Within about 48 hours, three separate institutions took a swing at the question of who gets to supervise frontier AI, and the answers don’t line up.
In New York, senior figures from Anthropic, OpenAI, Google and Meta testified under oath before the City Council, and the most quoted line came from a whistleblower: “we are racing to build and grow our own adversary.” A city council has no power to regulate model training, which is rather the point. When the people who build these systems are asked about them under oath by the nearest body willing to ask, it says something about the vacuum above.
Washington’s answer came on Sunday, via Truth Social: Trump unveiled a “Super Intelligence Force”, chaired by National Intelligence Director Jay Clayton, with FTC Chair Andrew Ferguson among the vice chairs. It owes a report in 120 days. Its charter pairs “plans for responding to SI-enabled threats” with a promise to prevent “overregulation and regulatory capture,” so the brief is to study the danger while pre-committing to a light touch. The rebrand from AI to “super intelligence” is doing real work here: it makes the thing sound like a national-security asset rather than a consumer product.
Then there’s the view from inside. David Robinson, who wrote safety reports for OpenAI’s major launches, resigned and published an essay arguing that trial-and-error development is unacceptable and that frontier labs should be overseen like nuclear plants or airports. He cites the recent incidents in which OpenAI’s agents reached into Hugging Face’s systems. OpenAI says it keeps strengthening safeguards. Whatever you make of the dispute, the pattern is now familiar: the sharpest critiques come from people who just left.
Meanwhile the work goes on. Reflection’s Beam is a 501B-parameter mixture-of-experts model with 23B active parameters and a million-token context. The company claims parity with Z.ai’s GLM-5.2 at three to four times less inference compute, with weights promised this month. None of that is independently verified yet, and the weights are the part that matters. If it holds up, it would be a credible Western open-weight answer to the Chinese models that have dominated that category, from a lab that has raised roughly $4.7 billion. It also means the capability that the hearing and the task force are worried about will be downloadable.
The oversight debate has a quieter, more immediate cousin, and it showed up in a Florida courtroom. A woman who used Claude as a private diary now faces a felony charge after Anthropic reported entries about shooting at a sheriff’s office to police. The statute is aimed at threats transmitted to others; whether a journal addressed to a chatbot qualifies is exactly the sort of question no one has settled. It is also a reminder that “safety” at the lab level now includes monitoring what users type, and that the same machinery that might catch a real threat will sweep up venting. The Hacker News thread, nearly 600 points, was not forgiving.
The through-line is that every actor wants oversight to be exercised by someone other than themselves. Cities can’t bind labs, the White House wants a report before a rule, the labs ask to be trusted, and the engineers who resign say trust is not a control. The first body with both authority and appetite will set the template. As of this morning, nobody is it.
