company That Builds Worlds, One Day Before Signing the Safety Accord
Lisa Su had a busy forty-eight hours. On Monday AMD announced it is acquiring World Labs, Fei-Fei Li’s spatial-intelligence startup, in an all-stock deal worth about $8.2 billion — its largest acquisition since Xilinx — expected to close by the end of the year subject to regulators. World Labs builds models that generate, reconstruct and simulate interactive 3D environments from text, images and video, plus the simulation layer robots learn in. Li joins AMD as executive vice president and chief scientist, reporting directly to Su. “Building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving,” Su said, which is the honest version of the strategy: Nvidia has spent two years selling the argument that the chip company closest to the models wins, and AMD has now bought a research team rather than argue the point. Note what the purchase is of. Not a chatbot, not a coding agent — a company whose products are physical spaces that don’t exist, made to order, for machines to practise moving through. On Tuesday, Su signed a document committing to monitor her models for cyber, bio and chemical risk.
Source: AMD, September 28, 2026
OpenAI Wants $30 Billion and Is Giving Safety as the Reason It Won’t Go Public
Yesterday this section covered OpenAI shelving GPT-6.1 Astra on safety grounds. The new part is that the same reason is now being offered to the markets. OpenAI is seeking close to $30 billion in a new private round at a valuation of roughly $1.4 trillion, after Sam Altman put the public listing back: “Right now would be an ill-advised moment to go public, and we don’t feel pressure on that,” he said, and, asked when — “I would say not 2026.” His stated grounds: “We have a lot of work to do, including meeting the safety and alignment requirements of this moment.” The business behind that sentence is not struggling. Annualised revenue has passed $40 billion, roughly double where it ended 2025, and Greg Brockman put the July run-rate increase alone at over 20%. In March the company raised $122 billion at an $852 billion post-money valuation; six months later it is asking for a quarter of that money at a valuation 64% higher. An IPO would bring quarterly disclosure, an investor base that can sell, and a legal duty to explain the October you cancelled a model. Private capital asks for none of that, and this week it is offering $30 billion not to.
Anthropic Is Interviewing Its Users Again, and This Time They Can Publish the Transcript
A smaller item, and a structurally odd one. Anthropic has reopened “What Do You Want from AI?”, a study in which an AI system called Anthropic Interviewer conducts roughly fifteen-minute interviews with Free, Pro and Max users whose accounts are at least two weeks old. It runs from 29 September to 6 October. The December 2025 round drew 81,000 participants and its results have since been quoted in the places where AI policy gets discussed. What’s new is the release valve: for the first time, a participant can choose to make their own full interview public, published with their country attached, instead of trusting the aggregate. That matters more than it sounds. Every previous version of this exercise produced a summary written by the company that commissioned it, and the only way to check a summary is to read what it summarised. Now some of that will be readable. An AI interviews you about AI, and the transparency mechanism is that you get to keep a copy.
Source: Anthropic, September 29, 2026