29 Oct 2026 · 53m · Sam Altman, John Collison · 20 quoted facts
Sam Altman argues that the safest, most useful AI future is many moderately superhuman AI agents working in teams, not one enormous superintelligence. Teams of agents check each other and echo the checks and balances that make human society resilient. Altman says this matches how AI is already being used: huge amounts of compute thrown at hard problems like cyber defense and math, with the benefits shared widely through iterative deployment.
The central tension is between speed and control. Models are finding serious software vulnerabilities and solving research-level math, and Altman is both thrilled and wistful about it. He admits that sandboxing, meaning containing a far smarter model, is unsolved. Meanwhile compute stays scarce and expensive, so the episode turns to supply chains, token pricing, and how businesses must adapt to personal agents.
Altman says people had not yet pointed models at the huge amount of code on the internet, and that models keep improving fast. He expects far more serious vulnerabilities than even informed people anticipate, which is why he and Collison favor re-architecting systems rather than only patching them.
Altman says sandboxing works up to a point, and monitoring (including reading the model's internal activations) works a bit further. Beyond that you must trust the model's alignment completely, which is hard to do for something far smarter than us.
Altman says teams of agents scale well and also have useful alignment properties, since agents can check each other the way people do in society. He would rather have a trillion slightly superhuman AIs than one AI with a trillion IQ.
OpenAI let researchers spend freely for about a week on math, and the Navier-Stokes work cost roughly $10-20 million. Altman bets it could be redone for about one-hundredth of the cost within a year, because new paradigms typically deliver around 10x efficiency gains per year.
Altman says memory makers remember decades of boom-and-bust and are cautious about investing in new capacity, so high prices have not yet produced a supply flood. He wants the five-year cost of a gigawatt of AI capacity to fall from about $70 billion toward $20 billion, which needs far more supply and cheaper energy.
Altman says people had not really aimed these models at the enormous amount of code online, and the models keep getting better quickly. He expects many more serious vulnerabilities to turn up than most people predict.
Many AI agents work on the same problem in parallel and check each other, rather than one model working alone. Altman prefers the word 'teams' and thinks they scale well and make AI safer because the agents keep each other in check.
A sandbox is a locked-down environment meant to contain a model. Altman says that for an AI a million times smarter than the best security researcher, you would have to completely trust its alignment (its goals matching ours), and nobody can yet do that.
Altman says roughly $10-20 million for that one proof. He says the company spent hundreds of millions of dollars on math work overall, and he thinks the Navier-Stokes proof could be redone for about one-hundredth the cost within a year.
Collison argues that customers will want to ask their own AI assistant to do things, such as refund an order, rather than use each company's website or app. Altman broadly agrees, saying that brand trust and fulfillment, as with Amazon, matter more than a company's interface.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Sam Altman said that when Codex went down during an internal incident at OpenAI, staff briefly seemed to have forgotten how to code without it.
“There was a crazy moment at OpenAI where we had a sev and codex went down. And everyone had forgotten how to code.”
Sam Altman said OpenAI has kept to iterative deployment of its AI tools even when the policy was unpopular in the AI safety community.
“even when it has been hugely unpopular in the AI safety community for us to push for iterative deployment and putting these tools in the hands of people, we have really stuck to that principle.”
Sam Altman said that a few years ago it was not possible to usefully apply more than about a dollar of compute to a single problem.
“a few years ago, it was impossible to throw more than a dollar of compute at a problem and have it help you at all.”
Sam Altman said the field had not yet worked out how to use large amounts of test-time compute for better results until chain of thought reasoning came along.
“we had not figured out how to use a lot of test time compute to get you a better result”
Sam Altman said OpenAI avoids rewarding employees for high compute use and that ranking high on a token leaderboard can count slightly against them in promotions.
“It's actually slightly bad to be high on the token leaderboard, certainly on a promotion.”
Sam Altman said that when OpenAI got a new model, staff made an internal channel where they threw the hardest research problems at it, and the model solved many of them.
“The name of the channel was model name dash hold my beer. People were just throwing that at the hardest problems they could.”
Sam Altman estimated that OpenAI's Navier-Stokes proof cost on the order of $10 million to $20 million in compute.
“Order 10 million for Navier-Stokes, I think. Okay. Maybe a little more. 10 million-ish. 10, 20.”
Sam Altman predicted OpenAI could repeat its Navier-Stokes proof at about one-hundredth of the original cost a year later, thanks to better models and approaches.
“I bet we could redo Navier-Stokes at 1 100th of the cost a year from when it was done.”
Sam Altman recalled that many companies once added a Google Calendar feature showing the approximate cost of each meeting based on attendees' pay.
“There was a time when a lot of companies had this thing in their Google Calendar which showed the approximate cost of each meeting.”
Sam Altman said that for a long time DRAM makers would invest about $20 billion in capital spending to earn about $19 billion, making memory a poor business over the long term.
“for a very long time, if you were a producer of DRAM you would invest $20 billion of CapEx to make $19 billion”
Sam Altman said memory manufacturers are now on track to each earn about $100 billion in free cash flow in a year.
“and then all of a sudden you know these guys are each going to make like $100 billion of like free cash flow in a year”
Sam Altman argued that local AI inference wastes scarce DRAM because memory in personal computers sits idle most of the time, whereas cloud data centers can run at much higher utilization.
“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”
Sam Altman said that only one company in the world seems to know how to do leading-edge chip lithography.
“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.”
Sam Altman said chip design, meaning producing the file sent for tape-out, is well understood and a good problem for AI.
“In terms of how to do the chip design, like to build the file that you eventually send it tape out, that is well understood.”
Sam Altman used roughly $70 billion as the current five-year total cost of ownership for a gigawatt of AI compute, and said he would be much happier if it were closer to $20 billion.
“People give different numbers for the kind of the cost per gigawatt measured in like the five-year total cost of ownership right now, but like let's say, let's pick $70 billion as a number.”
John Collison said OpenAI launched a personal agent called Dots a few weeks before the conversation, as other companies also launched personal agents.
“you guys launched Dots a few weeks back at this stage, and there's a bunch of other people also launching in the field.”
Sam Altman said that when shopping through ChatGPT he would choose Amazon over an unknown vendor every time, even at a higher price.
“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”
Sam Altman recalled that when the iPhone App Store launched, most Y Combinator startups rushed to build for it, while big companies felt they had to follow despite handing the customer relationship to Apple.
“most of the YC batch was like, I'm going to build an iPhone now. This is like new distribution.”
Sam Altman said Emmett Shear once described the ideal AI model as a thoughtful, challenging but still respectful senior coworker.
“Emmett Shear said to me years ago that what he wanted was a really thoughtful and challenging but still respectful senior coworker.”
Sam Altman said OpenAI nearly decided not to price its next model in tokens, but customers reacted negatively because they wanted to know what they were paying for.
“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.”
Altman explains why AI-found bugs matter and why rebuilding software architectures may be the durable fix, not just patching.
“we are in a world of many more serious vulnerabilities than we expect”
Altman defends releasing AI tools gradually to the public despite criticism from the AI safety community, arguing the benefit must be explained to ordinary people.
“push for iterative deployment and putting these tools in the hands of people”
Altman says containing a far smarter model is an unsolved problem and that trust in alignment is the real crux.
“sandboxing works to a point. Then I think monitoring works to a further point”
Altman describes how reasoning models made it possible to spend ever more compute on one problem, widening the gap between average and advanced AI users.
“how much compute can we productively throw at a single problem”
OpenAI let compute spending run loose for a week on math, which produced big results and some sadness among researchers about a chapter ending.
“we understood absolutely nothing about the scaling laws of teams of agents”
Altman argues for pluralism: many slightly superhuman agents checking each other would mirror the resilience of human society.
“1 trillion AIs of, like, slightly superhuman intelligence join society than one AI”
Altman explains why high prices have not yet brought new supply and why OpenAI is building its own chip and robotics program.
“relatively little talk about self-replicating data centers”
Altman says the cost of AI infrastructure should fall as supply expands, and puts today's five-year cost per gigawatt near $70 billion.
“let's say, let's pick $70 billion as a number”
Collison and Altman discuss how consumers will increasingly ask agents to act for them, putting pressure on companies' websites and apps as sources of defensibility.
“these days that is below your pay grade”
Altman says customers wanted tokens as a billing unit even though tokens are an imperfect measure of value.
“I'd much rather you inflate the tokens and pay more”
We're going off script. We're just having a beer. It's not the first non-Guinness, but it is the first IPA. And so, it's the first time for everything. Cheers! Cheers! I want to ask about PHP Week. Yeah. First off, how did it... Why did it take so long to find these bugs? Because it sounds like they were kind of... Why did it take, like, the 20 years, or why did it take the last few months? No, why did it take a year?
Maybe start by explaining the bug. Or the problem.
So the problem is that although we have known for a while that these models are getting good at finding vulnerabilities in the wild, A, they continue to get better fairly rapidly, and B, people just hadn't really used them to go look at everything out there. There's a lot of service area and a lot of code on the internet. And as people have started to do that, I think we're going to find that we are in a world of many more serious vulnerabilities than we expect, even though many people who have been following this for a while expect a lot.
Yeah, I just feel like a lot of the discussion ends up being around we need to fix the software. And sure, yeah, we should fix the software, but also we maybe need to change the architectures, which is harder, but a much more durable fix.
Like that seems to me like what needs to happen. It would seem really hard, except we have this amazing new tool to do it with. So I kind of think that's what's going to happen. And you've already seen a new generation of startups talking about building this new kind of internet or new kind of application architecture for this world where you assume that the models get smarter and smarter.
What do you think the global software upgrade looks like? For example, is a PHP code base just a bad idea and you should rewrite a nice clean version in Rust or do we upgrade PHP? But yeah, everyone thinks that we need to upgrade all hand-spun software. Just what form will that actually take? Is this huge for Infosys and all those guys? Because it's going to be a lot more software consulting project.
I mean, people are talking more and more about just like, let's re-architect what it means to be a router or an operating system or an application. And I've seen some of these new startups saying that they're just going to make a whole new stack. And that would have seemed insane two years ago. You would have been like, there's no way. there's too much you can't write all that 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 what else reminds you of the early internet days the thing that most reminded me of the early internet was uh you know there is there is this
like spirit of just like unbelievable creative unleash in people using these tools um you see it in kids particularly and like and again maybe this is just like because i remember what it was like when I first got a computer with internet and I could like do all these things that adults just couldn't even think to ask for and like watching a you know eight or ten year old make a complete video game with some crazy ideas they have that like no adult would really have and expect that to just be the way that technology is going to work for the rest of their lives in fact getting better and better and just like this constant I'm going to make
useful stuff for me and show it to my friends yes like we had to suffer when I was a kid to make a piece of software yes They barely did anything. And now you can have like a full 3D video game, like unbelievable anything you could think of in 10 minutes. That reminds me of the early internet, including that just like these people, they're never going to be mind blown by that because that's always what they expect out of the world.
Like any ideas they can come up with, anything they want to create, anything that they think might be useful, anything that they just need to make their lives better, like the computer's supposed to make it. There was a crazy moment at OpenAI where we had a sev and codex went down. And everyone had forgotten how to code. I mean, not all the way, but it was like a, whoa, what if you couldn't do it? I remember when I was a student and I would meet the really old guys and they would once in a while pull out their gigantic Xerox Alto disc or they would pull out their line by line written of assembly or they would pull out their
punch cards. And I felt like it is incredible to me that you all bootstrapped us to this point, doing it this way.
Yes, no, I agree. And keeping programs in your head, right? You know, programs running the first time after, you know, hundreds, thousands of lines were written. No, there's a few lost arts like that. Going back to the cyber side of things, in retrospect, you know, was the cyber RL a mistake? Or did we just have to go there? It's a little like the SNL thing of, you know, we urge you to urge us to stop. It's like the models are getting very good at cyber capabilities.
Like, yeah, I mean, they are getting very good, but not just totally emergently.
Look, the reason that SNL thing was so funny is because it was so true. Like, humor like that doesn't stick unless there's an element. It's like clearly pointing out a ridiculous but true thing about society or someone. and and the cyber thing there is a point which is like okay you know sure the models are really good at cyber but like who asked for this why are you doing this like one of my big disappointments criticisms are by our field or whatever and i think maybe the cyber models getting good at cyber is an example of this is 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 like the the kind of caricature of this is people feel like we're like all right you know fine you say the model's going to cure cancer someday maybe it will and if so i guess that's good and you say it's going to like make stuff cheaper and okay maybe i already see that's good and i get access to medical advice or legal advice but but then you paint this picture of the future where it's like a new industrial evolution and maybe the last one wasn't that good and this one sounds worse and i will have no job, no ability to like have control or choice or freedom of my future and AI is so dangerous
that you're gonna have to like have a crazy surveillance state so we don't all die. This is like a terrible sales pitch. And one of the things that I am most proud of about OpenAI is even when it has been hugely unpopular in the AI safety community for us to push for iterative deployment and putting these tools in the hands of people, we have really stuck to that principle. And I think this is the only kind of good path.
I think people, it's not enough to tell people, like, you're going to get the model to do these great things for you. It's got to be like, we are going to enable you to have a better life that you define for yourself to do useful things. You're not going to just sit and, you know, get a basic income and have a bunch of people sitting in a few square miles of San Francisco
make all the decisions about the future. Iterative deployment is this idea that you just have to get the advances out there and you can't kind of stopper them up inside. I mean, not all like,
you still have to make them like safe before you put them out there. You can't like cause havoc in the world. But that building up a capability overhang is a thing to fight against. That directionally speaking, yes, you want to get this into the world. You want to get this into people's hands and you want to let people use the technology to figure out all these wonderful things.
The exploits get a lot of headlines. The prevented exploits do not. How quickly are people making progress with CyberBlue?
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. Relative to my worst fears, if I can think back 10 days ago, the first day of that went much better than I expected. And we heard a bunch of stories from people that we're using our models to, I mean, not only build patches and build cyber defense, but like individual companies who are just like, I need to do a quick scan deployment.
Like, I'm going to take my thing down for an hour. I'm going to use this thing and I want to put it back up because my application's offline. Now I feel like I really jinxed it and I shouldn't have said that. But the first day did go a lot better than I thought. And I think people have been able to use that. 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 and they're like look if if everything is what we think and hardware exploits are as hard as we think they are then yes we can see how this is
going to work but if ai is going to discover really deep novel bugs in hardware that require physics we didn't understand but all of a sudden means some assumption doesn't hold then i think people aren't sure what happens one thing that
always strikes me is that when I talk to very knowledgeable people inside OpenAI and similar companies, there is a pessimism that sandboxing is possible, basically. And that would not have been my mental model going in. It would have been this, yeah, you have to build a really tight sandbox, and we haven't kind of hardened them historically enough compared to the current capabilities, but ultimately you can have a maximum security prison for the model to work in and keep it inside there.
It seems like you guys have kind of a pessimism that sandboxing models is possible.
My instinct, well, first of all, I think we knew much better than we're doing now, and we should. I think we clearly, that is the main thing to focus on right now is a better security model. If you imagine a theoretical AI that is like a million times smarter than the smartest security researcher, would you like really be willing to trust that sandbox, especially if the model can like discover new science that people don't already understand?
You can use the models to construct the sandboxes. I think you can. Well, so this gets, let me come back to that. 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 in their chain of thought, but we can do activation monitoring and try to really understand like are the models are we happy that the models are not deceiving us and then eventually i think you need to believe with huge confidence that the models are truly aligned and are going to value and try faithfully to do to follow their actual the intent of the people operating them and that they're going to
value sort of people being in control of the future forever and if you had a model that you could fully trust the alignment that you were absolutely certain of the alignment of then yes i think you can have that to help you build better and better sandboxes but like 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 i i don't know like i think that just gets strange.
Because the alignment becomes hard.
Because you need to really, really trust the alignment to trust that these sandboxes staying ahead of
capabilities. Can you not give models competing incentives and give them incentives to defect from each other? I think this is a promising direction.
I suspect that this idea, I hate the word agent swarms. Let's call them agent teams. But it seems to be Taking over. I suspect that teams of agents have very interesting scaling law properties, but also very interesting alignment properties. And I think that's like a direction we should push in more, not just from the capabilities, but from a safety perspective, too.
I know it when you talk about agent teams, not swarms. A swarm is such a scary... I know, it's a locus, yeah. Not good. Not good branding. in keeping with many historical AI industry traditions. We are so good at this. We'll get back to branding. But I note that there seems to be a big gap in what the leading AI users are doing and kind of standard corporate use in the use of swarms where like all the advanced math, sorry, teams, all the advanced math uses agent teams, all the advanced hacking uses agent teams, and then like common or garden corporate stuff is not using big swarms.
Why has that not done? Because we've had this for many months. Not to the same degree.
One way to look at this problem or this question is how much compute can we productively throw at a single problem? And a few years ago, we forget how fast this is all gone, but a few years ago, it was impossible to throw more than a dollar of compute at a problem and have it help you at all. Why? you could have GPT-4 like grind out for you, you know, like 100 ,000 tokens or something.
And it was going to answer your question. And it certainly was not going to be able to go right productively, like write you a piece of software. We just, we had not figured out how to use a lot of test time compute to get you a better result. And then we figured out chain of thought and the models could think hard about a problem. And now, and then like people really started using a lot of compute.
and now you can throw a whole team at something
and you can do that 10 ,000 times in parallel. You need chain of thought and test time compute before you can get to swarms because you're kind of a useful swarm with a bunch of G3s.
But so the effect of this, I think this is not being well-followed, paid attention to in the world, but 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 are using AI the most productively for has just continued to spread and spread and spread. How do you allocate
compute budgets now? Because presumably, even though researchers are paid a lot, it's eclipsed by compute. And so capital allocation at OpenAI is sort of compute allocation. And it's hard to a priori know how much it will cost to go do something. Yeah. It's actually also hard to just control the, like you can't quite prompt an agent where it's like, go do a really good job of this and don't spend more than $10 ,000. You can. Oh, OK.
Will it do a good job of it?
Yeah, if you say do the best job you can with $10 ,000 of compute, yes, it will understand that. Huh. I mean, there's another kind of hard question for society about what does it mean when you have two companies competing and one says you can spend $10 ,000, and one says you can spend $100 ,000 on the same problem. It's clear what's going to happen there, but there's a question about how society thinks about compute allocation and wealth and things like that.
But we have our own version of this inside of OpenAI. We allocate a lot of compute to internal productivity, and then people get very wide latitude to make their own decisions unless they're really being abusive.
But do you just encourage people to do useful stuff, and if they spend $10 million in compute, it's like, oh, interesting.
Correct. We try never to reward people for using a lot of compute. Sure. Like no token leaderboards. It's actually slightly bad to be high on the token leaderboard, certainly on a promotion. But it's like, it's a skeptical default and if you can show that you're doing something important, you will have nothing but appreciation.
But at a certain point, this will need to change because you're describing the increasing orders of magnitude and right now you have a researcher who does some very interesting math for $10 million. And you're like, a little pricey, but cool math. And then $100 million. And they're doing the billion-dollar math. You're like, look, I'm really happy that you're making so much math progress, but I really need you to cool the jets with the math. We did let it get out of control for like one week on the math.
Intentionally so, because you have to be like... And it worked. Well, you don't know how... You don't know how far it's going to go until you try. Yeah, yeah. And I think it's important to have... When there's a new paradigm, I think it's important to let people be fairly unconstrained. So when we did the Navier-Stokes proof, we understood absolutely nothing about the scaling laws of teams of agents. Are you supposed to like - The very recent Navier-Stokes proof.
Yes. Are you supposed to give them the smartest model? Are you supposed to give them many copies, fewer copies, longer rollouts, shorter rollouts? We did not understand that at all. And now we understand it a little bit. I think we can still understand it much better. But like has happened in every other new paradigm we've discovered, I am confident that we will find our 10x improvements per year at how we allocate dollars or computer flops, whatever you want to measure, into useful things.
So I think this is a thing that our company is particularly good at, but also like fits an important property of the field, which is 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. I, and, you know, directionally speaking, I would rather us be more willing to let people use crazy amounts of compute and let users have extremely high-end plans to figure out the edge of the possible and then figure out how to make that efficient at a mass scale.
But when does the optimization function come in?
Well, I think we're already doing it. I bet we could redo Navier-Stokes at 1 100th of the cost a year from when it was done.
So we did it in like... Not just knowing how the proof works, but just with better approaches. Yes.
Smarter models and better approaches. How much did all the math cost? Order 10 million for Navier-Stokes, I think. Okay. Maybe a little more. 10 million-ish. 10, 20.
That order makes you. Yeah, yeah. But for one proof. And so times 700?
Well, no, no, because then we kept getting better so fast. Like, we understood more.
Got it. We spent way less than that on the average one of those.
But like hundreds of millions of dollars of math.
Yeah, yeah, yeah, that kind of work. Interesting.
But that includes like much more other discovery than what we have published. Yes. So it was like, maybe we spent that much.
More than the 700 proofs?
Of course. When do we get those? We are trying to learn how to... This is like a real tension. I think on one hand, very clearly, and I feel unequivocal on this point, humanity deserves fundamental knowledge. And we should not be sitting on a major fundamental science discovery. Like, the world deserves that and gets to know. And I think that outweighs another very important equity, which is if you are a scientist, your career and, you know, a significant amount of your enjoyment and fulfillment in life comes from discovery and science.
And so if we're going to continue to do this for math or begin to do this for other fields, then I think 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 benefit from the knowledge and they also get the best version of the future that they can have. So we are working through that right now.
I did not. I think you may be too contrite. I did not interpret that moment as particularly bad for AI where you had, I think everyone saw the mathematicians union up in arms. And I think everyone else says, yeah, I mean, have you paid attention to 2026?
As I said, I don't think it was bad for AI. And I think on net, the world deserves to understand how the universe works. But I do have empathy for the math. I think we need to do this. I think we We can go understand how to, you know, a theory of all of physics or how to cure a bunch of diseases. Like, clearly, that is knowledge the world should have. But I would like to figure out a way to do this that is as sensitive to the communities that have driven these fields forward for centuries as we can.
And it's just a hard thing.
I wasn't that worried about the math. I thought it was great. But you agree the mathematicians are not that happy. Yeah. It was fine. The longshoremen union weren't happy about the Porsche automation.
Why do I feel differently about this? There was one weekend when we got a new model. I'm probably not supposed to say this, but I will since we're drinking beers. Exactly. We're half-pined in. The name of the channel was model name dash hold my beer. People were just throwing that at the hardest problems they could. and people just like didn't think the model was going to do these things and and we had this collective experience of people's like hardest research problems from their academic career the model was just knocking down one after the other many different fields it was incredible to watch and kind of two nights in a row i was up most of the night because i
was so excited i couldn't go to bed it was it was like many of those problems that we published many other things and 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 so again I agree this is the right thing on behalf of the world and and also I agree that this is just like the shape of reality and nothing that anyone can stop or should stop.
But I do have, like, real...
I think it's a wistfulness and nostalgia you can feel. It's more than nostalgia. Wistfulness. Yeah. Like, my dad describes how, you know, growing up in Ireland in the 1950s, he observed people around him still using horses productively in industry. In the 50s? Yeah. And, I mean, obviously there were also tractors, but, you know, it was being phased out. but it still for sure happened and observing that this was a technology that humans had been using for thousands of years and he was witnessing the very tail end of it and yeah I think there's always that feeling of witnessing the closing of a chapter that leads to some wistfulness but it's okay
it's people's lives and it's something that has brought them great joy and you know maybe most Most, I mean, certainly most mathematicians never solve a millennium prize problem. But the idea that you might be able to, and the idea that there is stuff out there that we don't understand yet, that through great effort and, you know, pushing aside other things in your life and dedicating yourself to, you can still solve and understand. Like, there's something beautiful about that.
What else are you excited about from a safety point of view? We talked about this a little bit earlier, but I want to like emphasize it again. I'll just call it Swarm since that seems like the term right now. I do think the properties of Swarm match a lot of the properties of the resilience of society or the robustness of society. That I kind of, I think we're just missing from the single rollout, single-brained version of the world.
Like, for all of the deep flaws of society, it is unbelievably robust. and this would not happen without a lot of people keeping each other in check and collaborating and kind of these values of pluralism and democracy. And I had been sort of sad that that was going to trend away in AI. And now I'm excited about this as a direction.
You're saying basically it shouldn't lead in a centralized direction because you will have many agents and agents pushing against each other in a dynamic equilibrium. And, you know, offense, defense is a good example of this. I would much rather intuitively have
1 trillion AIs of, like, slightly superhuman intelligence join society than one AI of 1 trillion IQ join society.
Yeah. And you think we're headed in that direction of the multiple?
All of these things are things we have to push for. Like, all of these things are like, we have this incredible technology. it can be pointed and deployed in many ways. I think there are a lot of people who may have a different opinion about how it should be used and what is the best version of the world, but this is the one that I think is right.
Let's talk about starting a business. Thanks to coding models, your weekend side hustle can quickly become a serious undertaking. And we've been busy here at Stripe making that process as simple as possible. Stripe Atlas helps you with everything you need to start a business, like company formation, tax IDs, founder equity, and all the other bits that usually require someone in a suit. So if you've got a company in the making, take a look at stripe .com slash Atlas.
How's the compute shortage going?
I mean, there are all these ways where I think things at various times along the curve are going to get somewhat out of hand and there will be a moment, Like with any new thing of this importance, there are going to be moments of irrational exuberance. There are going to be moments of bust. There are going to be moments of constraint and glut and up and down. But I think the trend is toward a, we always just want more compute.
Like the demand for sufficiently capable models at a sufficiently low price, I think that trend only goes one way. and we're doing a lot of things to try to be prepared for this. We made a chip that is an incredible performance at like, you know, units of intelligence, performing at 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.
People talk a lot about... The real RSI. Yeah, people talk a lot about RSI as like a AI that does AI research. There's very little, relatively little talk about self-replicating data centers. This idea that you can use AI to control robots, and run fabs, and build turbines, or solar panels, and build data centers. But I think we have that within grasp.
And so we'll try to make a lot of really cheap compute. But aren't data centers already not labor bottlenecked? They're permitting bottleneck and component bottleneck. Can we have more components than robots? Eventually. Yeah, yeah, yeah. And all of this is how we solve this over a long period of time. My observation to compute stuff is they say the solution to high prices is high prices. That's like, OK, step one, we have the high prices for memory and GPUs and everything.
And the solutions should be here any moment now. How are we doing on the solutions front? Bad. And I'm not saying this is OpenAI's job. I feel like this is the invisible hands job, right? Like, memory's gotten really expensive, and so new memory makers are going to flood the market any moment now.
I think the Invisible Hand is failing right now. I mean, the Invisible Hand, it has like a checkered track record at best. But I think it's like not quite working here.
Why not?
Like, is it just hysteresis? I think the specifics of each area matter here. But since you started with the memory guys, like, you know, for a very long time, if you were a producer of DRAM you would invest $20 billion of CapEx to make $19 billion and you would go through these like booms and busts like maybe some years you would make $21 billion some years you would make $18 billion but like until right now I don't think I don't think making memory has ever been a great business over a long period of time and then 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 But I think they all remember their history and how many times they almost went bankrupt by making too many fabs and being promised this time it was different.
Which leads them to be conservative in their capex. Clearly. It feels like what you should expect to see is that every part of the supply chain, both new startups coming along, and maybe more relevantly, people with relevant experience getting into it. So, you know, Intel makes chips. They can probably make, they can get back into the memory game, as is tradition. As is tradition. Switch back.
Oh, I think there are like all of these ways where the supply chain, it's just like super non-obvious. An example of this is people, a lot of people want local inference and somehow they assume that like the problem is solved now because we're going to have open source models we're going to be able to run them on our computers and so we're not going to need all the stuff in the data center but 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 like you want maximum utilization in the cloud if you're
trying to deal with this very supply-constrained environment. And somehow, many people's intuition is the opposite of that.
Where else have we seen the market respond to high prices, not just in memory, but across the supply chain? Turbines? Batteries? Concrete? Sorry, the market is responding there in high prices. Are you seeing any interesting market responses in terms of high prices pulling in innovation? Some 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.
There's new actuator makers and that kind of stuff.
Yeah, honestly, I mean, maybe if OpenAI sells our chip, maybe that would be an example.
Okay, yeah, I want to get to your chip because that was going to be my example. I think the solution to high prices and short, you know, limited demand, or sorry, limited supply, excuse me, is you make your own chip. Presumably you're using a lot of AI to make the chip. I would have thought that's hard because I would have thought that chip design is one of those places limited by the public tokens you can get, where there aren't actually that many blog posts on the public internet on how to build a, you know, leading edge two nanometer chip.
Well, 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. In terms of how to do the chip design, like to build the file that you eventually send it tape out, that is well understood. And that is a great problem for AI. And given that the world is going to be like limited in its silicon capacity, we all, we have like a shared interest.
One of the things that I'm proud of about OpenAI, and maybe it sounds obvious but it was maybe like not as obvious at the time is 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. What are other examples of that? AI research, software engineering but also increasingly like procurement or things that people think of like very outside the realm of what most people are using AI to do.
How does AI-driven AI research actually work? Like the math proofs where you say, or the find the PHP vulnerability where you say, please find me some AI advances.
So I think we don't know yet how it truly is going to work. We know today how to say, like, find me a 5 % computer efficiency win.
But it's that high level. Like, I could be an incredibly successful AI researcher now.
I don't know about incredibly. But yeah, if you were looking for a new job, you can come do it. What I don't think anyone knows how to do yet is to say, like, go discover a new architecture that is as big of a single leap forward as the Transformer was. And I think that's where things are heading. And what's missing there? Smarter models that are smarter in this dimension. The high-level planning and conceptual reasoning. I think if you apply the same pipeline that we applied to get our models to be so good at math to AI research, like, it will probably work.
Now, per an early part of this conversation, I kind of think that, 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 go. But a lack of progress will come from that decision, not from the AI's inability to do good research.
Concretely, if I'm using AI to develop improved AI, because we should expect GBD7 to be really good now, because you have GBD6 and whatever unreleased version of it to develop it, and so my expectations are sky high. But I come into work, you know, with my backpack, and I come into the OpenAI office in San Francisco, and I sit down at my desk. Like, am I saying, find me more intelligence in this model? Am I giving a real direction of research?
How directed does it get? What does the iteration loop look like? Just what does RSI, because you have a front row seat to RSI, and I'm curious what it looks like.
one of the sort of interesting things that people talk a lot about in OpenAI, which sounds very obvious, sounds like a good answer, just like, well, build me build me RSI or whatever, is what are the questions that you're supposed to ask of GPT-7? Probably it's not enough to just say, like, make me RSI. And also, like, it's probably not enough to say, like, buy me a 5 % compute efficiency win. I think this is, like, one of the most interesting debates going on right now.
And the answer is not just what question to ask you, but it's things like agent swarms. Like, how are we going to use these models in a way that actually maximizes what we can get out of them? I think we're going to learn a lot about that in the next three months.
OK, but it sounds like you guys do not have a codified OpenAI standard way of doing an AI-enabled AI development. It's still lots of experimentation. It's lots of experimentation. Yeah. When we're talking about the token budgets, it just strikes me that this is one of the big unresolved questions for AI use in enterprises generally is how, you know, we don't have hard budgets for individual engineers. You don't get cut off at a certain point. You're like, sorry, you've spent all your arcade tokens.
Maybe we should, but we don't. And again, I think we all believe in the scale up of AI. But if you take that seriously, it means the dollars of compute or AI per person will go up. And so I feel like we're pretty soon to now actually just running into this question of how you actually decide how much spending power to give individual people inside the organization.
There was a time when a lot of companies had this thing in their Google Calendar which showed the approximate cost of each meeting. in terms of like everybody's comp that had to be there. And I think that was always still an underestimate of it. And there was this observation that like Paul Graham made it at one point, you know, and many other people made it too, which is that there was like, there are people at companies that have incredible power to save like, you know, $17 on the snack budget of the office every week.
And yeah, you can have these meetings for like unbelievable amounts of money that are adding no value and that there's kind of like no equivalent of the ability to stop that. I think that most use... Okay, so I'm like really talking about my book here and I'm very biased, but I genuinely believe it. That use of AI is like not a company's biggest spend problem relative to other things that happen. And even if it's like somewhat inefficient, it is not where I would focus.
Although hang on, that's not scaling law pilled enough. Like, I'm not... Okay, two years from now.
You want to go send all Stripe's revenue, all Stripe's profit on open AI services and then complain about this? I'd be delighted to have a conversation.
But we're on a path to that. Great. Yeah. We now spend 1 % a shorter magnitude of Stripe's OPEX is tokens. That is obviously... That feels too low. Yeah. So we're working on fixing it. Okay. And it'll get to 5 % and 10 % and 20%. And then you have a real question. Do you have an intuition on what the right number is? Like what it should be today? What it should be today is maybe tricky, but steady state, 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?
I think for many companies they answered us clearly yes and and I think there will come a time when the right thing to do is to really optimize that and also like another thing I'm proud of about OpenAI is we've tried to really drive the cost per unit of intelligence down and down and down I think we've been like a good spokesperson and good champion of this but still Still, I think the problem, even at current costs, I think there are more companies the problem of not using enough AI rather than using too much AI.
Do the dollars per gigawatt of leading edge AI go up or down? That should go down. Down? That should go down. Why? Because it's gone up to this point, right?
I mean, the reason it could go up is that if AI just gets increases in value per what it can do per watt so quickly, and there's such a supply shortage that people are willing and to spend much more.
But you would view that as compute not keeping up.
But I would view that as compute not keeping up. And if I look at the fundamental cost of turning the sand into intelligence, that should go down. The cost of energy should go down. The cost of chips should go down. The rest of the cost of the supply chain should go down. If it doesn't go down, I'll be very sad. But for it to go down, given how much value there is in this, supply has got to go way up.
OK, so you would like to see a compute glut where we start producing a lot more memory, a lot more data centers, a lot more of all the inputs. And as a result, anyone who wants a computer for whatever reason can have one.
People give different numbers for the kind of the cost per gigawatt measured in like the five-year total cost of ownership right now, but like let's say, let's pick $70 billion as a number. I think if that number were... That's the cost to install. And run it for five years. Yep, okay. I think if that number were like 20 instead of 70, we should be much happier. And we should still demand to get 10x better performance out of it with smarter models every year.
Presumably you don't think 2 is a good number. Something's gone wrong if we're seeing 2 because it would be surprising if we were... Or is 2 better than 20?
Why is it not? I don't think that will happen. But why do we not want that number to be as low as possible?
I don't know. This is your theory. You have a view that generally deflationary technology is good.
I do believe in that. Yeah, I'm pro-deflation in a sense. Pro-deflationary technology.
So how are we going to get there? What's going to happen between now? $2 a gigawatt. I can't imagine. What's going to happen between now and $20 a gigawatt? $20 billion.
$20 a gigawatt would be really great. Even to get down to $20, you're getting into the realm of energy's got to get cheaper. But certainly, 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.
You could just get there with eliminating profits, or do you need to also make the energy cheaper? I don't know the math off the top of my head.
I mean, if you assumed like all materials were free and no one made a profit, maybe you could get there in the current energy envelope, but only just. I see.
Personal agents are the most exciting thing happening in AI right now. There's many exciting things happening in AI, but one of the most exciting things, you guys launched Dots a few weeks back at this stage, and there's a bunch of other people also launching in the field. And what will be true of the winning personal agent?
It's like intelligence, context, tools. It's a limited number of things. And someone is going. I think no one has yet figured out the thing that people... People like these, but this is not something that people are like, I cannot imagine my life without this thing. In a way that when people talk about coding agents or even ChaiGPT, they do say. and someone, I think someone soon is going to figure out what that looks like and you talked about one of the reasons why this doesn't quite work yet but I think this is like a small number of ideas away from a super mainstream thing
Oh interesting, you think we have not had the Macintosh moment for personal agents? I think we have not Do you have any sense of what the shape of it will be?
I think it'll look a lot like the current you know, Dots or Muse or GrokBot or whatever but it'll just work like it'll just work it'll be reliable you'll like never have you'll just like give it more
complex tasks and it'll do all of those things computer use dynamic is very interesting to me because not everyone is embracing the new agentic future with open arms but it's not totally within their control because you can just have an agent using a computer Where do you think that dynamic goes, the sort of cat and mouse?
Like rough instinct is the world will... I mean, people are going to fight this in both directions, but eventually converge on something like you're allowed to tell an AI to use your computer on your behalf, including software.
People always talk about that, but isn't it a bit like the semantic web? It's like a really nice conceptual idea. Yeah, I was going to say, I don't think this is so black and white.
I don't think I have a super high degree of confidence of how this is going to play out. As a consumer, I feel like somewhat annoyed if I can't have an AI automate tasks on my own computer. Like I want it to feel like it's my agent doing what I ask, but then it gets like I don't have to like go do something.
Yeah, I note a very interesting dynamic right now, which is every company is really eager to add AI to their products in a very specific way, which is there is a tab or pop-up widget or something that gives you a messaging type UI. And it's often surprisingly capable. And so you can say, hey, I actually want to return my most recent order. And then you're kind of surprisingly impressed when it actually is able to find your most recent order.
And, you know, it's a bit janky and it can do all those things. However, no consumer really wants this UI because consumers have strong, not immovable, but strong preferences for their existing primary chat app, the kind of top of wallet app that they open when they want to go to do something. And what they want is to be able to say, hey, go find my recent order on this business and refund it or something like that.
And I noticed that businesses are really reluctant to allow people to use those businesses inside another AI app because like, oh, we're going to lose control on our motive, whatever. And it just strikes me as fine because your defensibility probably did not come from your ability to put together HTML. Your defensibility came from the brand loyalty and the product selection. And the new UI is that agentic interaction inside ChatGPT or Gemini or whatever it is. But all businesses disagree with me.
And they're very good at business.
I mean, again, I have a bias here. But I so directionally agree with you. I have such an extreme preference to buy something from Amazon versus some random e-commerce vendor that I don't know if I can return to or I don't know if I can like trust to deliver it quickly or something like 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 pay a higher price I have just like such yes and it's nothing to do with like my experience of clicking around amazon .com exactly and now I think you could say we have not done a good enough job offering partners
a platform inside of ChatGPT if they don't want this to happen. But, and again, maybe I'm really biased here. I think Amazon's like, I think they're in a great place of commerce moves like much more agentic.
Amazon in particular has this amazing defensibility via the fulfillment network. And so you can use Amazon kind of from anywhere. I super trust that. Yeah. No, I'm just really struck by this dynamic. Like my new thinking, in particular with regards to agentic commerce, but also anything, is any time there is a web form field that you as a human are filling out, these days that is below your pay grade. Because you are probably not inventing that information for the first time.
Like we've all filled out the medical forms where it's like, oh, your family history or whatever. And that's so frustrating because how many times have you filled out your dad's and your mom's medical history and things like that? It just feels so wasteful. But the million versions of that, you are never inventing new information for a web form.
I would be so happy to never do fake work at a computer again. Yeah.
And my argument is that any web form is basically fake work.
I agree. But I was one of the very earliest adopters that I knew of, of those plugins that put in your name and address very clumsily. Because I was just like, I cannot do this again. But I kind of feel like that using most of the web at this point.
Yeah, I just find this phenomenally interesting where businesses right now are really attached to their HTML and mobile apps as sources of business defensibility. And it just feels very trans... Like, either they're wrong and it's not a source of defensibility, or it may be, but it's transitory and they just need to get over it. The last example of this that I think we really witnessed
was, like, the app stores. The iPhone app stores specifically. And I remember every startup was like... or many startups, like most of the YC batch was like, I'm going to build an iPhone now. This is like new distribution. This is awesome. I can do a new thing I couldn't do before. And many new startups got created. And all of the big companies felt like they had to make one, but they kind of didn't want to because I'm like, oh, I'm going to hand my relationship with my consumer over to Apple.
And it was great for most people.
It was just a new UI. Yeah. Okay, I'm going to go more rapid fire. People complain about bench maxing. Is that a real concern? Are OpenAI models bench maxed? Are you actually Rick Rubin-ing it where you just feel like the vibe of this model is right and you don't care what Sweebench gives it?
Maybe there are two things we should really care about models being good at. One is scientific discovery and doing things that people cannot do. And the other is being enjoyable to work with. And I don't think we have great benchmarks for being enjoyable to work with. I'm actually not sure we have great benchmarks for the smartest scientist either. but we endeavor not to bench max and I kind of I keep thinking that bench maxing is like over as a phenomenon and people keep doing it
so are you better than other labs at not getting sucked in by the siren call of bench maxing? I believe we are better at that it feels like I don't think we've always been better at that in history
I believe we're better than that right now
it feels like the industry is still figuring out the being nice to work with where there was the glazing era and more recently there was the stroppy ornery era of models that were anti-glazing and as a result tried to disagree with you and nitpick at you on every minor thing the current ones are pretty good actually it's gotten I think the latest models
are much better than we've had in a long time
but it also feels like there's probably higher peaks to its end
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.
Yes. After working with him for many years, what's your grand unified theory of Greg Brockman? I think Greg Brockman is like one of a kind in the best sense. He's singular on the Greg Brockman scale. You just kind of have to be there to understand what that is.
But I think it's like a combination of like, yeah, intensity, like monster productivity, extreme dedication, and like focus on succeeding at the mission. It's very cool.
you hate tokens as a metric for anything really um pricing or uh just kind of it's not the standard unit of ai and yet we found ourselves there you know we did and you might say outcomes but i don't think it can be outcomes because outcomes is too yeah different across different use cases it's
what's the right standardized unit 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. And you're just going to give it a prompt, and we're going to give you a price back. And we're going to do it. And it's going to be like AWS. You just do something and we send you a bill. But the customer response on that is so negative.
They're like, I need to know what I'm paying for somehow. Yes. And I asked a few people, would you rather pay more for the same output but have it priced in some number of consistent tokens or something? even if we inflated the tokens. And I'd much rather you inflate the tokens and pay more because I know there's a real thing there. So I think you are right that people don't just want us to give them a price because they feel like over the long term they're going to get screwed somehow.
In the same way that maybe people would rather pay for CPU hours rather than just getting their thing done, even if it was less expensive.
You know the way the chip gate length? Yeah, the nanometers has very little relation to reality these days. Will we get something like that in tokens where you guys should start saying, here is an intelligence equivalent token like, I can't believe it's not butter.
Of like the smallest feature. It's like made up, I hope not. I mean, it offends me that we like use the word nanometers for something that has nothing to do with nanometers.
I know, it's very silly. You're not like that.
I would rather not make up a fake unit. tokens are a kind of fake unit. I mean, they're real in some sense,
but they're not a good unit. They're fake from a business value point of view. Isn't the only defense of tokens that if you're wasting tokens, we should fix that pretty soon? Like, we shouldn't be streaming. Claude Shannon would not approve of us just streaming noise rather than signal. And so there will be a... There's a floor of the information theoretic most useful tokens on both sides. and we'll kind of come back to that.
Yeah, I mean, right now you have to think about whether you should want the 10 cent per million tokens out of a small model or $10 per token, tokens out of a big model, and then whether, like, how long the chain of thought is likely to be. And that's just like cognitive overhead that I don't think actually adds any value. Is the chain of thought included in the price? In what we charge? Most people do. By the way, the Guinness was not as bad.
It definitely not as good as an IPA, but not as bad. It's good. Yeah. I mean, IPA is great, but it's pretty good. We like it in this. Yeah.
At least in this podcast. Thank you.