Sam Altman on Why is against a single

1. Key Themes

AI-Found Vulnerabilities Will Force a Re-Architecture of Software, Not Just Patching

Altman argues the cyber discovery wave is only beginning because nobody has pointed the models at the whole internet yet. The durable response is to rebuild the stack, which sounded insane two years ago and is now plausible. John Collison pushed the point that patching is the shallow fix and architecture is the durable one.

Altman: "I think we're going to find that we are in a world of many, many more serious vulnerabilities than we expect even though many people who have been following this for a while expect a lot." 00:00:31

On rebuilding: "I've seen some of these new startups saying that they're just going to like make a whole new stack. And that would have seemed insane 2 years ago... And now, maybe soon you can. I would bet that we have a completely new sort of model of how we make systems and networks." 00:02:04

On the threat model: "The thing that I've heard a lot of cyber experts talk about is they just no longer are confident that we have the right assumptions baked into our threat model." 00:07:42

Pluralism Over Singularity: "A Trillion Slightly Superhuman AIs" Beats One Giant One

This is the episode's central thesis. Altman prefers teams of agents because their properties resemble the robustness of society, with many actors checking one another, rather than a single centralized mind.

Altman: "I would much rather intuitively have a trillion AIs of like slightly superhuman intelligence join society than one AI of 1 trillion IQ join society." 00:24:13

On why: "I do think the properties of swarms match a lot of the properties of the resilience of society or the robustness of society, that I think were just missing from the like single rollout, single-brained version of the world... for all of the deep flaws of society, it is unbelievably robust." 00:23:19

He also concedes a limit: "at least for a while, maybe for a long while, we are going to have to agree to some limits about how smart we let these single models get. But a lack of progress will come from that decision, not from the AI's inability to do good research." 00:32:33

Security Model Stack: Sandboxing, Then Monitoring, Then Alignment

Altman lays out a layered view of containment. Sandboxes work only up to a point, monitoring (chain of thought plus activation monitoring) extends it, and at the limit you need real confidence in alignment.

Altman: "I think sandboxing works to a point. Then I think monitoring works to a further point where we can look at what the models are doing not only their chain of thought, but we can do activation monitoring... And then eventually I think you need to believe with huge confidence that the models are truly aligned." 00:10:07

On the superintelligence case: "if you really let yourself project forward into the like, what does it mean to have something a million times smarter than we are? I don't know. Like I think that's just gets strange." 00:10:14

Test-Time Compute Created the Agent Team Era, and a Widening Usage Gap

Agent teams only became useful after chain of thought and test-time compute. The exponential in how much compute can productively attack a single problem is under-appreciated, and the gap between average and frontier AI users keeps widening.

Altman: "a few years ago, it was like impossible to throw more than a dollar of compute at a problem and have it help you at all." 00:12:50

"The incredible exponential of how much compute you can throw at a single problem and get better results is one thing. 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." 00:13:41

Let People Explore at the Compute Frontier, Then Optimize

OpenAI intentionally let a team's compute use run unconstrained for a week on the Navier-Stokes proof to learn the scaling laws, trusting optimization pressure to compress costs later.

Altman: "let people explore at the limit of what is possible, and then trust that the incredible optimization pressure will figure out how to do this in an efficient way." 00:17:05

On the cost curve: "I bet we could redo Navier-Stokes at 1/100th of the cost a year from when it was done." 00:17:45 He put the cost at "Order 10 million" per proof, with "hundreds of millions of dollars of math" in total. 00:18:04

Iterative Deployment and the "Terrible Sales Pitch" Problem

Altman thinks the AI field has failed to explain why the upside justifies the downside, and that the surveillance-state-plus-basic-income framing is a poor pitch. OpenAI's defense is iterative deployment and fighting capability overhang, even when unpopular with the safety community.

Altman: "I don't think we have done a great job of explaining to people why we think all of the potential downside of this technology is worth it." 00:05:04 "This is like a terrible sales pitch." 00:06:24

"even when it has been hugely unpopular in the AI safety community for us to push for iterative deployment and putting these tools into the hands of people, we have really stuck to that principle." 00:06:26

"building up a capability overhang is a thing to fight against." 00:07:13

The Compute Shortage: Market Signals Are Failing and Supply Must Rise

High prices have not yet produced supply response. Altman explains why (memory makers' scar tissue) and wants costs per gigawatt to fall dramatically.

Altman: "if you were a producer of DRAM, you would invest $20 billion of CapEx to make $19 billion... they all like remember their history and how many times they almost went bankrupt by making too many fabs and being promised this time it was different." 00:27:48

On the invisible hand: "I think the invisible hand is failing right now." 00:27:26

On cost targets: "let's pick $70 billion as a number... I think if that number were like 20 instead of 70, we should be much happier." 00:38:51

Self-Replicating Data Centers: The Underdiscussed Cousin of RSI

Altman flags an under-discussed path to abundance: AI controlling robots that build fabs, turbines, solar panels, and data centers.

Altman: "People talk a lot about RSI as like a AI that does AI research. There's very little relatively little talk about like self-replicating data centers... you can use AI to control robots, and run fabs, and build turbines or solar panels, and, you know, build data centers. But I think we have that like within grasp." 00:26:20

Personal Agents Haven't Had Their "Macintosh Moment" Yet

Altman believes the winning personal agent needs only intelligence, context, and tools, and is "a small number of ideas away" from mainstream. Reliability is the unlock.

Altman: "I think this is like a small number of ideas away from a super mainstream thing." 00:41:06 "it'll just work. It'll be reliable." 00:41:34

Agentic Commerce: Brand and Fulfillment Survive, Web Forms Die

Collison argues defensibility lies in brand and selection, not HTML. Altman agrees, using his own behavior as evidence.

Altman: "if I ask ChatGPT to buy me something and I can choose Amazon versus some other thing, like 10 out of 10 times, I click Amazon. I pay a higher price." 00:44:32

Collison: "anytime there is a web form field that you as a human are filling out, these days, that is below your pay grade." 00:45:31

2. Contrarian Perspectives

Don't Let Everyone Run Models Locally: It's Globally Wasteful

The intuition that open-source local inference relieves the data-center squeeze is backward in a supply-constrained world.

Altman: "from like a global allocation perspective, actually, it's quite terrible to have everybody like have a bunch of the world's limited supply of DRAM sitting in their own computer and being used 5% of the time... you want maximum utilization in the cloud if you're trying to like deal with this very supply-constrained environment. And somehow everyone's Many people's intuition is the opposite of that." 00:28:59

Pay Less Attention to AI Spend Than to Meetings

Counter to the enterprise worry about token budgets, Altman says most companies are under-using AI, and wasteful meetings are the bigger bleed.

Altman: "use of AI is like not a company's biggest spend problem relative to other things that happen... near-more companies have the problem of not using enough AI rather than using too much AI." 00:35:32 He also noted OpenAI discourages gaming: "Like no token leaderboards. It's actually slightly bad to be high on the token leaderboard." 00:15:31

Don't Slow-Roll Fundamental Discoveries to Protect Scientists, But Do Feel the Loss

Altman holds that the world deserves scientific knowledge even at the cost of a human tradition. He also admits unexpected grief at watching models solve researchers' hardest problems.

Altman: "humanity deserves fundamental knowledge. And we should not be sitting on a major fundamental science discovery." 00:18:39 On the experience: "I was tremendously excited and happy, but I also had like much more sadness, and melancholy, and like a feeling of like something ending than I was expecting." 00:20:55

Businesses Clinging to Their UI as a Moat Are Making the App Store Mistake Again

Collison argues that businesses resisting agentic access are protecting a transitory layer. Altman draws the parallel to the App Store era, when big companies reluctantly ceded their consumer relationship to Apple.

Altman: "all of the big companies felt like they had to make one, but they kind of didn't want to, because they're like, 'Oh, I'm going to like hand my, you know, relationship with my consumer over to Apple.'" 00:46:59

Customers Prefer Inflated, Legible Tokens Over Opaque Outcome Pricing

OpenAI nearly priced a model per-prompt, AWS-style, but customers recoiled.

Altman: "We almost made a very What would have been bad call uh to say we're not going to price our next model in tokens at all... the customer response on that is so negative." 00:49:51 "'Oh, I'd much rather inflate you inflate the tokens and pay more, because I know there's like a real thing there.'" 00:50:16

3. Companies Identified

OpenAI. Frontier AI lab behind ChatGPT, Codex, and GPT-series models. Mentioned as Altman's own company; relevant for its stated strategy: iterative deployment, custom chip, self-built data centers, robotics program, an internal culture of AI adoption, and cost-per-intelligence reduction. Altman: "Our chip... is an incredible performance at like, you know, units of intelligence per watt. We are learning how to build our own data centers. We have a robotics program that we're going to point at making lots of data centers." 00:25:21 On culture: "we have a culture where people, even if they've done something the old-fashioned way for a long time, are super willing to like try using AI to do it better." 00:31:13

Stripe. Payments and business-infrastructure company (Collison's company). Mentioned for its AI adoption and Stripe Atlas. Collison: "We now spend, you know, 1%-ish, order of magnitude, of Stripe's OpEx is tokens." 00:36:34 Altman: "That feels too low." 00:36:43

Stripe Atlas. Stripe's company-formation product handling incorporation, tax IDs, and founder equity. Ad read: "Thanks to coding models, your weekend side hustle can quickly become a serious undertaking... Stripe Atlas helps you with everything you need to start a business." 00:24:52

Amazon. E-commerce and fulfillment giant. Highlighted as a likely winner in agentic commerce because of trust and fulfillment. Altman: "I think they're in a great place if commerce moves like much more agentic." 00:45:06 Collison: "Amazon, in particular, has this amazing defensibility via the fulfillment network." 00:45:23

Dots (personal agent product) and similar agents ("Use," "Grokbot"). Personal agent products launched recently; Altman sees the category as promising but not yet "just works" reliable. Altman: "I think it'll look a lot like the current, you know, Dots, or Use, or Grokbot, or whatever, but it'll just work." 00:41:26

DRAM / memory makers. The memory industry is a historically poor business now seeing windfalls. Altman: "all of a sudden, you know, these guys are each going to make like $100 billion of like free cash flow in a year. Unbelievable." 00:27:48

Intel. Chipmaker; Collison suggested it could re-enter memory given relevant experience. Collison: "Intel makes chips. They can probably get back into the memory game, as is tradition." 00:28:30

ASML (implied). Altman referenced the lithography chokepoint without naming it: "in terms of like how to do the lithography, yeah, I think there's like one company on Earth that seems to know how to do that." 00:30:52

Infosys. IT consulting firm; Collison speculated that the global software upgrade may be a boon for such consultancies. Collison: "Is this huge for Infosys and all those guys because there's going to be a lot more software consulting projects?" 00:01:38

Stripe-style new-stack startups (unnamed). A generation of startups rebuilding routers, operating systems, and applications for an AI-native threat model. Altman: "you're already seeing like a new generation of startups talking about building this like new kind of internet or new kind of application architecture." 00:01:26

Industrial-robot and actuator makers. Mentioned as an area where high prices are pulling in innovation. Altman: "Components for robots. And not like the humanoid like, you know, kind of like fake demos, but like the industrial robots that can just make stuff faster." 00:30:05

Turbines, batteries, concrete suppliers. Supply chains currently seeing price pressure from compute buildout. Altman: "Turbines, batteries, concrete." 00:29:51

PHP. Open-source language at the center of the AI-found-vulnerabilities episode. Collison: "I want to ask about PHP week." 00:00:13

YC (Y Combinator). Referenced in the App Store analogy. Altman: "most of the YC batch was like, 'I'm going to build an iPhone app.'" 00:46:59

Apple. Mentioned for the iPhone App Store distribution shift. Altman: "The iPhone App Store specifically." 00:46:59

ChatGPT / Gemini. Frontier consumer assistants framed as the new "top of wallet" interface. Collison: "The new UI is that agentic interaction inside ChatGPT or Gemini or or whatever it is." 00:42:43

Xerox. Referenced for the Xerox Alto computer in Altman's reflections on lost arts of computing. Altman: "they would like pull out their gigantic Xerox Alto disk." 00:04:04

AWS. Used as the model for opaque, outcome-based pricing that OpenAI considered and rejected. Altman: "it's going to be like AWS, you just do something, and we send you a bill." 00:50:02

Google. Referenced for the practice of showing meeting costs in Google Calendar. Altman: "There was a time when a lot of companies had this like thing in their Google Calendar, which showed the like approximate cost of each meeting." 00:35:24

4. People Identified

Sam Altman. CEO of OpenAI and the guest. Offers the episode's core ideas on pluralistic AI, compute, and safety. Altman: "I would much rather intuitively have a trillion AIs of like slightly superhuman intelligence join society than one AI of 1 trillion IQ join society." 00:24:13

John Collison. Co-founder of Stripe and host. Contributes sharp reframings, such as that web forms are "below your pay grade" and that token spend will reach tens of percent of OpEx. Collison: "isn't it going to be, depending on the kind of company, isn't it going to be tens of percentage points of their OpEx?" 00:36:57

Greg Brockman. OpenAI co-founder and president. Altman called him "one of a kind, in the best sense." 00:49:10 "a combination of like yeah, intensity, like monster productivity, extreme dedication, and like focus on succeeding at the mission." 00:49:17

Emmett Shear. Entrepreneur and former interim OpenAI CEO. Altman credits him with the best description of the ideal AI collaborator. Altman: "Emmett Shear said to me years ago that what he wanted was a really thoughtful and challenging, but still respectful, senior coworker. And I have still not heard a better like brief encapsulation than that." 00:48:52

Paul Graham. Y Combinator co-founder. Cited for the observation that companies obsess over trivial savings while tolerating large wasteful costs like meetings. Altman: "there was this observation that like Paul Graham made it at one point... that there are people at companies that have incredible power to save like, you know, $17 on the snack budget of the office every week." 00:35:32

Claude Shannon. Information theory pioneer. Collison invoked him in the token-efficiency discussion. Collison: "Claude Shannon would not approve of us just streaming noise rather than signal." 00:51:27

Collison's father (unnamed). Observed horses still used in Irish industry in the 1950s, used to illustrate the wistfulness of witnessing a technological chapter close. Collison: "growing up in Ireland in the 1950s, he observed people around him still using horses productively in industry." 00:22:04

5. Operating Insights

Never Reward Raw Compute Usage; Reward Demonstrated Importance

OpenAI's internal compute policy gives wide latitude but a skeptical default, and refuses vanity metrics. This keeps experimentation open without incentivizing waste.

Altman: "We try never to reward people for using a lot of compute. Like no token leaderboards. It's actually slightly bad to be high on the token leaderboard... it's like a skeptical default, and if you can show that you're doing something important, you will have nothing but appreciation." 00:15:29

In a New Paradigm, Deliberately Run Unconstrained Experiments to Learn the Scaling Laws

When no one understands the cost-benefit curve yet, constraining early experiments forecloses learning. The Navier-Stokes effort ran without constraints for a week to learn how to allocate model size, copies, and rollout length.

Altman: "when there's a new paradigm, I think it's important to let people be fairly unconstrained... we understood absolutely nothing about the scaling laws of teams of agents." 00:16:19 He expects "10x improvements per year at how we allocate dollars." 00:16:36

Tell Agents a Budget in Natural Language

Collison assumed agents couldn't be bounded by spend. Altman said they can.

Altman: "If you say, 'Do the best job you can with $10,000 of compute,' like, yes, it will understand that." 00:14:42

Expect AI Spend to Hit Tens of Percent of OpEx, and Bias Toward Under-Spending Risk

Both agree the steady-state share of OpEx going to AI will be large. Collison is already at roughly 1% and says "it'll get to 5% and 10% and 20%." 00:36:45 Altman's guidance is to worry about under-adoption first.

Let Adoption Culture, Not Mandates, Drive AI Use Into Non-Obvious Functions

Altman credits OpenAI's willingness to try AI on established workflows, including procurement, as a cultural advantage. Altman: "AI research, software engineering, but also increasingly like procurement, or things that people think of like very outside the realm of what most people use AI to do." 00:31:37

6. Overlooked Insights

The Real Cost Lever Is Gross-Margin Stacking Between Sand and Tokens

Altman dropped a throwaway line that reframes where compute costs come from: the compounding of margins across every layer of the supply chain, not just energy or chips. To get from roughly $70 billion to $20 billion per gigawatt over five years, "all of the gross margin stacking that happens between melting sand, or the equivalent of melting sand, and getting the token back, a lot of that's got to get compressed." 00:39:54 He added that even with free materials and zero profit, hitting $20 billion would be possible "in the current energy envelope, but only just." 00:40:12 That implies investable leverage in vertical integration of the stack (chips, data centers, robotics-built infrastructure), which is exactly what OpenAI says it is doing.

Agent Teams May Be an Alignment Strategy, Not Just a Capability Strategy

Mentioned almost offhandedly in the cyber discussion and then returned to as the "most excited" safety item: Altman suspects that "teams of agents have very interesting scaling law properties, but also very interesting alignment properties." 00:11:38 Combined with his stated willingness to cap the intelligence of any single model, this signals a possible strategic direction: distribute frontier capability across many mid-sized, mutually checking agents rather than one monolith. If that becomes the dominant paradigm, value accrues to orchestration, monitoring, inter-agent incentive design, and the infrastructure to run thousands of parallel rollouts, rather than purely to whoever trains the single largest model.