Patrick Collison — Why Silicon Valley's most talented should leave
What this covers
Patrick Collison, Stripe's co-founder, sits down with Dwarkesh Patel to argue that progress across science, institutions, and business is bottlenecked less by funding or raw talent than by the structures and cultures that organize human effort. Rather than asking what prevents more money from producing more output, Collison traces the constraints to organizational design, problem selection, craft standards, and the durability of foundational systems. The conversation moves across payment systems, biotechnology, grant-making reform, and corporate governance, asking in each case: what infrastructure allows talented people to do their best work over decades?
The discussion extends through several distinct domains. In science, Collison challenges the assumption that moneyism — treating funding as a proxy for output — correctly explains research productivity, noting instead that CRISPR funding rejections and Advance Market Commitments suggest the binding constraint is often institutional rigidity, not dollars. On payment systems, he traces how interchange emerged as a distribution mechanism and examines why correct foundational architecture — such as the systems Dee Hock designed — produces compounding value across decades and use cases. The conversation also covers biomedical homogeneity and alternative funding models, the role of Say's Law in creating markets where none existed, and why companies like Apple and LVMH succeed partly through uncompromising craft. A recurring theme: Conquest's Laws — the idea that organizational goals drift as founders depart — partly explains why even well-intentioned institutions misalign with their stated missions over time.
Collison argues that progress is constrained more by cultural, organizational, and institutional structures than by money or raw talent, and that building durable systems — whether scientific institutions, payment APIs, or companies — requires taking craft, alignment, and the goal itself seriously over decades.
- Mapping inputs (dollars) to outputs is the wrong frame for science; the binding constraints are usually organizational and cultural, not financial
- Moats are overrated and 'just doing the thing' well is underrated; durable advantage comes from caring about the problem and craft
- Long-lived abstractions (APIs, institutions) compound value for decades when designed correctly
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The observed exponentially diminishing returns in science could arise simply from adding ever more unproductive capacity rather than from genuinely harder-to-find ideas; this alternative explanation weakens the case that low-hanging fruit is exhausted and gives weight to the possibility that the slowdown is fundamentally structural, cultural, or organizational.
“the fact that you could get these exponentially diminishing returns through the addition of ever more nonproductive capacity makes me not persuaded that the low-hanging case is necessarily true”
Frontier, a $1B advance market commitment modeled on the vaccine AMC, contracted with 40-50 carbon removal companies, most of which didn't exist when it started; an anonymous survey found 74% said Frontier played a causal role in their deciding to start the company, demonstrating that demand-side inducement effects can be significant.
“I think it was 74% of the companies that said that Frontier played a causal role in their starting the company in the first place. So these inducement effects can be pretty significant.”
Writing-oriented ('textual') organizational cultures are fundamentally different from oral ones — benefiting both readers (efficient, intertemporal access to the through-line of reasoning) and writers (organizing one's own thoughts) — analogous to how Latour argued the printing revolution helped cause the scientific revolution by making knowledge rigid enough that discrepancies between theory and reality became visible.
“the nature of oral cultures and textual cultures are just quite different. The kinds of collaboration that are possible, kinds of consistency, that can be achieved, are just fundamentally different.”
One of the most important and subtle skills in science is problem selection — choosing something important and hard enough to matter if solved, but not so complex that progress is unachievable — and this judgment, along with knowledge of true high standards, is what mentees learn from great mentors rather than from explicit instruction.
“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.”
Conquest's third law — model organizations as if run by a cabal of their enemies — captures a real dynamic: organizations start out genuinely pursuing their stated goals, but as founders depart and successive generations replace them, the new people's local incentives increasingly misalign with the original constitutional goals, even absent any individual fault.
“There's Conquest's third law, which is that: 'One should model organizations as if they're run by a cabal of their enemies.'”
Beyond direct therapeutic use, CRISPR can serve as a new kind of telescope: systematically perturbing each of the ~20,000 genes one by one in cell cultures under stressors lets researchers map gene effects and generate synthetic data, which is especially valuable for 'complex' diseases (non-infectious, non-monogenic — most cancers, autoimmune, cardiovascular, neurodegenerative) where understanding the genetic pathway could reveal how the disease arises and enable conventional treatments.
“with CRISPR, you can systematically go and perturb each gene one by one, mashing all the keys in sequence, and try to figure out what the effects of perturbing this versus that are.”
API design is an understudied discipline that can be decisive for platforms: getting it right produces compounding benefits and enduring abstractions that survive for decades amid frenzied surrounding change — as evidenced by iOS classes still prefixed 'NS' from NeXT in the 1990s and Unix's architecture working for over half a century — which is why Stripe designs for whether it can still stand behind something in 2044.
“When you get API design and architecture right, it can be so enduring over literally multiple decades, even in the face of what are otherwise frenzied evolutions in everything around it.”
Gerty and Carl Cori ran a lab at Washington University St. Louis that produced six Nobel laureate students despite not being the most prestigious lab able to cherry-pick top talent, suggesting their success rested on replicable organizational structures and cultural practices rather than on access to rare 'seven-sigma' genius — which matters because such practices are, in principle, replicable.
“six of their students, if I recall correctly, went on to win Nobel prizes... it's not like they got to cherry-pick every year the single most promising person, so something was going on there.”
Advance market commitments could address cases where socially beneficial innovation lacks a patent-based incentive — like testing whether mannose (an unpatentable generic sugar) could selectively starve tumors, or developing absent vaccines like Lyme disease — because no actor has the incentive to fund the work despite potential large benefit.
“there's a lot of innovation that seems socially beneficial, that patents don't provide a way to cover the cost of.”
Once you layer in customer support, consumer protection, fraud prevention, anti-money laundering controls, and credit, payment system costs seem to asymptote around 2-3% for hard-to-understand reasons, and much of this is remitted back to consumers as rewards, so it is unclear whether interchange should be counted as a transaction tax; evidence does not suggest the ~2% level is massively inefficient.
“once you layer in the customer support, consumer protection, fraud prevention, anti-money laundering controls and the credit, things seem to asymptote at around 2% or 3%.”
Shareholder capitalism may attenuate the lifespan of corporations, while in places like Denmark many firms (Novo Nordisk, Maersk, Lego) are controlled by non-profit foundations that legally embed mission into a binding constitution — e.g. Novo Nordisk being obligated to keep insulin cheap in Scandinavia and reinvest profits in R&D — which may plausibly be causal in remarkable discoveries like GLP-1 agonists.
“because of the tax code there, a lot of organizations are either controlled or substantially held by non-profit foundations.”
Big business is underrated: while people view startups and small businesses favorably and big business relatively negatively, established businesses tend to pay better, be more efficient, produce more consumer surplus, and account for a large aggregate share of innovation (turbines, fabs, insulation, relentless iterative improvement like Moore's Law and solar costs) — the startup's conspicuousness creates a cognitive bias understating incumbents' contribution.
“As a general class, andTylerwrote a book on this, big business is underrated.”
Treating funding as a proxy for output ('moneyism') falsely assumes a constant elasticity between dollars invested and outcomes produced; in reality the conversion rate is not a cosmological constant, and the binding constraint may be permits, labor shortages, or other non-financial factors, making spending a poor measure of R&D or scientific output.
“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.”
Biomedical science suffers from excessive homogeneity in its model — basic research done in academic labs by PIs applying for NIH grants reviewed by rigid study sections — and homogeneity is bad in any ecosystem, especially one seeking tail outcomes, so alternative complementary models like Arc should exist.
“Homogeneity is bad in basically any ecosystem, especially ecosystems where you're producing or seeking tail outcomes.”
The binding constraint on harmful biotechnology use is not frontier biological capabilities — highly capable malicious actors could already cause tremendous harm by applying known techniques — but rather accessibility; the real emerging risk is technologies (like a sufficiently capable LLM) that make existing dangerous knowledge accessible to anyone, not advances at the frontier of biology research.
“I don't think that the binding constraint on harmful use of biotechnology or bioweapons today is pure biological capabilities.”
For most products and businesses, things can be done much better and moats are typically overrated; payments looks like it should be highly defensible (network effects, fraud data scale, regulatory barriers) yet Stripe and a whole fintech ecosystem emerged, so the real binding constraint on how many effective organizations exist is sociocultural — motivation, ideas, and willingness to organize talent.
“My general view is: 'For most products and most businesses, things can just be done much better'.Moatsare typically overrated.”
Contrary to assuming craft is incompatible with scale and speed, many of the most successful large companies (LVMH, Tesla, Apple, TSMC) are distinguished by their appreciation for and skill in realizing craft and beauty, and craft may be as important as it's ever been — partly because aesthetics demonstrably influence customers and partly because the best people consider themselves craftspeople and want to work only with other top people.
“a lot of the most successful companies are those that are distinguished by the extent to which they exhibit appreciation for and skill in realizing craft and beauty.”
Interchange is best understood as a distribution incentive fee — paying other entities for recruiting customers, getting them to acquire and maintain cards and pay them off — plus the cost of actual credit issuance, customer support, and dispute handling; comparing to systems like Germany (no broad cards) or China (cheap transfers but weak consumer credit) reveals what the card networks made possible.
“It is a distribution incentive fee, where you're paying other entities for recruiting customers, convincing them to get a card”
Visa and MasterCard, designed by Dee Hock, got their architecture so right that one core design successfully addressed successive use cases over decades — originally replacing store credit, then supplanting travelers checks with jet-age tourism, then enabling internet transactions — demonstrating the enduring value of correct foundational architecture even when later criticisms (like interchange) are valid.
“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.”
Organizations are generally not good at objective self-reflection, partly because the founders are long gone and the current leaders' incentives are not clearly aligned with honestly assessing what was done well or poorly — so it is unclear who would even have the incentive to take stock critically after an event like COVID.
“In general, organizations are not awesome at self-reflection.”
Major breakthroughs have repeatedly been rejected by conventional grant review: Jennifer Doudna's CRISPR work was funded by DARPA after NIH rejections, and Katalin Kariko's mRNA vaccine applications were famously rejected — suggesting speculative, important work is systematically disadvantaged in the standard funding environment.
“Jennifer Doudna's work was, if I recall correctly, funded byDARPA, because herCRISPRNIH applications were rejected.Katalin Kariko's NIH applications formRNA vaccinework were famously rejected.”
A survey of Fast Grants recipients found that 79% (four out of five) said their research agenda would change 'a lot' if they had flexible funding they could direct however they wanted, suggesting the binding constraint on scientific output may be the rigidity of how grantees can choose their agenda rather than the funding level itself.
“Four out of five(79%) said that their research agenda would change a lot if this constraint was removed.”
A core hunch behind Stripe and Stripe Climate is a version of Say's Law — demand creates supply — where aggregating global demand can have expansionary effects on the ensuing supply, and an advance market commitment can confer credibility and create a market where none existed.
“isSay's lawabout demand creating supply. In as much as Stripe aggregates more and more global demand... this aggregation of demand can have important expansionary effects with respect to the ensuing supply.”
While Smith and Ricardo themselves were less dogmatically attached to free trade than modern interpreters assume and thinkers like Friedrich List are relatively underrated, activist trade policy (tariffs, protected single-firm sectors forcing quality like East Asian export-led growth) only plausibly benefits in narrow cases with clear goals and a credible path to success — not most sectors in most countries at most times.
“People likeFriedrich Listand other, not quite contemporaries, but quasi-contemporaries are underrated on a relative basis.”
Reflexive contrarianism for its own sake is not genuine independent thinking — if you simply oppose the prevailing mood, you are still following the herd, just with a 'sign bit' inversion; truly resisting the herd is extremely difficult in practice.
“If you're just contrarian to the prevailing mood, then you're following the prevailing mood but with a sign bit inversion. I don't endorse that either.”
While teenagers may benefit from going to San Francisco, people in their 20s pursuing deep technical expertise should reconsider, because San Francisco culturally valorizes iconoclastic founding over the multi-decade accumulation of deep technical knowledge that many of the most valuable inventions require.
“I said that people in their teens should go to San Francisco. I wonder if people in their 20s shouldn't go to San Francisco.”
Stripe's goal of 'growing the GDP of the internet' is not merely migrating existing economic activity onto its rails but causing new economic activity that wouldn't otherwise occur — making it possible for people who weren't monetizing, or businesses in underserved markets like Albania, to transact, where the counterfactual impact is far larger than in well-served markets like the US.
“it's not like: 'Get the existing GDP onto our rails', — it's sort of: 'Where on the margin can we cause there to be economic activity that isn't already occurring?'”
Stripe extends capital to businesses not to profit from loans but because the businesses it extends capital to then grow faster on a persistent subsequent basis, illustrating how optimizing the under-explored 'meta system of business' yields large gains.
“the businesses, whom we extend the capital for, then just grow faster on a persistent subsequent basis.”
Arc Institute differs from the standard model in three ways: it funds scientists themselves to pursue curiosity-driven research rather than funding specific projects; it builds shared in-house infrastructure so labs aren't circumscribed by their own scale; and it provides career paths for people who love research but don't want to become principal investigators.
“scientists are funded themselves to pursue whatever they want. So it's curiosity-driven research, whereas NIH grants are given for projects.”
The 'adaptability premium' — the value of being able to synthesize information and respond competently to crises rather than predict them in advance — is likely to rise substantially over the next decade, a lesson reinforced by COVID and Fast Grants where competent individuals and new institutions mattered more than advance prediction.
“The adaptability premium is probably going to go way up over the next decade.”
Autonomous transactions already exist (usage-based billing where no human pushes a button) and will increase along a smooth continuum as LLMs take over more decisions — not as a sudden 'bots unleashed' moment — raising practical questions about bot liability, ownership responsibility, transaction velocity, and which rails suit which patterns; crypto may play a role as the part of financial services de facto exempt from AML by design.
“It's not that we're going to wake up some month and be like: 'Oh my god, suddenly the bots have been unleashed.'”
Despite many capable San Francisco technology companies already operating in payments, Stripe's emergence is fundamentally a story of market inefficiency — it could plausibly have been built earlier, and the puzzle of why no one did it before points to deeper questions about what really constrains the number of effective organizations in a sector.
“Mostly my story of Stripe is one of market inefficiency. I do wonder why much of this didn't happen sooner.”
The biggest emerging change for internet funds flows is not transaction fees but tax law: jurisdictions increasingly impose sales taxes on businesses with no physical presence there, creating a combinatorial problem of buyer jurisdictions and product types where rates of 5-10% dwarf the basis-point scale of transaction fees.
“I think changes in tax law are actually a much bigger deal than anything about the transactional economics.”
There were roughly 1% as many practicing professional scientists in the US pre-World War II as post-1950, with papers and patents following similar ratios, yet a great deal of high-quality science emerged in the first half of the century — so despite spending increasing by two-plus orders of magnitude, there is no clear linear relationship between scientific spending and output.
“there were around1% as many practicing professional scientists in the US pre-World War II as there were post-World War II”
The constancy of US GDP growth is one of the strangest economic phenomena, hard to explain — dismissing it as overdetermined fails because other countries manifestly don't exhibit it, raising the unanswered question of what is weird and special about the US.
“The constancy of US GDP growth is one of the weirdest things. I don't know if I've got an explanation for it.”
Stripe processes roughly 1% of global GDP (about a trillion dollars a year) while deploying core charge-flow services around 1,000 times a day at roughly five-and-a-half-nines (99.9995%) reliability — about two to two-and-a-half minutes of unavailability per year — achieved through meticulous incremental traffic rollout, automated measurement of what they care about, and upstream secondary controls that detect deviations before they cause production problems.
“we deploy production services that are in the core charge flow around 1,000 times a day... about 1,000 deploys per day at roughly or somewhat in excess of five-and-a-half-nines — 99.9995% reliability”
Much of the tech industry doesn't place much value on process and operational excellence — culturally prizing the spontaneous, creative, iconoclastic, and path-breaking — so building mechanisms for reliable provision of important services at scale and removing sources of variability doesn't get as much cultural credit, even though tenacious multi-year adoption of well-understood production practices yields very high returns.
“a lot of the tech industry doesn't place a lot of value on process and operational excellence. We culturally value the spontaneous, the creative, the iconoclastic, the path-breaking.”
The right way to run a major crisis initiative is to find and hire the best domain expert rather than a generalist administrator, as Operation Warp Speed did with Moncef Slaoui to remarkable success — the recommendation being 'figure out who the Moncef is and go hire Moncef.'
“They chose a super effective domain expert,Moncef Slaoui, to run that and it was monstrously successful, truly remarkable.”
To the extent Stripe has a moat, it is organizational and cultural: a deep understanding of the domain and people who genuinely care about solving the problems they claim to solve, who are continually paranoid about what could supplant Stripe's approach and try to build that first.
“In as much as we have a moat, it's because we have a very good understanding of our domain. We have a set of people who actually care about solving the problems”
Stripe's deliberate strategy from the beginning has been to plug into every existing system, rail, and domestic organization rather than supplant them, because a self-contained financial island is far less valuable than enhancing an existing ecosystem — capturing Metcalfe's-law network value rather than building a closed system.
“We would much prefer that Stripe plugged into every existing system, rail, domestic organization, rather than that we tried to come along and supplant them.”
Stripe's customers are currently outgrowing the internet economy, but for mathematical reasons these must eventually converge, meaning Stripe is ultimately doomed to grow at the rate of the economy — though abundant low-hanging fruit in optimizing the under-digitized 'meta system of business' means this convergence may be many decades away.
“Stripe is doomed to eventually grow at the rate of the economy, there is only a question of how long it takes to get there.”
Contrary to common advice against working with friends or partners, working with people you're very close to is underrated — all of Collison's significant ventures (Stripe with John, Arc with Patrick Hsu and Silvana, Fast Grants with Tyler and Silvana) have been with people he's close to, providing both meaning and real complementarity, and these relationships can endure for decades and outlast the ventures.
“The general thing I'd say is: working with people you're close to is underrated.”
AI progress over the last eight years has been very difficult to forecast — verdicts given in 2019-2021 would look quite different now — so the appropriate stance is humility, with the central open question being to what degree scaling laws hold and what capability level they approach, the answer being highly sensitive to curve parameters no one yet knows.
“the basic position to a first order has to be some degree of humility. As your blog post identifies, the big question right now is: 'To what degree scaling laws hold?'”
Stripe does not draw a sharp distinction between interface and implementation because customers are essentially purchasing Stripe's architecture — what it makes easy and possible — analogous to Mathematica selling a self-contained universe of primitives for modeling, rather than just a GUI.
“We think of the interface to Stripe as being the architecture.”
Stripe's internal LLM tooling found surprisingly high returns from letting people share prompts across the organization (e.g. a SQL-query-optimization prompt others discover and reuse) and from routing all LLM access through a central bus, enabling experimentation across models with observability into performance — now running millions of invocations per day across dozens of production use cases.
“we added sharing prompts across different people, so that somebody might discover these prompts... the collaborative abilities there have proven surprisingly high return.”
Information dissemination via podcasts is valuable because it increases the 'catalytic surface area' of certain kinds of information, citing the example of a prize-leading contender who learned about an opportunity by listening to the podcast.
“Increasing the catalytic surface area of certain kinds of information is a valuable thing in the world, so I'm very glad you're doing the podcast.”
The host introduces Patrick Collison, CEO of Stripe, as the guest.
“Today I have the pleasure of speaking withPatrick Collison, CEO ofStripe.”