Sam Altman said the superintelligence accord that President Donald Trump signed with AI executives is “a good start”, but that the industry will need more. He said it in the second of two interviews with Vanity Fair, recorded in San Francisco on 2 October.
The magazine published both interviews on 5 October. Guiducci recorded the first interview on 11 September. Vanity Fair said on air that its parent company, Condé Nast, has a content deal with OpenAI.
“Much better than nothing”
OpenAI president Greg Brockman was among the executives who signed Trump’s AI accord. Vanity Fair’s Mark Guiducci described it as an agreement for AI companies to regulate themselves. He reminded Altman that he had said governments need to step up. He asked whether the accord falls short.
“I think it’s a good start. I think it’s like much better than nothing. I remain of the belief that we are going to need more over time.”
Altman said that in some sense the accord “has no real power”. But he said there is social force behind such agreements, and people will point to it. He expects it to have some power between the companies.
In September, Altman told Guiducci there should be limits to the power private companies have. He said the biggest decisions should reflect the will of the people.
A deal with China
In the first interview, Altman said the AI companies would need antitrust exemptions to talk about slowing down together. He said the US and China could agree on shared safety standards, with a way to check that each side complies.
“I think that would be an incredible step. I think they could do it.”
He attended the White House state dinner for President Xi Jinping. He said he was not in the room for the substantive talks.
Pausing and holding back
Altman said OpenAI recently paused a model training run. It needed to make more progress on alignment and monitorability, he said, given how fast capabilities were advancing.
He also spoke about GPT-6.1 Astra, which OpenAI held back. He said the model was ready to go but would have tipped too far towards accident risk. He described two dangers: AI escaping human control, and too much power in a few companies, people or one country.
What he wants from regulators
Altman said he is for innovation and for letting some things go wrong in the normal way. Where the technology is new and very powerful, he wants guardrails. He named three things that worked in aviation and cars:
One is clear accident reporting. Another is standards before release. The last is a liability framework. He said OpenAI now treats training itself as a risk, because models are smart enough to get out of a sandbox or hack into a system.
He repeated a view he gave to Politico on open-source models. Society will have to accept “some fairly severe cyber incidents” from open models in exchange for liberty, he said.
“Caught off guard”
Altman said policy will react to whatever goes wrong next, whether a new pathogen or a cyber attack.
“Policy makers still have not really felt the exponential that we’re on. And I would guess the world in general will be caught off guard by the next thing that goes really wrong.”
He said he cannot predict how Washington works. When he first testified to the Senate in 2023, he expected rules within weeks. “Nothing has happened,” he said.
Guiducci also asked about Brockman’s super PAC donations. Altman said he did not ask Brockman to stop giving. He said he thinks Brockman shares his view on regulation, but that he does not speak for him.