Sam Altman says he does not want the AI future to belong to a single, enormously smart system. “I would much rather intuitively have a trillion AIs of slightly superhuman intelligence join society than one AI of 1 trillion IQ join society,” the OpenAI chief executive told Stripe co-founder John Collison on the Cheeky Pint podcast.
The argument is about safety, and it comes from an analogy to people. Altman said society holds together because huge numbers of people keep each other in check, and he linked that to pluralism and democracy. He had worried that AI would move away from that arrangement. Teams of cooperating and competing agents, he said, now make him hopeful again. He also asked that they be called “agent teams” because “swarms” sounds like locusts, a point Collison cheerfully conceded.
The reasoning gets sharper when Collison pushes on containment. Several people inside OpenAI and similar labs, Collison noted, doubt that a sandbox can hold a very capable model. Altman agreed the current security model can be far better, and said sandboxes work up to a point, with monitoring of the model’s internals extending that point further. Past it, everything depends on trusting that the model is truly aligned. He asked whether anyone would trust a sandbox built against an AI far beyond the best security researcher, a million times over. His own answer was that he does not know, and that the picture gets strange.
That is a notable thing for the head of a frontier lab to say out loud. Altman went further: for a while, maybe a long while, he thinks the industry will have to agree limits on how smart single models are allowed to get. In his telling, any slowdown in AI research would then come from that choice, not from models being unable to do the work. He gave no mechanism for how such limits would be agreed or enforced, and OpenAI is one of the companies that would be asked to accept them.
The interview also put numbers on how much compute is being thrown at hard problems. Altman said OpenAI let a team of agents run with few constraints for a week of math work, and that the Navier-Stokes proof the company announced cost on the order of $10 million. He added that the average result cost well under that, and bet that the same proof could be redone at one hundredth of the cost a year later with better models and methods. OpenAI, he said, deliberately avoids rewarding staff for burning compute and keeps no token leaderboard.
On hardware economics he used $70 billion as a round figure for the five-year cost of owning and running a gigawatt of AI capacity. Getting that to $20 billion would make everyone happier, he said, and would require cheaper energy, cheaper chips, and a squeeze on the stacked margins between raw materials and finished tokens. He called the tokens used to price AI “a kind of fake unit.” OpenAI came close to dropping token pricing for its next model, he said, but customers reacted so badly to a bill with no unit behind it that the idea died.
Taken together, the conversation shows an executive arguing for distribution over concentration, while his own company builds the most capable models available. The contradiction is partly resolved by Altman’s insistence on iterative deployment, putting tools in many hands early. It is not resolved by anything he said about who sets the ceiling on intelligence.
Companies building multi-agent products now have a stated design preference from the largest lab: many cheaper, bounded workers with oversight between them, and a budget line that will keep growing past the roughly 1 percent of operating costs that Collison said Stripe spends on tokens today.
Reported from the Cheeky Pint podcast interview of Sam Altman by John Collison; the source file carries no publication date.
