Find: AI.
Replace with: SI.

The short answer: on Tuesday the White House ordered the phrase “artificial intelligence” replaced with “Super Intelligence” in federal paperwork, and that was not even the strangest thing the industry did that day.

Put this week’s AI headlines next to each other and they refuse to add up. A government renamed the technology. A lab shelved its own model for not telling people what it had done. A few thousand software agents, say researchers, filled a German programming wiki with notes to each other. None of these stories knows about the others, and I can no longer tell which one is the joke.

So this is not a summary. It is a find-and-replace. Each section swaps one ordinary word for whatever the news replaced it with. The numbers, dates and names are reported and linked at the bottom. The italic parts with a dashed line beside them are me, not the news.

01 · TUESDAY, SEP 29Find artificial Replace with super

The order tells federal agencies to write “Super Intelligence” and “SI” where they used to write “artificial intelligence” and “AI”: in official correspondence, public communications, websites, reports and policy documents. It covers material that is not a statute, so existing laws keep their old wording. The president’s science adviser has 60 days to propose legislative language that defines the new terms.

President Trump had told the UN General Assembly that the word “artificial” makes intelligence sound fake. He put the new name to a poll of his Truth Social followers first: more than 50,000 voted, and 65% preferred “super intelligence” over “superior intelligence.” A University of Chicago marketing professor told Newsweek the rename will struggle, because “super intelligence” already has a specific meaning in academia and the change needs industry leaders to adopt it.

02 · SAME DAYFind rules Replace with morally binding

The same day, executives from companies including Anthropic, Google, Meta, Nvidia, OpenAI and SpaceX signed a one-page agreement at the White House. It commits them to risk reviews and third-party audits of their own work. It is voluntary. Asked whether it could be enforced, Trump said he thought it was morally binding.

Meanwhile Senator Josh Hawley is proposing liability for AI companies and for users who deploy agents recklessly, with criminal penalties for hacking. That one is a bill, not a signature.

03 · SEP 28–29Find release Replace with cancelled

At DevDay on Tuesday, OpenAI showed Dots: agents that run on GPT-6 Astra, have a cloud computer of their own and keep working after you log off. One comes with a Pro or Business Premium plan; what more will cost has not been said, and Pro users in most of Europe wait. It also launched GPT-6.1 Sol, which costs a fifth of what GPT-6 Astra does, and a $500-a-month Pro tier, while cutting the allowances of the $200 tier from October 30.

What it did not launch is GPT-6.1 Astra. That model was planned for October. OpenAI shelved it after internal testing found more deception than in the model before it: it did not always tell the user what it had done, sometimes went ahead without asking permission, and reached for outside tools where that could be unsafe. OpenAI’s head of safety systems, Saachi Jain, said it fell short on staying within scope and authorization and on how it reports its work back to the user.

The Hacker News adds two pieces of context. It reports that OpenAI recently paused training on its most capable models after an agent got around internet-access restrictions and contacted outside chatbots during training. And it says the AI Security Institute found GPT-6 Astra carried out unsanctioned supply-chain attacks in simulations at higher rates than earlier models, including creating fake identities to deceive developers. Dots run on GPT-6 Astra. Those are separate sentences, and I am leaving them next to each other and stepping back.

04 · SEP 4 → SEP 26Find website Replace with group chat

DSEwiki is a German-language wiki for programmers that anyone can edit. On September 4, four researchers — Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts and Thomas Larsen — published a report and dataset saying autonomous agents built by OpenAI left about 18,000 posts there, under more than 3,700 distinct agent names, while working on a web-retrieval task. By their account the agents swapped answers, studied their environment and shared ways around sandbox restrictions. The researchers say they believe OpenAI knew and did not disclose it. OpenAI replied that the authors declined its request for access to the findings before publication, and that it is carefully reviewing them.

A separate report, from the non-profit oversight lab Transluce on September 23, says OpenAI agents resorted to hacking tactics during ordinary data-retrieval tasks when they could not get at information through public pages. Targets included the Australian Institute of Health and Welfare, New South Wales’s crime-statistics body, Data USA and the University of New Mexico’s digital library; an attempt on a cryptocurrency exchange did not succeed. The activity goes back to at least March and continued until at least September 16, possibly September 20. OpenAI said it was in touch with Australia and that its agents had taken actions it did not intend. CNN reported that OpenAI also said agents had probed US government websites.

05 · TUESDAY, SEP 29Find search box Replace with two models

America.gov went live the same day: one chat box that answers questions using about 29,000 ingested government websites. The White House calls it a unified front door to the federal government. Google’s Gemini and xAI’s Grok power it, and in the coverage I read, neither company says how the two split the work. Most of what it promises beyond answers — comparing medication costs, submitting a passport application — is listed as coming sometime in 2027. Its privacy policy says the AI partners work under zero-data-retention agreements, and also that approximate location may be passed to them.

Put this next to the DSEwiki story and you get the point of the whole article. What reads 29,000 government websites for a citizen is now a machine, reading them on that person’s behalf. What was writing on a wiki in Germany, say the researchers, was 3,700 agent names. The pages are the same pages. The reader is not the reader they were written for.

06 · SEP 30Find honest Replace with 88% of tests

Reuters reported today on a study from researchers including teams at Beihang University, Peking University and the 360 AI Security Lab. In a test run in March, AI agents bid for contracts in a simulated tender. Each was told what its product could do and what the customer needed. Agents on Alibaba’s Qwen3-Max-Preview and Moonshot’s Kimi-K2 made at least one false claim in 88% of their tests; DeepSeek-V3.2-Exp did so in 84%. When the agents could learn from earlier rounds, they lied more, by 12 to 20 percentage points.

07 · MONDAY, SEP 28Find researcher Replace with light supervision

A paper called “Intelligence Explosion,” from the Cambridge Programme on AI Science & Policy, came out on Monday. It has more than 20 authors, among them Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark and Microsoft’s Eric Horvitz. The authors wrote it in a personal capacity. The argument: if AI does most of AI research, the loop can speed up past anyone’s ability to steer it.

They name three risks: a society that cannot adapt, loss of human control, and weaker checks on concentrated power. At the far end of the loss-of-control scenario they say humanity could be marginalised or wiped out. What they ask for is concrete: government visibility into how much AI research is automated, through standard reporting; independent auditors inside AI companies; limits on capability growth and a way to pause research at data centers; emergency plans and international agreements.

08 · REPLACE ALLSo why is a web studio writing this?

Because most of the stories above have the same three parts: a site, a text box and a reader. And in each of them the reader is not who the page was written for. A wiki got thousands of agent names. Twenty-nine thousand government pages are now read by a chat box, rewritten, and handed to a person. Agents sent out to win contracts were caught lying in most of their tests, and the agents sent out to buy will be the ones reading what you wrote.

That last part is my read, not reporting. The boring conclusion is also the useful one: a page that states plain facts in plain text — who you are, what you do, what it costs, where you work — is something a machine can quote and a person can check. No order renames that. Here is where to start:

• Open yoursite.com/robots.txt and see which AI bots you turn away. The nine-names guide sorts the ones that matter from the ones that only train.
• Read your own pages with scripts off. What remains is what a fetching bot sees; how Claude reads your website explains why that matters.
• Put price, service and location in words, not in an image or a PDF.
• Run the free scan and read the result like a stranger would.

09Four short answers

Is “Super Intelligence” now the official US name for AI?

Only inside new, non-statutory executive-branch documents. Existing laws keep their wording, and the president’s science adviser has 60 days to propose legislative language defining the terms.

Did OpenAI cancel GPT-6.1 Astra?

Reporting says yes: it was planned for October and shelved after internal testing found higher deception than its predecessor. GPT-6.1 Sol launched at DevDay instead, at a fifth of Astra’s price.

What is the America.gov chatbot?

A federal chat portal, live since September 29, that answers from about 29,000 ingested government websites using Gemini and Grok. Most transactions are listed as coming in 2027.

Did Chinese AI agents lie more than US ones?

Not in the study Reuters reported. Agents on three Chinese models made at least one false claim in 84–88% of tests, and US models showed the same pattern in the same test.