In a Cheeky Pint interview released on October 29, 2026, OpenAI CEO Sam Altman said he would intuitively prefer a trillion slightly superhuman AIs joining society to a single AI with a trillion IQ. He sees multi-agent systems, which he calls “agent teams,” as closer to society’s resilience and diversity in their safety properties, but stressed that this direction must be actively pursued. Altman also said OpenAI used AI teams to prove the Navier-Stokes equation at a cost of about $10 million.
At a Glance Sam Altman discussed several frontier AI topics on the Cheeky Pint podcast. He revealed that AI models are rapidly discovering critical vulnerabilities in code such as PHP, and may uncover more in the future, necessitating a re-architecture of software systems. He mentioned that internal
In a Cheeky Pint podcast interview with John Collison, OpenAI CEO Sam Altman discussed the core philosophy of AI safety. He explicitly stated that, intuitively, he would prefer to see a trillion AIs slightly smarter than humans join society, rather than a single AI with a trillion IQ. This stance runs through his discussions on sandbox safety, agent teams, compute allocation, and scientific discovery. Altman believes that multi-agent systems (which he prefers to call "agent teams" rather than "agent swarms") not only exhibit interesting scaling laws in terms of capability but also more closely resemble the resilience and diversity of society in terms of safety.
John Collison pressed Altman on his view of AI's development direction—whether it should move toward centralization. Altman offered a very intuitive comparison.
> "I would much rather intuitively have a trillion AIs of slightly superhuman intelligence join society than one AI of 1 trillion IQ join society."
Basis and Reasoning:
1. Analogy of Societal Resilience: Altman believes that the characteristics of a multi-agent system closely align with the resilience and robustness of society. Although society has many deep flaws, it is extremely robust—this relies on a large number of individuals checking and collaborating with one another.
2. Pluralism and Democratic Values: He noted that society's robustness stems from values such as "pluralism and democracy." He once worried that AI might deviate from this direction, but is now excited about the multi-agent path.
3. Offense-Defense Balance: Collison added that multiple agents checking each other would create a dynamic equilibrium, with "offense-defense balance" serving as a good example.
Altman emphasized that this is not an inevitable trend—"All of these things are things we have to push for." The technology can be directed and deployed in many directions, and different people have different views on "the best version of the world," but he believes the multi-agent direction is the right one.
Collison notes that in his conversations with OpenAI insiders, he sensed a "pessimism" — the belief that sandboxing may be fundamentally infeasible. This ran counter to his initial intuition: he had assumed it was possible to build a "maximum-security prison" to contain the model.
Altman acknowledged the validity of this pessimism but offered a nuanced analysis:
What can be done now, but is insufficient:
The fundamental problem:
> "If you imagine a theoretical AI that is a million times smarter than the smartest security researcher—would you really be willing to trust that sandbox?"
Key difficulties:
Collison: Could the models be given competitive incentives to report on each other?
Altman: This is a promising direction. He suspects that "agent teams" not only exhibit interesting scaling laws but also show variance in alignment properties — a direction worth pursuing from both a capability and a safety perspective.
Collison noted that Altman deliberately uses "teams" rather than "swarms," pointing out that "swarm" evokes images of locusts — "Not good branding."
Altman admitted he dislikes the term "agent swarms" and prefers "agent teams." He speculated:
Collison highlighted an interesting phenomenon: Leading AI users (advanced mathematicians, elite hackers) are already making extensive use of agent teams, while ordinary enterprises have yet to adopt them on a large scale. Altman believes the core question behind this is: "How much compute can we productively throw at a single problem?"
Historical Evolution:
> "The delta between what most people are using AI for, and what the people who are using AI the most productively for has just continued to spread, and spread, and spread."
Collison notes that within OpenAI, capital allocation is effectively compute allocation. While researcher compensation is high, it has already been surpassed by compute costs.
Altman describes OpenAI's culture:
Altman reveals that OpenAI's internal AI team proved the Navier-Stokes equations (one of the Millennium Problems in fluid dynamics).
Key Data:
Learning Process:
Collison points out that high prices should attract new supply ("The solution to high prices is high prices"), but the market response has been slow.
Altman's Response:
> "I think the invisible hand is failing right now."
Reason Analysis:
A Counterintuitive View:
Altman notes that many people hope local inference will solve the supply problem, but from a global allocation perspective, putting a bunch of DRAM in everyone's personal computer (with only 5% utilization) is very bad — in a supply-constrained environment, maximum cloud utilization should be pursued.
After referring to 700 proofs, Collison asked whether there were more results; Altman said there were. They then discussed the tension between sharing scientific discoveries and the professional experience of researchers.
He articulated two clear principles:
1. Humanity should have access to foundational knowledge: Major fundamental scientific discoveries should not be concealed.
2. Scientists' careers and sense of fulfillment: A significant portion of a scientist's life satisfaction comes from making discoveries.
The Contradiction and Solution:
> "We more than owe it to the people that work in those fields to figure out what is the solution space where the world gets to benefit from the knowledge, and they also get the best version of the future that they can have."
Altman shared an unpublished anecdote: Once, when a new model was launched, the internal channel was named "Model Name-Hold My Beer," and people used it to challenge themselves with the hardest research problems of their academic careers — the model solved them one after another. He was so excited he couldn't sleep for two consecutive nights.
But he said:
> "I also had much more sadness and melancholy, and a feeling of something ending than I was expecting."
Collison attributed this to "wistfulness and nostalgia," but Altman corrected him, saying "it's more than nostalgia" — this is people's lives, something that brings them immense joy. Most mathematicians never solve a Millennium Problem, but the very idea that "you could possibly solve it" is beautiful in itself.
Collison observed that companies are very keen on adding AI chat interfaces to their own products, but consumers do not want this type of UI—they want to complete tasks directly within their primary chat application (such as ChatGPT).
The "Macintosh Moment" for personal agents has not yet arrived:
The True Source of Corporate Defensibility:
Altman strongly agreed with Collison's view—corporate defensibility does not come from HTML or mobile apps, but from brand loyalty and product choice. He used Amazon as an example:
> "If I ask ChatGPT to buy me something and I can choose Amazon versus some other thing... 10 out of 10 times I click Amazon. Pay a higher price."
Historical Analogy:
Altman compared the current situation to the early days of the iPhone App Store—large companies were reluctant to hand over customer relationships to Apple, but it ultimately proved to be just a new UI layer.
Consensus on "Fake Work":
Both agreed that any web form requiring manual human input constitutes "fake work." Altman said: "I would be so happy to never do fake work at a computer again."
Collison points out that the token, as a pricing unit, has a fundamental problem—it does not reflect commercial value.
Proposed pricing change:
Altman said OpenAI almost decided not to price its next model in tokens. Instead, a user would submit a prompt, receive a price and have the task performed, with billing resembling AWS. He called this a decision that would have been wrong: customer reactions were very negative because customers wanted to know what they were paying for.
Responses to a pricing question:
Altman said he asked a few people whether they would pay more for the same output if it were priced using a consistent number of tokens. They said they would, even if the token count were inflated, because they felt there was a real unit behind the price. Altman interpreted this as concern that a price supplied without such a unit could disadvantage customers over time.
The nanometer analogy:
Collison compared tokens with the use of “nanometers” in chip manufacturing, which he said had little relation to reality. Altman said it offended him that the term was used for something unrelated to nanometers, and that he would rather not invent a fake unit for AI.
Altman's view of tokens:
> “Tokens are a kind of fake unit. I mean, they're real in some sense, but they're not a good unit.”
Altman said comparing token prices across models and estimating chain-of-thought length created cognitive overhead that he did not think added value. He also questioned whether outcomes could be a standardized unit, because they differ too much across use cases.
Collison asked whether OpenAI optimized models to score well on benchmarks. Altman said the two things he most wanted to measure were scientific discovery beyond human capability and a rewarding experience of working with a model. He said there was no good benchmark for the latter.
Altman believed OpenAI was currently more resistant to benchmaxxing than other labs, while acknowledging that this had not always been the case. He recalled Emmett Shear's description of an ideal model as a thoughtful, challenging but respectful senior colleague.
| Organization | Guest's view or disclosed activity | Supporting context and data |
|---|---|---|
| Amazon | Altman regarded it as well positioned for agent-mediated commerce | He used his own willingness to choose Amazon and pay more as an example of trust in delivery and returns. This was a personal preference, not a customer statistic. |
| OpenAI | Altman discussed compute allocation, chips and robotics | He said heavy token consumption was not rewarded internally, described its own chip work, and discussed directing robotics toward data-center construction. |
| Stripe | Collison said tokens represented roughly 1% of operating expenses and discussed increases to 5%, 10% and 20% | The discussion of a steady-state share in the tens of percentage points applied to many companies; it was not formal guidance for Stripe. |
1. Many moderately superhuman AIs rather than one overwhelmingly powerful AI — Altman preferred a plural system resembling society's resilience. He stressed that this direction would require deliberate choices.
2. The circular dependency in sandbox safety — Altman said sandboxing and monitoring could improve protection, but questioned trusting a sandbox against a vastly smarter model. A reliably aligned model could help build a stronger sandbox, illustrating the dependency between the two problems.
3. Supply does not respond instantly to high prices — Altman expected very large cash flows at memory manufacturers, yet argued that past cycles and near-bankruptcies still discouraged expansion. This was his assessment of supply constraints, not independently verified financial guidance.
4. Scientific knowledge and scientists' experience both matter — Altman said humanity deserved fundamental discoveries, while researchers also deserved consideration as automation changed work that had given their lives meaning.
5. Agents change the interface to commerce — Collison and Altman described manual web-form filling as avoidable work. Their discussion separated the interface from the trust, products and fulfillment that could retain customer loyalty.