Main Topic: The conversation revolves around various topics, including career advice for people in their 20s, the importance of accumulating expertise in a domain, the limitations of modern universities, and the potential for institutions like ARC to provide a more effective model for scientific research. Additionally, the discussion touches on the challenges of scaling AI, the importance of craft and beauty in software development, and the potential for Stripe to play a significant role in facilitating global trade.
Key Points:
Participants:
Takeaways:
We discuss:
* what it takes to process $1 trillion/year
* how to build multi-decade APIs, companies, and relationships
* what's next for Stripe (increasing the GDP of the internet is quite an open ended prompt, and the Collison brothers are just getting started).
Plus the amazing stuff they're doing at Arc Institute, the financial infrastructure for AI agents, playing devil's advocate against progress studies, and much more.
Watch on YouTube. Listen on Apple Podcasts, Spotify, or any other podcast platform. Read the full transcript here. Follow me on Twitter for updates on future episodes.
Timestamps
(00:00:00) - Advice for 20-30 year olds
(00:12:12) - Progress studies
(00:22:21) - Arc Institute
(00:34:27) - AI & Fast Grants
(00:43:46) - Stripe history
(00:55:44) - Stripe Climate
(01:01:39) - Beauty & APIs
(01:11:51) - Financial innards
(01:28:16) - Stripe culture & future
(01:41:56) - Virtues of big businesses
(01:51:41) - John
Dwarkesh Patel
Today I have the pleasure of speaking with Patrick Collison, CEO of Stripe. Patrick, first question. You have an excellent compilation of advice on your blog for people 10 to 20, and you say there that once you turn 35 you'll write some for people in their 20s. What advice do you have for us now, the people in our 20s now? When's it coming?
Patrick Collison
I haven't really thought about that. The one I've been wondering about recently is, I said for that advice for people in their teens, they should go to San Francisco. And I wonder for people in their 20s if they shouldn't go to San Francisco. I mean that glibly, and I think there's a significant set of people who should, in fact, go to San Francisco.
But the thing that I wonder about is there is a set of career paths that some people ought to pursue and would derive most fulfillment from pursuing, and that are really valuable for the world if pursued, that require accumulating a lot of expertise and really studying a domain in tremendous depth. San Francisco valorizes—and this is also San Francisco's great virtue—a kind of striking out on your own, iconoclastically dismissing the received wisdom and the founding archetypes and lore of Steve Jobs and Bill Gates and all the rest. I'm way less successful than those people, but to some extent, Stripe, as much as it fits a pattern, is an instance of that pattern. That's great, and I'm happy that phenomenon exists in the world. But I don't think the world needs lots of other things.
I don't think San Francisco particularly—I'm using San Francisco as a kind of metonym for a cultural orientation—encourages the pursuit of really deep technical depth. We're recording this in South San Francisco, which is most noteworthy in the corporate world for being the headquarters of Genentech. Genentech was co-founded by Bob Swanson and Herb Boyer, and they produced cheap insulin for the first time with recombinant DNA. Herb Boyer couldn't have done that at age 23. He first had to accumulate all the knowledge and skills required to be able to invent that over the course of a multi-decade career. I don't know what age he was when he finally went and invented it, but he was not in his 20s. I feel San Francisco perhaps doesn't culturally encourage one to become Herb Boyer.
Yesterday, at the time of recording this podcast, Patrick Hsu, one of the co-founders of ARC—which maybe we'll speak about later in the show, a biomedical research organization we started a few years ago—announced this new phenomenon of bridge editing, a new recombinase where you can insert DNA into a genome. It's pretty early, but it might turn out to be quite consequential. In order to do something like that, you have to study for a long time and acquire a lot of basic and not-so-basic technical skills.
As I think about advice for people in their 20s, I'm not going to normatively pretend or presume to know in which direction one should go in one's life. Obviously there are successful examples of basically every strategy. I'm really glad that you're doing what you're doing at age 23, 33. Information dissemination is a really valuable thing in the world. The guy who last I heard was in the lead for the Scroll Prize—Nash told me he learned about the Scroll Prize listening to your podcast. Creating increasing catalytic surface area of certain kinds of information is a valuable thing in the world. I'm very glad you're putting the podcast out.
I don't presume to know what people should do with their lives. But in as much as trying to give advice, especially if they're reading my advice and not someone else's, maybe they're thinking about career paths that look directionally like mine. My advice might be that maybe you should do something like what I did or am trying to do, but there are other paths as well. A lot of really important invention in the world, and a lot of the things I'm most happy are happening, actually require a very different trajectory. There were counterfactual versions of my life where I pursued that path, and who knows how well it would have worked.
Last point on this: San Francisco is just very status-oriented. Maybe everything is status-oriented, so that's tautological. But maybe what I'm really saying is I feel San Francisco's entrepreneurs are held in excessively high regard in my view. I like entrepreneurs. Entrepreneurs as an aggregate group in the world—all the companies built in Stripe—I think are great. But there's a strange version of it in San Francisco that should not be people's only fixation.
Dwarkesh Patel
I like this and I like you because you have this sense of contrarianism. The things people are expecting to hear from you in any given moment, you just really want to tell them the opposite. When EA was a little bit more popular, you were like, "Here's the problems, here's why progress is important." And when it was down in its depths, like, "Hey guys, pay attention." On this particular piece of advice, Michael Nielsen says that every field in science has way too many adherents or way too few, but the market is almost never in the right equilibrium.
Patrick Collison
I think the reflexive contrarian way, I would say that. But reflexive contrarianism for the sake of it is also tired. If you're just contrarian to the prevailing mood, then you're just following the prevailing mood with a sign bit inversion. I don't endorse that either. But I think the herd is a really powerful phenomenon. One of the learnings of my adult life has been that everyone knows and says you should be very wary of following the prevailing tides and moods and winds, but it's freaking hard to do in practice.
Dwarkesh Patel
What practically does that look like to hone your craft in any of these disciplines that take a long time? You've spoken and tweeted about some of the problems with modern universities. Is that still the de facto path if you want to be the great biologist at Arc Institute or something?
Patrick Collison
In many domains, I don't know. In hardware, which is not a small domain—most things in the world involve stuff and things—I have no facility with or experience with doing things in hardware. If you want to become a super skilled practitioner there, what's the best career path? I don't know. Maybe it's to drop out and join SpaceX or something. I'm not necessarily endorsing pursuing the most establishment and credential-oriented path. I think people should try to find the gradient of maximal learning in whatever it is they care most about.
For biology, not that I'm a biologist, but it is very clear that in order to do really good work, there are a lot of bench skills one has to acquire. There is a lot of actual specific knowledge. Any kind of life wasn't designed with neat fundamental principles the way that maybe physics was. A lot of it is evolved and contingent and messy and complicated. There is a lot of specific factual stuff to learn. For those two reasons, there are very few successful pure autodidacts in biology. In virtually every case that I'm aware of, you have to have had direct experience in and with a top lab where you're seeing how people actually do it in practice.
Maybe this ties back to what we were discussing previously about founders and what they learn from each other. There's an interesting book, *Apprentice to Genius*, that follows three or four generations of scientists—someone who mentored somebody else, who in turn mentored another scientist. They're all extremely successful. The book is a reflection on what they all did, but also on what was transferred. One of the most important and subtle questions in science is problem selection. How do you choose what to work on? No one tells you what to do, and you have to answer this question multiple times. With a company, in some sense you just have to decide once and then it's an iterative process. Whereas in science, you're frequently pursuing completely new problems. You need to choose something sufficiently important and hard that it would be important if you succeeded, but also not so intractable that you can't make any progress. The book describes how this is part of what the mentees learn from their mentors.
Another thing they talk about is learning about high standards and what high standards actually are. When I talk to people in other domains, this is so frequently the thing I hear—that when they worked with X person or Y organization or in Z environment, they learned what great actually is. That permanently changed their sense for what their own standard for their work ought to be.
Maybe one version of what people in their twenties should do is get some ideas about domains you're interested in or care about, but then figure out where you can learn the highest standards. Where are the highest standards embodied? Where can you go and experience that firsthand?
Dwarkesh Patel
Before we get back to Stripe and ARC Institute, I want to touch on the progress studies stuff for a second. There's a view that says if we improve the NIH 10% or whatever percent, are we really making a dent in the fact that ideas are getting harder to find over time? How much of a difference do institutions make anyway? If it's just about the number of researchers and how many people in your society you can put into research, it's not like Singapore can have a much more effective scientific institution that lets it compete with America in science. What's wrong with that intuition?
Patrick Collison
Noah Smith and others have talked about something like "moneyism"—I can't remember the term he used. The idea that we assume there is some kind of constant elasticity between investment in some particular outcome, like building a semiconductor factory in Arizona or a new bridge, and the outcome of the factory or the bridge. One, the conversion rate between those inputs and the output is not a cosmological constant. Maybe any of these things could be done for half or a tenth of the cost. Two, there are even deeper questions about whether it's possible at all, or what else would have to change for it to be possible. By just talking about these things in funding and dollar terms, you're making the implicit assumption that the only relevant constraint is the financial one, when in practice maybe it's permits, labor shortages, or other things.
In the context of the NIH and science and R&D, I'm really skeptical of this same approach being brought to bear where we can just talk about the amount we're spending on R&D and think that's implicitly a useful measure of the output. To a fairly close approximation, there were around 1% as many practicing professional scientists in the US pre-World War II as post-World War II, or say even 1950. The other epiphenomena in papers or patents and so forth tend to follow pretty similar ratios. We got a lot of pretty good stuff in the first half of the century. Despite increasing the amount we spend by between two and maybe slightly more than two orders of magnitude—not quite three—it's not clear to me that there is a direct linear relationship.
When analyzing the NIH or how we should pursue any of this stuff, I'm inclined to try to get way more concrete and tactile. Try to think: what would success here look like? What is happening today at the micro scale? What are the actual problems? What could success look like at the micro scale? What might it look like to scale that up?
To give one pointed example: we ran a survey of the FastGrants grant recipients after FastGrants, asking about their normal work—not about anything to do with FastGrants itself. We asked them if they had flexible funding—if they could spend their current research dollars wherever they wanted, not if they had more research dollars, just if they could direct their current dollars however they wanted—how much their associated research program would change. We gave them three options: not much, a little, and a lot. 79% said a lot. Four out of five said their research agenda would change a lot if this constraint was removed.
Should the NIH funding level be X or 1.1X or 1.2X? That seems like a bad way to analyze this question compared to, for example, how bound and constrained should an NIH grantee be in choosing their research agenda? If their judgment was way better than that of the committees—not saying it is, but maybe it is, who knows—maybe there's a 5x improvement to be generated just by making that one switch. I'm very skeptical of these financially oriented frameworks.
Dwarkesh Patel
If it was particular to the scientific institutions, you'd expect things that are disconnected from the NIH-specific structures to show it. You've talked a lot about it getting harder to find papers. Sector through sector, it's not like NIH is running Moore's Law progress. But even there, you see you need exponentially more researchers to keep up the same level of progress. It does seem important to have these level effects that are one-time, like in the case of COVID where we needed that level effect right now. But when we're framing it in terms of hundreds of years from now, this is going to be the thing that increases growth rates—a framing that is also supplied when talking about progress. Does that make sense in that context, when all these sectors are seeing these slowdowns that seem consistent with just how the economy and science progresses over time?
Patrick Collison
I don't know is the short answer. I think it's really puzzling. The constancy of US GDP growth is one of the weirdest things. I don't have a good explanation for it. But I also don't think the obvious thing to do would be to shrug and say it's overdetermined and that's just how countries work. You can look at other countries—it's obviously manifestly not the case. So what is it that's weird and special about the US?
The thing I wonder about in a lot of these cases is you could get many of the observed system characteristics if we weren't actually adding productive capacity. That's a simple way to explain a lot of it. If you're just adding exponentially more unproductive capacity, then on a stylized level, a lot of this stuff would just fall out of it. I'm not saying we're necessarily doing that, but there are lots of ways where that could be what's effectively going on, even if it's not the case that the marginal people or things or organizations themselves are bad. It's just somehow how the components interact.
The fact that you could get these exponentially diminishing returns through the addition of ever more non-productive capacity makes me not persuaded that the low-hanging fruit case is necessarily true. I give some weight to the prospect that it is fundamentally structural, cultural, or organizational.
To give a micro example—a very basic and obvious one—I think it's interesting to compare the SpaceX R&D budget and the NASA R&D budget and actually look at those two time series together. Maybe we're just returning to the financial point again, but it seems pretty clear that the trajectory of NASA's efficacy has not fully followed the trajectory of its inputs.
Dwarkesh Patel
The point about the marginal inputs we put into science have not been as highly effectively used or as high quality as what was before. The 1X is a much higher quality 1X than a 100X. It's not clear what you do to fix that. If it's just a case that there's a limited amount of John von Neumanns in your society that are part of the pre-World War II 1X, it's not like we can just put 100X more John von Neumann-type physicists into science. If the binding constraint is the number of John von Neumanns, then that's bad news. There's not a lot we can do on the margin.
Patrick Collison
I'm not sure that it is. I keep going back to the cultural and sociological point. Gerty and Carl Cori ran a lab at the University of Washington, St. Louis, and six of their students, if I recall correctly, went on to win Nobel prizes. They had a well-known lab and they got good students, but they weren't the most prestigious lab in the world. It's not like they got to cherry pick every year the single most promising person. Something was going on there. There's a book about it that tries to get into this a little bit. I don't know that I can figure out quite what it was. There was also some good fortune where they got into molecular biology at a good time. But I think there are these hopeful data points where they were obviously extremely brilliant people, but the thing that distinguished them and their students was not that they were these seven-sigma Martians. I think rather that they found organizational structures and cultural practices that really worked. Those are at least in principle more replicable. You might still say, "Fine in theory, but how do you actually do that?" I think that's the big open question.
Dwarkesh Patel
I think that's a great point to talk about Arc Institute. I think you just answered this question basically. But it's not exactly like biology research is something that society has neglected. What's the theory of change here? Is it a story similar to Stripe in that if you get the right people, even though there's tens of billions of dollars of biology funding, getting the right people and the right culture and right dedication is what it takes?
Patrick Collison
Even though there are lots of scientists and lots of universities, there's a lot of homogeneity today in how science, and in particular biomedical science, is pursued—basic research in an academic context before there's any commercialization or prospect of it in sight. I don't know that the model is necessarily a bad one. We're not particularly claiming that it's a bad one. But the construct of universities, labs, PI—a principal investigator running the lab—that person applies for grants primarily to the NIH, maybe supplemented by other sources, and grants are reviewed by committees with pretty study sections, as they call them, with pretty rigid scoring criteria. That's the structure. It just seems suboptimal to me.
Homogeneity is bad in basically any ecosystem, especially ecosystems where you're seeking tail outcomes. We thought that from first principles, other models should be possible, and we had specific ideas as to how one particular model might be a good idea and complementary to the status quo.
In very short terms, what's different about ARC is: one, scientists are funded themselves to pursue whatever they want. It's curiosity-driven research, whereas NIH grants are given for projects. Two, we build a lot of in-house infrastructure so that scientists can draw upon other platforms and other capabilities that they don't have to go and build and maintain themselves. In the standard university academic context, scientists would virtually always have to do that in-house. Because of the natural skill constraints on any given lab, that effectively circumscribes the ambition of a possible research program. Three, we try to provide career paths for people to remain in science if they don't want to become principal investigators. The university structure kind of co-mingles the training purpose of academia with the execution, where the people doing the work are typically the grad students and the postdocs who are themselves, at least nominally, on the career path of eventually becoming principal investigators.
There are lots of people who, for all sorts of different very valid reasons, love science and love the pursuit of research but don't want to be a manager running a lab, choosing their own research programs, and dealing with all the overhead and typically grant applications that are concomitant with that. At ARC, we have a real emphasis on hiring scientists who have finished their postdocs, finished grad school, and that's just what they want to do in their lives. There isn't really a career path for them today.
One of the things that's really exciting about the discovery that came out yesterday—this new bridge editing technology—is that work was led by one of these senior scientists who'd finished his postdoc. It's not clear to me that he wanted to go on to become a PI, but he loved science and he's an amazing researcher. He's able to go and have that career at ARC. In addition, the prospect of these mobile elements being usable in this way for genomic insertion is a pretty speculative, out-there thing. Had he applied to the NIH to go and pursue that—he didn't, so I don't know what the outcome would have been—but Jennifer Doudna's work was, if I recall correctly, funded by DARPA because her CRISPR NIH applications were rejected. Carolyn Kuriko's NIH applications for mRNA vaccine work were famously rejected. It at least seems very plausible that it wouldn't have worked out.
All these things are random, and I can't make any definitive claims about what would have counterfactually happened. But it seems plausible to me that this thing announced yesterday wouldn't have happened, or would have been less likely to happen, in a different environment.
Dwarkesh Patel
When we think forward 10 years or 20 years, this specific line of research where you understand the effects of degenerative architecture around different traits, and also you can edit, invert, insert whatever the DNA arbitrarily—you saw sickle cell anemia, you've done the obvious things. What does that lead to? What are you excited about?
Patrick Collison
The thing I think is really interesting about it is using it as a new kind of telescope. When people hear about CRISPR, there's an obvious excitement and a legitimate excitement around using it to cure things directly in the body, using it as a kind of therapeutic. But you can also use CRISPR to try to figure out what's going on in cells and in cell cultures in a kind of structured way.
The body is interesting in that it has this switchboard of 20,000 genes. With CRISPR, you can systematically go and perturb each gene one by one, like mashing all the keys in sequence, and try to figure out what the effects of perturbing this versus that are. If you do that in a cell culture where you can subject the cells to some stressor or some treatment, you can kind of see differentially how different perturbations affect different cell outcomes. Or you can just use it for synthetic data generation more broadly, where you could perform all these perturbations and then sequence and see what's happening in the cells. Single cell sequencing has come a long way.
The point is, there's a lot you can do with all this gene editing stuff for discovery and for data generation in the broadest sense. That's really compelling because for a lot of diseases, they're complex in the field's jargon—meaning they're complex in the colloquial sense, but specifically complex in that they're not infectious. They're not just some pathogen getting into you. And they're not monogenic, like Huntington's where it's one specific mutation. Instead, it's some combination of environmental factors and maybe some genetic factors as well. It's somewhere in between. That includes most autoimmune diseases, most cancers, to some extent cardiovascular disease and neurodegenerative disease—the big ones we haven't yet solved.
Coming back to these functional genomics technologies, what's interesting is trying to figure out how the genetic component of those diseases happens and works. Even if that's only a small contributor, it can potentially shine light on what the general pathway is. This is speculative—none of this has actually happened—but by figuring out the genetic interactions between genes and, say, Alzheimer's, can you figure out how Alzheimer's arises, which we don't understand today? Once you understand how Alzheimer's arises, maybe you can use conventional technologies and targeting to figure out how to inhibit that or modulate those pathways. That's what we're really excited about from a functional genomics standpoint. There's kind of an AI angle as well that we could hook up if you want.
Dwarkesh Patel
How do you think about the dual-use possibilities of biotech?
I am sympathetic with the idea that if you think of prior technology, just like Google search or even just the computer itself, you could forecast in advance, "Oh, this has all this dual-use stuff," but for some reason, history has been kind to us and maybe we should just... Chester's meta-fence here is keep doing science. But with biotech, there are specific things you can think of with this specific technology where you can imagine some nefarious things. How do you think about that? Why not just focus on ameliorating the risks first?
Patrick Collison
Well, I don't think the binding constraint on harmful use of biotechnology or bioweapons today is pure biological capabilities. If some set of incredibly capable, intelligent people wanted to cause tremendous harm with pathogens, or with something biological, they wouldn't necessarily need to invent anything new. They would just need to apply currently known techniques in a malevolently directed fashion.
I think there are some concerns and risks with respect to things that don't invent new technologies, but do make them more accessible. The question is, how would the effect on the world be if there was a sufficiently sophisticated LLM that could help anybody synthesize and disperse smallpox? I don't know that the laws of physics prohibit such an LLM existing. I presume they don't. Would the world be fine if such an LLM was widely distributed? Maybe, but maybe not.
But my point is, I don't think knowledge at the frontier of biology is the relevant margin here. We don't need crazy AI risk to motivate this. The world is perfectly capable of originating really severe pandemics and pathogens itself, plus all the other diseases that are not pathogenic. Whether we care about the possible dual-use harms you mentioned, or we just care about things that already exist, to ameliorate both, we need enhancement of our capabilities. There are a lot of biological problems that we don't today know how to solve.
If one were to do what you're proposing and try to advance the defensive side, I don't know that what one would do would necessarily be that different. There are fundamental capabilities we would presumably need to have that we don't today have. By trying to solve current human diseases, you're probably also pursuing something pretty close to the best steps to solve the potential diseases that malicious actors could cause in the future.
Dwarkesh Patel
That makes sense. Zooming out from bio-risk in particular, how are you thinking about AI these days?
Patrick Collison
I think everyone has to be in a state of high perplexity. The verdict one might have given at the beginning of 2024, or at the beginning of 23, 22, 21—back over the last eight years—those would all have looked pretty different. Maybe Guern might have scored the best from 2019 onwards. But broadly speaking, it's been pretty difficult to forecast. The basic position has to be one of some degree of humility.
As your blog post identifies, the big question right now is to what degree scaling laws hold. If they hold, then what exactly is it that we're approaching? We don't necessarily know the shape of that thing. How one should feel needs to be very sensitive to the exact parameters of those curves, and I just don't think anyone knows what the true value of those parameters actually are. It's clearly going to be important—it's already important today—and it has a pretty central bearing on both Stripe and Arc. We'll see.
Dwarkesh Patel
I wonder if the meta lesson we got from COVID, with things like Fast Grants, was that you obviously can't predict these things in advance, but the most important thing is having competent individuals who can synthesize and organize information, and having new initiatives and institutions to get the right thing done when the thing is happening.
Patrick Collison
Yes, the adaptability premium is probably going to go way up over the next decade.
Dwarkesh Patel
With that in mind, and I know you already have a couple of day jobs, but something like Fast Grants—when the time comes down to it, you'd be one of the top people you could think of in terms of having expertise and respect in a wide range of domains and competency as a leader. Just keep it in the back of your mind, or maybe in the middle of your mind, given how far we are into the transition.
Patrick Collison
Fast Grants was three beloved squirrels in a trench coat. Or I guess I was one of the squirrels, so I'm self-fulfilled of it. It was also Tyler Cohen, who's an amazing person and a great friend, and my wife, who's also one of Arc's co-founders. Fast Grants was not this giant, impressive office that would qualify me for anything at all.
Dwarkesh Patel
But it doesn't have to be giant to have that kind of big impact.
Patrick Collison
As an objective matter, that's true. John and I try to be very self-aware of the limits of our expertise, which are very proximate to us. If something like that was necessary, look at Operation Warp Speed. They chose a super effective domain expert, Moncef Slaoui, to run that. It was monstrously successful, truly remarkable. I don't know who the Moncef Slaoui of whatever the problem in question would be, but my recommendation would be: figure out who Moncef is and go hire Moncef. I think anybody who deemed me the Moncef of that thing is probably mistaken.
Dwarkesh Patel
I don't think you're being too humble. But staying on Fast Grants, now we have the retrospective of how effective the Fast Grants recipients were compared to other grants given out by the NIH or NSF. To your knowledge, what has been the reaction of these institutions to the discrepancy between the speed and effectiveness of Fast Grants? Have they analyzed their protocols?
Patrick Collison
Not to my knowledge, but I don't want that to sound like an indictment. Maybe they've done a lot of reflection and I just don't know about it. I don't know anything about the response at CDC, FDA, NIH, NSF, or any of the relevant organizations or their international equivalents. None of what I'm saying should be taken as critical of them or even a comment on them. I just don't know what they did.
But in general, organizations are not awesome at self-reflection. I assume as a default prior that some of the dynamics we discussed at the beginning are rooted there, where none of the people who started those organizations are there today. What exactly are the incentives of those leaders? It's not clear to me who would have the incentive to really take stock in a fully objective and self-critical way to figure out what was done well and what was done poorly.
Dwarkesh Patel
I promise not to be too myopic about AI, but one more question. Long-term, we can forecast—maybe even medium-term we can't—but near-term, it looks like we might have things that look like AI agents and they might need to trade. What does a financial infrastructure for AI agents look like?
Patrick Collison
That's a really interesting question. Automated or autonomous transactions already exist to some extent today. Lots of services have usage-based billing. A lot of the expenses being incurred are autonomously incurred—no human is pushing a button when Stripe does most of what it does with cloud computing and incurs some cost with some cloud service. So it's happening today in an extremely primitive way.
I assume it will follow some gradient where some of those decisions are either directly or indirectly being made by an LLM or some equivalent. There'll be an almost unnoticeably smooth continuum up to very considerable degrees of autonomy. It's not that we're going to wake up some month and be like, "Oh my God, suddenly the bots have been unleashed."
There will be interesting questions around the legality of these transactions. Are these treated as the responsibility of the owner? Is there any degree of independence granted? How does liability work? Which rails are best suited? What kind of transaction velocities are we talking about? If it's a billion transactions a second, the properties of that system should look very different. Is it one giant clearing transaction every day?
If we use the analogy of usage-based services, those tend to incur liabilities in tiny increments but then settle on a monthly basis when you pay your bill. Maybe these agent transactions will have that character. There are a lot of practical applied questions, but what you're saying around these autonomous transactions conceivably being an important dimension is very true and real. It's one of the interesting ways in which the economy might change and expand over the next decade.
I think it's possible that crypto plays some role here. If we take KYC and AML very seriously for humans and we want to know the human associated with some particular financial activity, that's a murkier question in the context of an AI agent. If we look at crypto as the part of financial services that is de facto exempt from AML by design, then maybe that plays a role.
Dwarkesh Patel
How long before Stripe was founded do you think a product like Stripe could have been invented?
Patrick Collison
Depending on what exactly you define Stripe as being, conceivably decades earlier. On some level, PayPal is a kind of Stripe, and there are many payments companies before PayPal. You could go all the way back to cash registers or something. It depends on these definitional questions.
The particular secular tailwinds we benefited from around the rise of app stores, the on-demand economy, and maybe the startup boom post-YC and after the financial crisis—those particular tailwinds were idiosyncratic and specific to Stripe. The GFC was 2009 and Stripe was founded in 2010. In as much as you define those as being core, then not that much earlier. But mostly my story of Stripe is one of market inefficiency. I do wonder why much of this didn't happen sooner.
Dwarkesh Patel
I always find it really interesting when there are these cases where it wasn't even that it could have been started sooner but there was nobody in the market. There were many people in the market, and they weren't random people—there were technology companies headquartered in San Francisco who were in the market. Do you have some explanation for why it didn't occur to them?
Patrick Collison
I'm hesitant to generalize too much because I only have maybe n equals one experience. It's dangerous to over extrapolate from that. Maybe n equals two now with Arc as a very different kind of organization. Or if you include all the features of Stripe, n equals like 10 or 20 something.
My general view is that most products and most businesses can just be done much better. I think moats are typically overrated. Payments is a great example of a domain where on a logical basis, you would say there are so many sources of defensibility—the network effect of account holders, the data network effect and economy of scale for fraud, regulatory moats and barriers—and yet, not only does Stripe exist, but there's a whole fintech ecosystem today.
It gets down to deep questions of what the binding constraint is on the number of effective organizations that exist in the world. For any given sector, why is it that number of companies rather than twice that number? I think it's about motivation, ideas, people's willingness and determination to organize talent. These are more sociocultural explanations.
Hamilton Helmer is probably the leading scholar of the sources of defensibility for businesses. He has this niche book called *Seven Powers* that disaggregates all the various sources of market power. That is true and important insofar as it goes, but nonetheless, it's kind of strange to me that nobody had done Stripe before Stripe.
Dwarkesh Patel
When you think about the fact that moats are overrated and just doing the thing is underrated, does that make you think differently about Stripe's moat?
Patrick Collison
Yes. I do think that one can have organizational and cultural moats. Maybe this contradicts what I was just saying, or maybe it's consistent in the sense that it's a kind of cultural explanation. In as much as we have a moat, it's because we have a very good understanding of our domain and a set of people who actually care about solving the problems. They are continually paranoid at the prospect that we might be forgetting something important, instead of trying to figure out what the important thing that could supplant Stripe's approaches is and making sure we build those first.
You're familiar with Conquest's laws. Conquest's third law is that one should model organizations as if they're run by a cabal of their enemies. Obviously it's tongue in cheek, but it's interesting to think about the kernel of truth there. Most organizations, when they start out, are actually trying to achieve their stated goals. Somebody started the organization for a reason, probably for the stated reason. But over time, that person and that set of people depart, and new people come to take their place.
For the fifth generation, why are they there? To what degree do their particular local incentives align with the nominal originally stated goals of the organization? There can be a lot of misalignment. They're following a local path, and conceivably even the leader of the organization—not through any fault of their own per se, just that they're human and have their own incentives—the original constitutional incentives of the organization might be quite different. This phenomenon is a fact of life.
To your question, in as much as Stripe has a moat, what is it? Others can judge to what degree it's actually manifested and rooted in practice. I think it is, but I'm a biased observer. I think it would be that people at Stripe really care about solving the problems that we say we are trying to solve.
Dwarkesh Patel
The point about misalignment over generations is interesting. Do you have examples of institutions which have for decades or hundreds of years managed to keep their original mission statement and organizational competence? Tech companies, even the oldest ones, have not been around that long. The median age of the corporation is famously low. What is a good example here?
Patrick Collison
I think some of the explanations around the effects of shareholder capitalism have some credibility. It is very plausible that shareholder capitalism attenuates the duration of these organizations. I'm not saying that's definitively true, but I find it incredible the idea that it is. It's not clear to me that that's necessarily bad, even if it is true. Are we on the side of the humans, of aggregate innovation in the world, or of the corporations qua legal entities? It's not clear to me the answer should be the third.
At the same time, if you look at Europe or some other places like Denmark, for reasons related to the tax code, a lot of organizations are either controlled by or very substantially held by nonprofit foundations. Novo Nordisk, the GLP company, Maersk, the shipping company, I believe also Lego—a lot of these corporations have a lot of their stock held by foundations. That has the secondary effect in many cases where they actually embed in a legally binding constitution their mission.
I'm not an expert on Novo Nordisk, but I happened to get a book about it over Thanksgiving. It's enshrined in their constitution that they have to make insulin broadly available really cheaply, or at least cheaply in Scandinavian countries. They're allowed to charge market prices elsewhere. The rest of their profits, they're legally obligated to reinvest in R&D. Is that somehow causal in the fact that they then invented one of the most remarkable pharmacological discoveries of the last 20 years in these GLP-1 agonists? Plausibly.
I think these questions around why the median age of organizations and corporations is what it is are definitely interesting. I suspect it's a somewhat contingent aspect of how we've chosen to organize large corporations in the U.S. today.
Dwarkesh Patel
The thing you're mentioning about this firm seems very similar to export-led growth in Asia. You have tariffs. This one company, your task is making the cars, but you better make the cars good. You have no competition, but you had to invent the best car in the world.
Patrick Collison
Yes. We are all fans of Smith and Ricardo and all these characters, and even they are sort of less dogmatically attached to free trade than perhaps people today interpret them as being. People like Friedrich List and those quasi-contemporaries are maybe on a relative basis underrated. In as much as you believe the sociological, cultural, skill, even vague alignment—not in the AI sense, but in the more interpersonal sense—in as much as you think these are important and explanatory, then I think you end up thinking about some of the things you just raised.
Dwarkesh Patel
That's really interesting to hear you say that, because if you think about Stripe's mission—it's to facilitate global trade, to make sure that some firm from India can compete with any firm in Nigeria. The room for you to have this sort of learning curve where you're less efficient than the global competition should be less if Stripe exists. Isn't Stripe the anti-List company?
Patrick Collison
Well, it depends which version of List. To be clear, I'm not specifically endorsing these tariffs and trade barriers. The history associated with them is checkered at best. I think it's possible that if you have a specific sector where you have clear goals and a credible path to actually achieving some substantial degree of success, and probably some more kind of conjoined propositions, then maybe some degree of activist trade policy might be on net the beneficial thing to do. I don't think that describes most sectors in most countries at most times.
Dwarkesh Patel
That's so interesting. I think there's an interesting thread here in how it relates to Stripe Climate. You're subsidizing these learning curves that these East Asian countries did for their own internal companies. You haven't picked out a specific company that's going to necessarily be the key of carbon sequestration, but how do you think about this?
Patrick Collison
Maybe a way to unify the two points—and I'll speak about Stripe Climate in a second—is that I think it's Say's law about demand creating supply. In as much as Stripe aggregates more and more global demand, part of the—I don't know, it seems too self-aggrandizing to call it the theory of Stripe—but some vague hunch in Stripe is that that aggregation of demand can have important expansionary effects with respect to the ensuing supply. Stripe Climate is some version of this hypothesis applied on a much smaller scale than Stripe itself, but still real and we'll see, maybe important.
The basic idea, for folks who aren't familiar: We observed in 2018 that everyone seems to agree that carbon removal will be very important. Even if we decarbonize the economy on the most optimistic timeframes, there'll still be an accumulated stock of carbon that is a problem. It sounded pretty weird. There were virtually no carbon removal companies in the world in 2018. Maybe there were two or three. No companies had ever purchased from a carbon removal company. These were really science projects.
We thought, "Somebody got to start." It might be valuable to not only transfer some dollars but to confer some credibility on this sector. Not that Stripe is the world's most credible company, but it's better than nothing. We started contracting with some of these carbon removal companies. That went pretty well and they seemed appreciative.
We then in 2021 formed Frontier, which is an AMC—an advanced market commitment. Inspired by the first AMC, which was a pre-commitment to purchase vaccines for developing world countries for diseases where the profits weren't sufficient to pay for the program, we said to do this for carbon removal. We raised a billion dollars. Stripe was the first investor—we're not actually investing, we're just buying. We were joined by Shopify, Alphabet, Meta, JP Morgan, and a bunch of other companies.
Now there's a fairly active sector of carbon removal companies. Frontier has contracted with between 40 and 50 companies, the overwhelming majority of which didn't exist when we started out. We ran an anonymous survey back at the end of last year and asked them to what degree the existence of Frontier was causal in their starting the company in the first place. 74% of the companies said that Frontier played a causal role in their starting the company. These inducement effects can be pretty significant.
Dwarkesh Patel
That's huge. What are other ideas you've come across where an AMC would be an effective instrument for moving forward with the tech?
Patrick Collison
That's a good question. We've actually been having some of that discussion internally. It's not that we plan on doing it ourselves necessarily, but I'm wondering if there are people we should share our experience with and try to help along.
There's still a lot of stuff in the biomedical field. Patents are pretty useful insofar as they go, but there's a lot of innovation that seems like it would be socially beneficial that patents don't provide a way to cover the cost of. There was some excitement a few years ago about mannose, which is a sugar. There was one paper that suggested maybe tumors will selectively take up mannose rather than glucose, but they can't actually metabolize it properly and they just die. Maybe this could be an effective onco-treatment of some sort.
But mannose is a generic sugar. It's been understood for more than a century. You couldn't patent it. It's not clear who has the incentive to even fund the work to test whether this would actually work in practice. This is not an endorsement of mannose, but there are things of this shape where you can clearly see it might be very beneficial, but it's not totally clear how the economic structure of the market can make it possible.
There are still a lot of those across the biomedical landscape. There are still a lot of vaccines that could in principle exist that don't. Lyme disease—there's one vaccine that was withdrawn from the market over some safety concerns that I think were misplaced, but it's still the vaccine.
Dwarkesh Patel
It's not even that well understood. People have chronic Lyme disease. We don't know if it's legit or not.
Patrick Collison
Exactly. It's a good question. Maybe some of your listeners will have ideas for fields where we sorely need an AMC.
Dwarkesh Patel
I want to go back to Stripe for a second. You're famously appreciative of craft and beauty, but also you appreciate the power of scale and growth. Is there a type of speed? Interesting. But is there a type of craft that is just not amenable to speed, growth, and scale? If you think of a Japanese chef learning to cook rice for a decade before he can move on to sushi, is that just not competitive in the modern world?
Craft, scale, and speed. I don't know. They're strictly necessarily intention in every case, but they're definitely frequently intentioned. So just yes, I think is kind of one short answer to that.
Patrick Collison
At the same time, a lot of the most successful companies are those that I think are distinguished by the extent to which they exhibit appreciation for and skill in realizing craft and beauty. LVMH is one of the largest companies in the world and that's literally their business. I think Tesla is pretty good at this. They're good at many things, but including this. Obviously there's Apple. TSMC is a kind of chip sushi chef in Taiwan. So much tacit knowledge and difficult to transfer skills.
I think it might be the case that craft and the pursuit of it is as important as it's ever been. As Stripe has gotten larger, I think we ourselves have come to greater conviction in this. Part of what's interesting about these aesthetic qualities is they're generally speaking unquantifiable. I don't know if they're intrinsically unquantifiable—maybe you could train a model to do so—but today they are broadly unquantifiable, and yet they influence people in significant ways. People demonstrably care about aesthetics. If they're a company, they care about the aesthetic characteristics of the products they use. On an intuitive level, people know that's true.
But it's difficult to manage that at an organizational level where there isn't a P&L associated with it. If you're screwing it up, you don't see a neat time series decline. Over the 14 years of Stripe, through a kind of not exactly trial and error, but by studying cases where things worked well and cases where things worked less well, it really seems clear to us that even in a domain like ours—selling primarily to businesses—this is something truly important.
Also, getting back to what we were discussing previously about the sociology and cultural explanations of defensibility: the best people consider themselves craftspeople in their domain. They really want to work with the best other people above almost all else. So it may almost be true that even if from a customer-facing standpoint craft was not valued by the market, you might still want to build an organization that indexes heavily on this because you just want the best people for other reasons. As it happens, I think customers do value it, and the evidence is broadly consistent with that. It's very hard to assemble groups of the best people if you don't take the practice of the work super seriously.
Dwarkesh Patel
What kind of beauty or craft or simplicity is more important—interface or implementation? There's famously that essay that Unix is successful because the implementation is simple and not the interface.
Patrick Collison
I guess the interface is kind of simple, but there are a lot of asterisks and caveats and edge cases that Unix doesn't handle for you. Stripe does, right? Presumably it depends what you're building. For TikTok, it's probably more important that their interface is simple, even if their implementation is a mess. That's probably okay. Whereas for Stripe, people are on some level purchasing our architecture, purchasing their ability to do certain things because of what our architecture makes easy and possible.
If by interface you mean the visual GUI interface, then maybe we can draw some separation there. But we don't really draw that distinction. We think of the interface to Stripe as being the architecture. No one else seems to agree with me, but I often think of Stripe as similar to Mathematica, where we're selling a self-contained universe to model whatever is of interest to you. We're providing primitives and interfaces and tools to enable your modeling. Fundamentally, we're helping you do something on your own terms. In that sense, I don't think the architecture and the interface are necessarily that separable.
Dwarkesh Patel
That's a really interesting analogy. Although with Mathematica, the entry it's giving you is to the platonic objects of math. Whereas for you guys, it's like the entry is to Visa error codes. The end object is not the platonic.
Patrick Collison
That's true, though in both cases. Yes, the analogy falls down in a few respects. But the idea of a transaction is pretty fundamental and is roughly as old as the quadratic equation or something. I guess the transaction is older. Mathematica today supports all kinds of crazy arcane stuff. If you go through the more obscure packages, you can definitely find things that are much less broadly employed than even Visa error codes. So these are not the same. I just find it to be an interesting source of intuition. I think what Wolfram has done with Mathematica is pretty amazing.
Dwarkesh Patel
Another way in which Mathematica maybe differs: if they had to make a change, big deal, somebody has to learn new syntax. If you make a change, it's like billions of dollars of transactions don't happen. How does that change the way you think about the initial architecture and the stakes?
Patrick Collison
It's a good question. First, on the point about beauty with respect to architecture: I think it's interesting that API design in general doesn't get more study as a discipline and as a practice. It plays a significant role in the fate of platforms. If you get it right, there can be compounding positive benefits. It's really striking that with mobile app development—one of the most dynamic ecosystems of the past 15 years—so many of the classes in iOS development are prefixed with NS. Less so now with Swift, but for much of the iPhone's history. NS refers to Next Step from Next in the 90s. When you get API design and architecture right, it can be so enduring over multiple decades, even in the face of frenzied evolutions in everything around it.
Unix is another example. It has tons of shortcomings, but the architecture has basically worked for more than half a century. We're trying to impress upon people at Stripe the importance of multi-decadal abstractions. People sometimes respond thinking that's some insanely lofty hyperbole. But no, that's actually just what happens when you get this stuff right. If you get it right, you can reap these incredible benefits for a very long time.
To the Mathematica point, they take backwards compatibility really seriously. You can run programs written 20 years ago unchanged in today's Mathematica. That really raises the stakes in API design. We have that same problem ourselves. When we think about introducing something new, it's not just, does this address the particular needs motivating it today? But do we think we can stand behind this in 2044? How do we think the world might evolve around us such that it all remains coherent? We certainly don't always get that right, but that's what we're trying to do.
Dwarkesh Patel
Is VisaNet an example of this? One might say that one of the downsides of being able to use implementation for many decades in the future is that even if it's self-sustainable and you have this ecosystem equilibrium set around it, if you can't modify it just because of people's local incentives, you get stuck in an equilibrium that's worse than it could be otherwise.
Patrick Collison
I think the card networks generally—Visa and MasterCard—are a pretty good equilibrium. It's easy to judge today with the world as it exists in 2024, but you have to look at the world as it was when they started out and the particular problems they were solving. When you compare the financial landscape in the U.S. or the Western world to those in other places, it's certainly not clear to me that the U.S. has gotten a bad hand or is somehow stuck in any meaningful way.
The card networks do a couple of things. They were originally designed to replace store credit. It was a credit card originally, not a debit card, and that was important. The availability of structured consumer credit is a pretty big deal and pretty beneficial, especially for lower income people. With the advent of jet travel and mass market tourism, they helped supplant travelers checks and various worse alternatives like carrying cash around. With the internet, they were substantially involved in enabling online transactions. The fact that they got the architecture so right that so many different use cases were able to be addressed by their core design is really impressive.
The guy who designed all this, Dee Hock, was a remarkable person. People complain about interchange. Stripe is on the—well, it depends how you look at it. Many people consider Stripe to be on the wrong side of the interchange cost equation, in the sense that we're giving away the interchange revenue to other companies. So I don't think I'm structurally biased in favor of interchange. Yet I will say it's pretty interesting what interchange made possible. It's a distribution incentive fee, paying other entities to recruit customers, convince them to get a card, maintain the card, pay it off at the end of the month. You're paying for pure distribution. There's a person at the end of the flight telling you to sign up for the United Credit Card. That's what interchange is paying for. That guy annoys me.
There's paying for the actual credit issuance itself, the customer support, the dispute handling. It is interesting to look at cases where the card networks didn't arise. Germany is a classic one. Dealing with the online economy in Germany compared to the U.S. is so much worse. If Stripe could push a button and have really broadly adopted cards in Germany, we would push the hell out of that button.
You can look at China, which has Alipay and WeChat payments that are really ubiquitous. They're very digitally enabled from a transactional standpoint. But those products don't tend to be as sophisticated with consumer credit. The transaction fees for transferring money you already have are super cheap. But you need to look at it on a fully loaded basis: what about the cost of actually getting the credit to make the purchase in the first place, as a credit card would enable? As you look at these counterfactuals in other places, one feels a kind of gratitude for what Dee Hock and Visa and MasterCard made possible. I'm not saying they're perfect. I'm most interested in critiques from people who've really studied the ecosystems of other countries, because it's easy to underestimate what we got in their invention.
Dwarkesh Patel
Maybe there's a sort of test and expense thing going on here. If you had to design payments from first principles now, does it make sense that all these things—taking on credit risk, the chance of fraud, dispute adjudication—should cost like 2% or 3% of each transaction? What would payments look like if you had to design that from first principles?
Patrick Collison
We're seeing a live version of this experiment play out for the first time in many years in a number of countries today. Central banks are becoming more active in designing national payment schemes. PIX in Brazil launched in late 2020. UPI in India was the instigator here—the central bank payment system tied up with Aadhaar and their national identity system. That inspired a lot of central bankers in other countries to build their own UPIs. PIX launched in 2020 and now a significant majority of all Brazilian adults are weekly active users. It had this incredibly rapid adoption curve. You have Swish in Sweden. Across East Asia, Japan, Thailand, Switzerland—central bank after central bank is deciding they should have their version of this.
This is a kind of reinvention of the payment system from scratch. For hard to understand reasons, things typically seem to asymptote around 2 or 3% once you layer in customer support, consumer protection, fraud prevention, anti-money laundering controls, and credit. It's also important to note that a lot of that 2% or 3%, beyond covering costs, ends up getting remitted to consumers in the form of rewards—not in every country, but in many. If you look at public reports from various banks in the U.S., their interchange revenue—a lot of it is going straight back out the door to the consumers themselves. So it's not clear how exactly one should think about the economics. If it's going back to the consumer, should you include that as a transaction tax, or is it just a weird circular relationship?
I've not seen any evidence to suggest that the 2% or thereabouts is massively inefficient in the scheme of things. I'm not saying it's the optimal level. Maybe 1% would be better. But within some range of 1% to 3%, it's probably reasonable.
The place where there's even more change at the moment that we find ourselves thinking more about is the changing structure of global tax. There's been a reasonable amount of innovation in the tax domain over the last century—income taxes got pretty high, then value-added taxes. The new thing in the online context is jurisdictions imposing sales taxes on businesses that don't have any kind of locus in the jurisdiction in question. So you're a podcaster in the Bay Area, and the Dwarkesh merch store will have to pay the town of Uppsala in Sweden a special tax on baseball caps. You will need to know about that particular tax, collect that amount from the buyer, report to Uppsala, and then figure out how to get that money to Uppsala. It's a combinatoric problem of buyer jurisdictions, product types, and all the different jurisdictions you have to remit to. Those amounts are not three basis points. The taxes in question are often 5% or 10%. So as I think about the funds flows on the internet and how that's evolving, changes in tax law are actually a much bigger deal than anything about the transactional economics.
Dwarkesh Patel
By the way, it's not the Dwarkesh podcast. It's Lunar Society Podcast LLC, registered on Stripe Atlas. So any merchandise I sell in the future, Stripe will take care of that.
Patrick Collison
Well, if there's ever any Stripe complaints, we will write.
Dwarkesh Patel
No, it's great. It's been super useful. It would have been much more difficult to get business operations going. Do you think—I know you're supposed to be interviewing me—but did Stripe play any counterfactual role in you charging for anything? This is the thing we're always interested in. When we talk about growing the GDP of the internet, it's not getting the existing GDP onto our rails. It's where on the margin can we cause economic activity that isn't already occurring. You started the podcast before incorporating, but were we causal in any fashion in the merch or anything of that nature?
To the extent that Substack would not be a convenient place to get payments from to begin with, that's definitely everything. Also, I wouldn't charge for the newsletter if Substack hadn't made it super easy. If I do an ad or something, I wouldn't even know how to begin getting the money if I didn't already have an LLC through Stripe with an associated bank account. So yeah, probably kind of factorially responsible for a lot of the monetization.
Patrick Collison
That's cool. Appreciate it.
Dwarkesh Patel
So what are some unexpected compliments to payment processing you see in the future? Atlas, identity fraud detection—in retrospect, it might not have been obvious back then. Now it does seem that way. What will be like this in five or ten years?
Patrick Collison
Honestly, our problem ends up being that too many things—more things than we can possibly pursue—look like compliments. Every business, almost by definition, has revenue. We obviously want to help them generate, accept, manage, and orchestrate everything pertaining to that revenue. Once you're in that flow and you go through the steps of running a business, a lot else looks relevant and connects quite directly.
When Stripe started out, Stripe seemed like it definitely wasn't cool. It was sort of the opposite for just a couple of us. We thought we could make this the superior payments API. For the vast majority of its history, Stripe has attracted people who are drawn to unglamorous infrastructure challenges. We're not a company that specializes in making beautiful cars. We make roads.
I bring that up because it's relevant to this compliment question. In our discussions internally, the significant majority is still about where there are actual practical shortcomings and limitations in even our core bread and butter. Payment processing might be a slightly too limited term. Maybe it's more about global programmable money orchestration. That includes consumer-to-business payments, business-to-business payments, payments with credit or lending involved, how you hold money, how you convert money between different currencies, how you represent money held by different legal entities, and how we make it possible for individuals or small businesses to act as micro multinationals.
Those problems are not small. If we could just solve those really effectively, Stripe will be a very consequential organization. The counterfactual importance of building some of this stuff as we go to newer markets that are more poorly served is actually increasing rather than shrinking. In the U.S., there were payments companies before Stripe. Maybe if Stripe had never done its thing, you'd eventually have found some way to monetize a newsletter. But if you're in Albania, the set of options available to you is far more restricted. The marginal impact as we expand globally increases quite a bit. Even though we are interested in and do pursue some of these direct adjacencies, the core problem of global money orchestration remains a really big and unsolved problem.
Dwarkesh Patel
Does that look like being a better interface for all these complexities, glossing them over under seven lines of code? Or does that look like actually replacing the rails and the infrastructure to make all this more efficient?
Patrick Collison
The former. It's just not that useful to build financial ecosystems that are self-contained. A financial island is not that helpful. It's much more valuable to build a financial air network or something. We would much prefer that Stripe plugged into every existing system, rail, and domestic organization rather than trying to come along and supplant them. This has been Stripe's strategy very deliberately from the beginning. There were lots of companies when Stripe started out that were trying to do their own thing and go their own way. Our belief was—it's classic Metcalfe's law stuff—by enhancing the capabilities of an existing ecosystem, you create quite a bit more value.
Dwarkesh Patel
Let's go back to Stripe. Is Stripe a writing culture for the benefit of the writer or the reader?
Patrick Collison
It can be both. But which one more so? I think there are actually really considerable benefits on both sides. For the reader, it's not just that it's maybe more efficient to communicate stuff through text, though in many cases it is. There's also an intertemporal benefit where future readers can try to understand the through line and the thought process that led us to this point. That's very considerable.
But it's also true that I write things, and lots of people write things, in order to organize one's own thoughts. If that ability was taken away from me, I think I'd be meaningfully less effective. How exactly those bounce out is hard to say. They're not actually separable. That's my answer. Literate cultures are just a different thing. I don't mean literate in some faux intellectual way. Maybe textual cultures is a better term.
Bruno Latour spoke about how he thinks part of how the printing revolution—Gutenberg's—caused the scientific revolution was by making knowledge more rigid. Before, if some observation didn't match some claim, you could always struggle and say, "I guess the person who transcribed that thing made a mistake." By making things more rigid, it's easier to break them. Then you can notice discrepancies between the theory or the claim and the actual reality.
There's some version of that organizationally. I'm not drawing that precise parallel, but there are analogous dynamics where the nature of oral cultures and textual cultures are just quite different. The kinds of collaboration that are possible and the kinds of consistency that can be achieved are fundamentally different. Is the front or rear wheel of the bicycle more valuable? Theoretically you can have a unicycle, but as a practical matter, you need both.
Dwarkesh Patel
I know I said no more questions, but on this particular point, it actually seems very legitimate to me that you might expect firms that have a lot of writing to be the first to experience the productivity gains of AI. Because there's all this context that the model doesn't have available readily. I don't know if that's something you anticipate.
Patrick Collison
I think that's probably true. If the model is really good, maybe you should be able to pick stuff up quickly. But I think most organizations are not recording all of their meetings for a variety of reasons. If they're not, then there is this question of what is the corpus? How do you get up to speed?
So yeah, my guess is that'll be true.
Dwarkesh Patel
Tell me about the internal LLM you built.
Patrick Collison
Oh, we didn't build an internal LLM. We built an internal LLM tool for making it very easy for people to integrate LLMs into production services, but also into their regular workflows as humans. So the ability to work directly with the LLM as a standard chat agent, as lots of people have built, but then also to integrate that with some of our tools for querying and accessing data, or maybe most interestingly, with sharing prompts across different people.
One of my favorite examples is somebody put together a prompt for optimizing SQL queries. It doesn't always work, but sometimes it does. It's very cheap to ask, "Got any ideas for optimizing this SQL query?" And sometimes it'll come up with some good stuff. The collaborative abilities there have proven surprisingly high return.
Having a central bus through which to route all access to these LLMs, so we can experiment with different models and have some degree of observability into performance trends and usage of different cases, has been valuable. Now, with the proliferation of LLMs themselves, this is proving quite valuable because we're able to figure out for different use cases which models—self-hosted models, who knows—are most effective. I don't know what the total number of invocations is, but I think we're making millions of invocations per day now. There are dozens of actual production use cases across Stripe.
The financial services ecosystem is in some way a giant analog-to-digital exercise. Humans are analog, intentions and identities have some degree of uncertainty and noise, but transactions are digital. We often find in these analog-to-digital conversions that LLMs can be a surprisingly interesting augmenting tool.
Dwarkesh Patel
On that point about the flexibility and the edge cases in the way humans interact with these systems—in some sense, Stripe is like a really high stakes bug bounty program. If somebody hacks it, money's in play. But if there's a reliability issue, not just because of a hack but because you deployed the wrong way, a significant percentage of world GDP would grind to a halt. How do you deal with that kind of responsibility? How do you keep the uptime and reliability while deploying fast?
Patrick Collison
This is one of the things we've spent the most time on. Back to this point about wanting to be the place with the best people and the value of focusing on craft. In the context of software development, developers hate two things: slow development cycles and being paged at 2 a.m.
Given the criticality of the businesses we serve—in rough terms, 1% of the global economy. It's not totally clear how to measure this because GDP is defined as final goods, and Stripe is not only selling final goods. There could be a bit of double counting, but Stripe is mostly selling final goods. We're not used for giant supply chain shipments. I think maybe there's a mismeasurement of 10% or 20%. Long story short, it works out to about 1% of global GDP, about a trillion dollars a year.
That makes us really terrified of outages. We work so hard to enable fast iteration and development cycles without having outages. To put some numbers on it, we deploy production services that are in the core charge flow around a thousand times a day. Most of these services are automatically deployed. When anybody makes any production-ready change, it goes into production, meticulously and carefully orchestrated so that it first runs on a small sliver of traffic, then incrementally more until it's everything.
About 1,000 deploys per day at roughly five and a half nines—99.995% reliability—which works out to about two and a half minutes of unavailability per year. It's not contiguous, but that's what it approximates, happening as background radiation throughout the year. Getting to that point takes a huge amount of investment. Then there are security properties that are less readily measured but analogous.
Silicon Valley doesn't tend to place a lot of value on process and operational excellence. We culturally value the spontaneous, the creative, the iconoclastic, the pathbreaking. But building mechanisms that enable reliable provision of important services at scale and removing sources of variability that can cause a bad day for a very large number of people—I don't think they get as much cultural credit.
We've found that there's a core feedback loop: defining what we care about, building automated measuring systems to measure it in practice, figuring out where we're not living up to that and why, then intervening to improve the system. Importantly, we build secondary controls that detect instances of deviation long before they cause a production problem, instrumenting the system in some upstream way.
Most of what I said was well understood by production engineers in 1930. I'm not claiming it's any radical breakthrough, but we've found that the adoption of these practices in a tenacious, multi-year form yields really high returns. There may be other organizations that both ship at that rate and maintain that developer velocity at this combination of scale, reliability, and security, but I don't think there are that many. It's a real testament to the remarkable folks at Stripe who made it happen.
Dwarkesh Patel
Last point on this—the fact that you have this huge internal tooling and testing means once you get the AI engineers, they can just push the commits and you have the infrastructure set up for it to be readily evaluated.
Patrick Collison
Across the board, so much comes back to what has to be true for us to actually build and take seriously this goal of building the best software. It's easy to say that as a lofty, vague, hand-wavy aspirational statement. But if you take that seriously as a goal and think about what you would have to measure if you were going to pursue it in earnest, and what the characteristics of organizations that produce it are—you get down to customers having to really like your stuff. So how can we measure that? How can we systematize the process of making sure there aren't regressions?
We have this concept of experience journeys—pathways through Stripe that we really care are always implemented at a high quality level. It has to be true that developers can iterate very quickly. A theme through everything we've talked about is actually taking the goal seriously. A lot of what we do at Stripe is not genius—it's the very earnest, repeated, serious, and long-term application of taking the goal seriously.
Dwarkesh Patel
A few more Stripe questions. One percent of global GDP is such a staggering number. When you think about where further growth for Stripe comes from, does it come from the internet economy expanding, or from Stripe becoming a larger share of the internet economy? To the extent that Stripe is growing faster than the internet, where is that alpha coming from?
Patrick Collison
That's a good question. The customers that Stripe serves are outgrowing the internet economy as a whole in aggregate. At some point, those have to converge for obvious mathematical reasons, but we're 14 years in and they haven't converged yet. I think there's a lot of headroom there.
Stripe is handling around a trillion dollars a year. When Stripe started out, the global economy was 60 to 70 trillion-ish. The global economy is now around 100 trillion. We still have quite a bit of headroom before the amount of activity coming through Stripe butts up against the ceiling of global economic growth. Of course, there's no ceiling on global economic growth—it could be vastly higher than it is. I don't even mean new technologies or AIs, but just basic per capita math: what if everybody had an income on par with the US?
One of the reasons I am so interested in working on Stripe is the old Lucas line about how when you start thinking about differential rates of development in countries, it's hard to think about anything else. Why does Brazil have the particular income and GDP level that it does? Why does Poland? Why did Ireland go from being the sick man of Europe to one of the wealthiest countries? Stripe is some applied version of this question in practice. You're building software products, but in some sense touching on these questions of why aren't there more companies, and what determines the growth rate of a company?
When you start a merch store, why is it at X level of buyers rather than 2X? I think those questions remain fruitful. We actually haven't optimized the meta system of business to any particularly great extent. For the vast majority of businesses, they've been offline, inefficient, analog. It's really only over the last one to two decades that a significant share has been meaningfully digitized. The prospects for efficiency gains and optimizations are still pretty significantly underexplored.
We find incredibly basic things like extending capital to businesses. The reason we do that is not to generate profit from the loans, but because we find that the businesses we extend capital to just grow faster on a persistent, subsequent basis. Or trying to figure out how a business decides which countries it sells in. You'll find for the smallest businesses through to some of the largest that these are very ad hoc questions. "Why don't you sell in Mexico and Brazil?" "Well, it seemed kind of complicated, so we didn't quite get around to it."
To your question about where the growth comes from, I think there's still an awful lot of low-hanging fruit in just asking some of these incredibly basic questions.
Dwarkesh Patel
When we think about how Stripe will continue to grow, it'll obviously involve a lot of big businesses. You're now processing a significant amount of Amazon volume. First, how does it make sense how an exponentially growing startup contributes to exponentially growing growth for Stripe? How does Stripe keep growing on the same trajectory when it's existing big businesses you're partnering with? Second, the case for why these startups matter is so compelling—a new thing is coming into the world and we should support it. Why is it compelling that Amazon can fulfill orders more efficiently?
Patrick Collison
Those are very good questions. On the first one, you're right. Stripe is doomed to eventually grow at the rate of the economy. It's just a question of how long it takes to get there. The good news is it can be a very long time because there is so much low-hanging fruit around different optimizations and improvements. It would be many decades before that happens, but it will eventually occur.
On the second question about supporting established businesses—we will misunderstand where a small business typically has its innovation. In the cases where we denote them startups, there's usually an embedded innovation. The innovation is all that the company is. They have a new idea and they're going to do something better or different. Generally speaking, we like innovation, so we have positive sentiments towards that startup.
But there's a lot of innovation that comes from large established businesses. That's not all they do—there's also just running the existing thing. It might be a smaller share, but the aggregate fraction of innovation that comes from established businesses is really large. We have to be cognizant of the cognitive bias of startups being somewhat more conspicuous. On a relative basis, the improvements in turbine technology, fab technology, or insulation technology that come from established businesses—a significant fraction of the important inventions over the last 10 or 20 years will have come from incumbents.
Tyler wrote a book on this—big business is underrated. If you look at survey data, people tend to have very positive sentiments towards startups and small business as a class, even though they have negative sentiments or relatively negative sentiments towards big business. Established businesses tend to pay better, they tend to be more efficient, more of the innovation in our economy comes from them, and they produce a lot of consumer surplus.
The specific case for Stripe working with them is typically they're coming to us not because they want to take what they're already doing and transpose it, but because either they want to do a new thing they're not doing today—associated with some new business line or innovation—or they've spotted the opportunity to meaningfully change how they provide an existing one in a fashion that yields consumer surplus.
That sounds very abstract, but in practice it means they want to take this thing they're selling in this market and sell it in many more markets. Or they've realized they should sell it in more convenient ways, like on mobile. Each of those, if successful, gives us a decentralized signal from the economy that something of value is being provided that wasn't there before.
As I take stock of the enterprises migrating to Stripe—whether large retailers, global manufacturing firms, or shipping companies—it typically has one of those two patterns: new product, or current product sold to people who weren't buying it before.
Dwarkesh Patel
If you think about the big trends in society that solve our big problems—Moore's law, the cost of solar—these are marginal improvements over many decades that big companies are able to invest a lot of money into R&D for. The rate of improvement is underrated.
Can I ask about John for a second?
Patrick Collison
Sure.
Dwarkesh Patel
You guys recently published Poor Charlie's Almanack, and subsequently Charlie Munger passed away. Did Munger ever comment on your relationship, or whether it reminded him of his and Buffett's?
Patrick Collison
Not to me, but he knew John better, so it's possible that he did to John. I don't know.
Dwarkesh Patel
What have you learned about marriage from John? This sort of co-equal, intense, lengthy partnership is the closest thing you have to marriage.
Patrick Collison
I'm relatively new to the practice of marriage, so maybe in a decade I'll be able to extract the generalizable commonalities. The general thing I would say is that working with people you're close to is underrated.
I'm doing Arc with Patrick Hsu and Silvana. Fast Grants was with Tyler and Silvana. Stripe is obviously with John. John was also instrumentally involved in Arc's formation—it would not have happened without him. For all the ventures of any significance in my life, they've not only been with others, but with people I'm very close to and where I would like to have an enduring relationship that outlives them.
Sometimes one hears the advice that you shouldn't work with friends or your partner. All these things are idiosyncratic and there are instances of every possible permutation, but for me it's been a really rewarding experience. John and I will probably work together for decades. It's been an important source of meaning and fulfillment, and I think there's a real complementarity. Stripe would be a less effective company without either of us. I don't know if you mean from a bandwidth standpoint, but we both bring different things to bear.
Dwarkesh Patel
Patrick, I think that's a great place to leave it. Thank you so much for coming on the podcast.
Patrick Collison
Thank you.
Outro music begins
Dwarkesh Patel
Hey everybody, I hope you enjoyed that episode. As always, the most helpful thing you can do is to share the podcast. Send it to people you think might enjoy it. Put it in Twitter, your group chats, etc. It just splits the world. Appreciate you listening. I'll see you next time. Cheers.