Altman Would Rather Have a Trillion Super AIs Than One AI with a Trillion IQ

OpenAI CEO Sam Altman appeared on Cheeky Pint, the podcast hosted by Stripe co-founder John Collison. He drank his first Guinness and had a 52-minute conversation. The video is below, and links to the original video and transcript are at the end.

The episode starts with PHP vulnerability week and goes on to discuss the early internet, Cyber Blue, sandboxes, a math weekend, agent swarms, compute shortages, personal agents, and AI deflation. In the rapid-fire segment, it also touches on benchmark gaming, Greg Brockman, and the idea that tokens are a fake unit.

In the PHP vulnerability week segment, Sam did not overstate his case. He first explained the problem: models have been getting steadily better at finding vulnerabilities, and not slowly, but until now few people had used them to scan all the code on the internet. The code surface is enormous, and once scanning begins, more serious vulnerabilities may be found than expected. Fixing software is of course necessary, but he also mentioned changing architecture—difficult, yet more durable. Two years ago, saying the entire stack should be rewritten would have sounded insane; now some people are starting to take that seriously. He even suggested that a new system, network, and application model will emerge in the future. The more notable point is that a global software upgrade has gone from a joke to an engineering problem.

The early-internet analogy was concrete. Sam said that seeing an 8- or 10-year-old build a complete 3D game with these tools still feels wild; adults would be less likely to think that way. This generation's expectations for what technology can do are high from the start, and they will not be shocked by making something in 10 minutes. On the other hand, there is an internal OpenAI joke: during one SEV, Codex went down and everyone suddenly forgot how to code. Dependence has reached the point where no further explanation is needed.

In the Cyber Blue part, he said the first day went better than expected, but he explicitly refused to declare victory. He does not think the PHP matter is clear yet, and he assumes bad things will happen. One security expert noted that the real question is whether the assumptions behind the threat model still hold. If AI can find hardware vulnerabilities that require new physics to understand, the old reasoning loosens. Sandboxes fall into the same category of problem. The host said that talking with people who know OpenAI and similar companies well reveals a kind of pessimism: sandboxes may not be able to contain superintelligence. Sam first stressed that much better can be done than the status quo, and that better safety models should be the focus. Then he pushed the question to an extreme: facing a theoretical AI a million times smarter than the best security researcher, would you trust that sandbox? That framing is more concrete than asking whether AI will escape.

The math weekend is another detail. The source material says OpenAI ran internal experiments using tens of millions of dollars' worth of compute to prove complex math problems, citing Navier-Stokes. When Sam talked about that weekend, there was excitement but also wistfulness. Agent swarms produced the most quotable line of the whole episode: he would rather have a trillion AIs slightly smarter than humans join society than one AI with a trillion IQ. That preference only becomes complete when seen alongside compute, cost, and personal agents.

The compute shortage segment lands on custom chips, self-replicating data centers, and why cost per gigawatt should fall. The show also discusses why AI brings deflation, which is on the same line as compute costs and token pricing. One commenter said the cost-per-gigawatt part was what they most wanted to hear. Whether data centers can be built cheaply and keep getting power will directly determine how much compute can run later. In the personal-agent part, he mentions that traditional web forms and mobile apps will be replaced by personal agents, and that many fake jobs that exist around forms will disappear. In the rapid-fire segment, tokens are called a fake unit—an ending that fits the podcast's pub atmosphere.

The original video is on YouTube, 52 minutes 45 seconds long: https://www.youtube.com/watch?v=iMsUcCYNKqY

Cheeky Pint's Substack has the full transcript: https://cheekypint.substack.com/p/why-sam-altman-is-against-a-single