Sam Altman told Axios the Trump administration's voluntary review of OpenAI's forthcoming Astra model was a "productive process." [1] "But we of course did it." [1] Maria Curi filed that confirmation on Thursday from Wednesday's tape at the G20 Innovation Ministerial in Chapel Hill, where Commerce Secretary Howard Lutnick interviewed Altman at lunch. [1] The government did not certify Astra. The company sat for a review it did not have to sit for, of a model it has already placed at a "critical" cyber threshold, under a framework the White House does not plan to publish.
Tuesday's Astra file held the product: a shipping model that can hunt zero-days without a person at each step, gated to a customer whitelist and a monitor that will also halt innocent work. Wednesday's Lutnick file was the executive declaring Anthropic "back on the right side" because the company did what they asked. Thursday is Altman confirming the Astra review happened. Those are not the same receipt as a public rule.
Axios said OpenAI is close to releasing Astra, the first model it has designated at its critical cybersecurity capability level, which the company says requires stronger safeguards during development and before release. [1] The administration has developed a voluntary framework for reviewing some advanced models before release, including potentially giving officials access for up to 30 days. The White House does not plan to release that framework publicly. [1] Opacity is the design. Labs inside the process get a private window. Companies outside it, and the public, get a lunch quote.
Altman told Axios that as models reach "a quite significant level of capability," close engagement with safety institutes in the United States, the United Kingdom, and elsewhere "will become more important." [1] He called Astra "a significant step forward." [1] On stage with Lutnick he cast OpenAI as "the pragmatic centrists," with people, not AI, running the future. [1] To Axios he attacked "doomerism — for the sake of trying to have more power, control or enforce a worldview" and, in the next breath, "blind optimism without an acknowledgment of the risks." [1] Jensen Huang, at the summit breakfast, warned ministers that "the worst outcome is that you don't take advantage of it [AI], that you are left behind." [1] Fear and hurry shared a building. Neither produced a statute.
Altman said he expected Congress to set a "basic framework" after his 2023 testimony. "I thought at that time that Congress was going to do something." [1] He blamed the pace of the technology, and added that "it appears to be very hard to get Congress to do big things." [1] Lutnick's one job in this file is the chair. Wednesday he told Axios he trusts Anthropic because they did what they asked. Today's heavier hand, such as it is, is private access to a critical model, not a public rule. Do not write that the government signed off. Take Altman at the sentence he offered. They of course did it. The of-course is the tell.
-- THEO KAPLAN, San Francisco