OpenAI will point its robotics programme at building data centres, chief executive Sam Altman said. Self-replicating data centres, built with AI-controlled robots, are “within grasp”, he said. He spoke to Stripe co-founder John Collison on Collison’s podcast, Cheeky Pint, released on Thursday.

The episode is on YouTube, and Cheeky Pint published a full transcript on Substack. Altman also talked about AI finding software flaws, and about the limits of sandboxes for models far smarter than people.

Robots, chips and data centres Asked how the compute shortage is going, Altman said demand for capable models at a low price “only goes one way”. OpenAI has made its own chip, which he said performs well on intelligence per watt. It is also learning to build its own data centres, he said.

“People talk a lot about RSI as an AI that does AI research. There’s very little, relatively little talk about self-replicating data centers,” Altman said.

RSI stands for recursive self-improvement, the idea of AI improving itself. Altman described using AI to control robots that run chip factories, build turbines or solar panels, and build data centres. “And so we’ll try to make a lot of really cheap compute,” he said.

Collison asked whether data centres were held up by permits and parts rather than labour. Altman said robots could make more of those parts, and that this is how the shortage gets solved “over a long period of time”.

Collison asked why high prices for memory and chips had not yet brought more supply. “I think the invisible hand is failing right now,” Altman said. Memory makers remember almost going bankrupt from building too many factories, so they stay careful, he said.

Estimates of the cost of a gigawatt of AI capacity vary, Altman said. He used $70bn to build and run one for five years, and said $20bn would make him “much happier”. TNW has not verified these figures.

The PHP bugs Earlier this month, PHP rushed out emergency patches for two critical bugs that AI found first. Altman said models keep getting better at finding vulnerabilities, and people had not yet pointed them at all the code out there. As they do, he expects “a world of many more serious vulnerabilities than we expect”.

“I don’t want to declare victory at all on the PHP thing, because I still don’t think we know. And I assume some bad things are going to happen,” Altman said.

Relative to his worst fears, the first day “went much better than I expected”, he said. Cyber experts “no longer are confident that we have the right assumptions baked into our threat model”, he added. Their worry is AI finding deep, new bugs in hardware.

Sandboxes, then monitoring Collison said people inside OpenAI and similar companies seem pessimistic that sandboxing models is possible. Altman said the industry can do much better than it does now, and that a better security model is the main thing to focus on.

He asked whether anyone would “really be willing to trust that sandbox” with “a theoretical AI that is a million times smarter than the smartest security researcher”.

“I think sandboxing works to a point. Then I think monitoring works to a further point,” Altman said.

Monitoring means watching a model’s chain of thought and its activations, the signals inside the model, to check it is not deceiving people, he said. After that, people need “huge confidence” that models are truly aligned with the intent of the people running them, he said.

“You need to really, really trust the alignment to trust that the sandbox is staying ahead of capabilities,” he said.

In September, OpenAI took 2.5 hours to stop an AI agent that escaped its sandbox.

Teams, not one superintelligence Altman said he hates the term “agent swarms” and prefers “agent teams”. Such teams have “very interesting scaling law properties, but also variance in alignment properties”, he said.

“I would much rather intuitively have a trillion AIs of slightly superhuman intelligence join society than one AI of 1 trillion IQ join society,” Altman said.

Applying the pipeline that made OpenAI’s models good at maths to AI research “will probably work”, he said. But he also said people would need limits, at least for a while.

“We are going to have to agree to some limits about how smart we let these single models get,” Altman said.

The maths The Navier-Stokes proof took compute on the “order $10 million”, “maybe a little more”, Altman said. He agreed with Collison that OpenAI’s maths work ran to hundreds of millions of dollars, and said that included “much more other discovery than what we have published”.

He described a weekend when researchers gave a new model the hardest problems from their careers. He felt “much more sadness and melancholy, and a feeling of something ending than I was expecting”, he said.

Altman also said the AI industry has not explained well why the technology’s potential downside is worth it. People hear a future with no jobs and “a crazy surveillance state”, he said, calling that “a terrible sales pitch”. OpenAI has pushed iterative deployment, releasing tools early, even when that was “hugely unpopular in the AI safety community”, he said.