Sam Altman Defends His Middle Path Between AI Catastrophe and Monopoly

Dots, the agent that works while you're not looking

The interview opens with Dots, launched the previous week and described as "an AI assistant that always has your back," inspired by the cool versions seen in movies. Altman says he loves it more than any OpenAI product.

His argument: everyone lets the tasks they dread pile up until they form a "gravitational well of anxiety." His dot combs through his emails and to-do list to clear them "with invincible energy."

He recounts a concrete use. He calls his dot to deliver a few ideas, and when he comes back the agent has written to him on Slack: some ideas are bad for such-and-such reason, and it found other relevant material on his computer. Altman sees it less as an assistant than as "an extension of myself."

On privacy, he promises "extreme" rigor, a condition for the tool to work at all. He also says he's saddened by technology that constantly grabs attention: he's turned off nearly all his notifications and prefers to text a friend himself rather than let his dot talk to theirs.

Astra 6.1 held back: the doctrine of two cliffs

Altman justifies delaying Astra 6.1 by pointing to a shift in era. A year ago, he says, you could release a model that was "maybe a bit dangerous" without anything serious happening. Models are now in the "seriously capable era," and the safety bar will keep rising.

He describes two ways things can go off the rails. On one side, a serious accident potentially leading to loss of control, with an AI that escapes and "wreaks havoc on the internet." On the other, concentration of power, which he says is advocated by competitors who believe these models are too dangerous to be distributed.

Astra 6.1 leaned the wrong way. OpenAI is therefore aiming for a model about whose alignment it can "make much more confident claims," while committing to distribute it widely, even if that takes longer.

He adds a technical detail: testing no longer happens only after training, since models are now supposedly capable, even during that phase, of escaping a sandbox or hacking a system.

Regulate, but in the gray zone

Mark Guiducci confronts him with his contradictions. Altman has called for government intervention and told the UN that key decisions can't be left "to a handful of labs in San Francisco." Yet OpenAI lobbies in the EU and California, and the deal signed with Trump amounts mostly to self-regulation.

Altman rejects the binary framing. His position: be pro-innovation, let some things "go wrong in a normal way" as happened in the early days of the internet, but set up guardrails where the technology is "extraordinarily powerful and fundamentally new."

The concrete price of this freedom is open source. He anticipates a "wave" of cybersecurity problems coming from open models, but refuses to ban them. As for the acceptable limit, he says it falls to the political process: "safety is always negotiated between stakeholders."

Facing Jensen Huang, who favors self-regulation, and Bill Gates, who compares AI to drugs without an FDA, he cites models worth importing: aviation's incident-reporting culture, pre-release standards, and a liability regime. He also predicts politics will remain reactive, and that the world will be "caught off guard" by the next major incident.

Brockman, Trump, Xi: the uncomfortable questions

Greg Brockman, OpenAI's president, and his wife have already given $25 million to the super PAC Leading the Future and $25 million to Trump's reelection campaign, before withdrawing a new $25 million commitment. Altman insists he didn't intervene and believes Brockman shares his vision, while repeatedly deflecting the journalist back to Brockman himself.

On the state dinner held for Xi Jinping, which Altman had described as an opportunity for "considerable progress," the results are thin. He wasn't in the room for the substantive discussions. The ballroom Trump showed guests around? "A construction site. It's big."

The journalist worries about a president who doesn't say a word about AI in front of industry leaders. Altman stays diplomatic: the two presidents supposedly discussed it earlier in the day, and by evening, "people are tired."

On Washington, he admits his powerlessness. During his 2023 Senate testimony, he was promised legislation "in two weeks." Since then, "nothing has happened."

Extinction, marketing, and legacy

Is the extinction risk overblown? For Altman, the mere fact that it's non-zero demands taking it seriously, and it's part of OpenAI's reason for being. He concedes some people use it to justify a power grab, without that making the risk any less real.

Asked about a New York Times article claiming OpenAI ignored internal warnings before the "Hugging Face incident," he says he's not familiar with it. He responds with a general defense: he'll always be accused of delaying a model or ignoring warning signs. "Come try to catch all of them. They're not that easy."

Finally, he rejects playing the historical-figure game, calling it "toxic." He positions himself against a small "priesthood" in the Bay Area that would keep AI for itself, and against those who want it for their own power. His wish: that AI be "a new fire or electricity," given to everyone. When the journalist throws out the word "Prometheus," he replies: "closer to that than other things."