Patrick Collison, CEO of Stripe and co-founder of Arc Institute, discusses career advice for people in their 20s, the Arc Institute’s approach to biomedical research, Stripe’s role in global commerce, and his thinking on AI, progress studies, and institutional design.
Career advice for people in their 20s
Collison questions whether his earlier advice to teenagers — to move to San Francisco — still applies to people in their 20s.
San Francisco culturally valorizes the founder archetype: striking out on your own, dismissing received wisdom, building companies like Steve Jobs or Bill Gates. Stripe itself fits this pattern to some degree.
But many of the world’s most important inventions require accumulating deep technical expertise over decades — a path San Francisco does not culturally encourage.
Herb Boyer co-founded Genentech and produced cheap insulin with recombinant DNA, but only after decades of accumulated knowledge; he was not in his 20s.
Patrick Hsu at Arc Institute announced “bridge editing,” a new recombinase technology for inserting DNA into genomes — work that required years of deep technical training.
Collison does not presume to know what any individual should do, but wants readers of his advice to know that paths requiring long periods of study and skill accumulation are equally valid and important.
On finding the right environment for learning:
People should try to find the “gradient of maximal learning” in whatever domain they care about.
In biology specifically, there are very few successful pure autodidacts; at some point, direct experience in a top lab is essential.
The book Apprentice to Genius follows three generations of scientists and examines what is transferred from mentor to mentee — including problem selection (choosing what to work on) and learning what high standards actually look like.
Practical advice: identify domains you’re interested in, then figure out where the highest standards are embodied and go experience them firsthand.
On contrarianism:
Collison pushes back on reflexive contrarianism — being contrary just because the prevailing mood has shifted is still following the herd, just with “a sign bit inversion.”
Following prevailing tides is “freaking hard to do in practice” even though everyone agrees you should resist it.
Progress studies and the NIH
Collison is skeptical of financially-oriented frameworks for analyzing science funding.
Noah Smith’s concept of “moneyism” — the presumption of constant elasticity between investment and outcomes — is flawed because the conversion rate between inputs and outputs is not a cosmological constant, and financial constraints are rarely the only binding constraint.
Pre-WWII, there were roughly 1% as many practicing professional scientists in the US as post-WWII, yet a lot of important work was done. The relationship between spending and output is not linear.
A better approach: ask what success looks like at the microscale and what is actually happening today.
Fast Grants surveyed grant recipients and found that 79% said their research agenda would change “a lot” if they had flexible funding they could direct however they wanted.
The question should not be “Should the NIH budget be X or 1.1X?” but rather “How constrained should an NIH grantee be in choosing their research agenda?”
On whether slowdowns in scientific progress are structural:
Collison finds the constancy of US GDP growth puzzling and does not have a clean explanation.
One possibility: many observed phenomena could be explained if society is adding exponentially more unproductive capacity rather than productive capacity — not because the marginal people are bad, but because of how components interact.
He is not persuaded that the problem is simply a limited number of geniuses (e.g., John von Neumanns). Cases like Gerty and Carl Cori’s lab at Washington University in St. Louis — where six students went on to win Nobel prizes — suggest that organizational structures and cultural practices matter and are, in principle, more replicable.
Arc Institute
Arc Institute is a biomedical research organization co-founded by Collison, Patrick Hsu, and Silvana Konermann.
Theory of change: the current system of academic biomedical research is homogeneous and suboptimal.
The standard model — universities, labs run by principal investigators, NIH grants reviewed by committees with rigid scoring criteria — is not necessarily bad, but homogeneity is bad in any ecosystem seeking tail outcomes.
Arc was designed as a complementary model with three key differences:
Curiosity-driven funding: Scientists are funded to pursue whatever they want, rather than for specific NIH-style projects.
In-house infrastructure: Scientists can draw on shared platforms and capabilities rather than building and maintaining everything themselves, which circumscribes ambition in standard academic labs.
Career paths for non-PIs: Arc hires scientists who have finished postdocs or grad school and want to do research without becoming principal investigators — a career path that barely exists in the university system.
Example: the bridge editing discovery was led by a senior scientist who had finished his postdoc and may not have wanted to become a PI. It is plausible this work would not have happened in a standard NIH-funded context, given that Jennifer Doudna’s CRISPR work was funded by DARPA after NIH rejections, and Katalin Kariko’s mRNA vaccine NIH applications were famously rejected.
On the long-term potential of functional genomics:
CRISPR can be used not just therapeutically but as a discovery tool — systematically perturbing each of ~20,000 genes to understand their effects, especially under stressors or treatments.
Most major unsolved diseases (autoimmune, cancers, cardiovascular, neurodegenerative) are “complex” — not monogenic or purely infectious, but involving combinations of genetic and environmental factors.
Understanding the genetic component could illuminate general pathways, which could then be targeted with conventional technologies.
On dual-use risks in biotech:
Collison does not think the binding constraint on harmful use of biotechnology is frontier biological knowledge — currently known techniques could be applied malevolently without new inventions.
The more concerning threat is something like a sufficiently sophisticated LLM that could help anyone synthesize and disperse a pathogen like smallpox.
However, the world already faces severe pandemic risks from naturally occurring pathogens, plus all non-pathogenic diseases. Advancing capabilities to solve current diseases is likely close to the same thing as advancing defenses against potential future biothreats.
AI and Fast Grants
On AI forecasting:
Collison emphasizes humility — predictions from 2021, 2023, or even early 2024 would have looked very different from each other.
The key open question is the degree to which scaling laws hold and what the models are approaching as they improve.
The meta-lesson from COVID and Fast Grants is that the most important thing is not predicting crises in advance but having competent individuals who can synthesize and organize information, and having institutions ready to act when crises occur. The “adaptability premium” will likely increase over the next decade.
On Fast Grants retrospectively:
Fast Grants was run by three people (Collison, Tyler Cowen, and Collison’s wife Silvana Konermann) — “three beloved squirrels in a trench coat.”
Collison does not know whether institutions like the NIH, NSF, CDC, or FDA have done retrospectives on their COVID-era performance, but notes that organizations are generally not good at self-reflection, and it is unclear who would have the incentive to do so objectively.
On AI agents and financial infrastructure:
Autonomous transactions already exist in primitive form (usage-based billing where no human pushes a button).
Collison expects a smooth continuum toward greater autonomy, not a sudden shift.
Interesting questions include: legal responsibility and liability for AI agents, which payment rails are best suited, transaction velocity (billions per second vs. one large transaction per day), and whether crypto — as the part of financial services de facto exempt from AML — might play a role.
Stripe history and moats
On when Stripe could have been founded:
Depending on definition, payment companies existed decades earlier (PayPal is a kind of Stripe), but the specific tailwinds Stripe benefited from — app stores, the on-demand economy, the post-YC startup boom — were idiosyncratic to the late 2000s/early 2010s.
Collison’s story of Stripe is primarily one of market inefficiency: many technology companies were already in the payments market, yet Stripe was able to build a large business.
On moats:
Collison believes moats are typically overrated. For most products and businesses, things can just be done much better.
Payments would seem to have many sources of defensibility (network effects, data economies of scale, regulatory barriers), yet Stripe exists alongside a whole fintech ecosystem.
To the extent Stripe has a moat, Collison thinks it is organizational and cultural: people at Stripe deeply care about solving the problems they say they are solving, and they are continually paranoid about what might supplant their approach.
He references Conquest’s third law — “organizations should be modeled as if run by a cabal of their enemies” — and notes that over generational turnover, the incentives of people in an organization can drift significantly from the original mission.
On institutional longevity:
Collison is interested in examples of organizations that have maintained their original mission and competence for decades or centuries.
He notes that in Denmark and parts of Europe, many large corporations (Novo Nordisk, Maersk, Lego) are controlled by non-profit foundations, which can embed mission into legally binding constitutions.
Novo Nordisk’s constitution requires making insulin broadly available and reinvesting profits in R&D — plausibly causal in their development of GLP-1 agonists.
He finds it plausible that shareholder capitalism attenuates the duration of organizations, though it is unclear whether that is necessarily bad.
Stripe Climate and AMCs
Stripe Climate began in 2018 when Collison observed that virtually no carbon removal companies existed and no one had ever purchased from them.
Stripe started contracting with early carbon removal companies to transfer dollars and confer credibility on the sector.
In 2021, Stripe formed Frontier, an Advanced Market Commitment (AMC) — inspired by the vaccine AMC for developing-world diseases — raising $1 billion with Stripe, Shopify, Alphabet, Meta, JP Morgan, and others.
Frontier has contracted with 40–50 carbon removal companies, the majority of which did not exist when the initiative started. In an anonymous survey, 74% of these companies said Frontier played a causal role in their decision to start.
On other areas where AMCs could be effective:
There are many biomedical innovations that are socially beneficial but not patentable, so no one has incentive to fund them.
Example: mannose, a generic sugar that early research suggested might selectively kill tumors, but cannot be patented, so it is unclear who would fund clinical testing.
There are still many vaccines that could exist but do not (e.g., Lyme disease — a vaccine was withdrawn over safety concerns that Collison thinks were misplaced).
Collison is open to sharing the AMC model with others who might apply it in different fields.
Craft, beauty, and API design
Collison is a well-known appreciator of craft and beauty, and sees them as compatible with scale and speed.
Many of the world’s most successful companies are distinguished by their appreciation for craft: LVMH, Tesla, Apple, TSMC.
Even in Stripe’s domain — selling primarily to businesses — aesthetic qualities matter significantly to customers.
These qualities are generally unquantifiable, yet people demonstrably care about them.
Even if customers did not value craft, building an organization that indexes heavily on it may be worthwhile because the best people want to work with the best other people.
On interface vs. implementation:
Collison does not draw a sharp distinction between the two. He thinks of Stripe’s interface as its architecture — similar to Mathematica, where Stripe provides primitives and tools that let users model whatever they need on their own terms.
The analogy is imperfect (Mathematica gives access to platonic mathematical objects; Stripe gives access to Visa error codes), but the idea of a self-contained universe for modeling is a useful source of intuition.
On API design as a discipline:
API design does not get enough study as a discipline, yet it can have compounding positive or negative effects on platforms over decades.
Example: iOS development objects prefixed with “NS” (from NeXT in the 1990s) persisted for much of the iPhone’s history.
Unix’s architecture has worked for over half a century despite many shortcomings.
Stripe tries to design for multi-decadal abstractions. When introducing something new, they ask: “Can we stand behind this in 2044?”
Financial infrastructure and global payments
On Visa and the card networks:
Collison has a nuanced view of Visa and MasterCard. While people complain about interchange fees, he notes that the card networks got their architecture so right that it has served vastly different use cases — replacing store credit, supplanting traveler’s checks, enabling online transactions — for decades.
Dee Hock, who designed Visa, was a remarkable person.
Interchange functions as a distribution incentive fee, paying for customer acquisition and maintenance. Much of it is remitted back to consumers as rewards.
Countries where card networks did not rise (e.g., Germany) have worse online payment experiences. China has ubiquitous digital payments (Alipay, WeChat) but less sophisticated consumer credit.
On designing payments from first principles:
Collison notes that when you layer in customer support, consumer protection, fraud prevention, AML controls, and credit, costs seem to asymptote at around 2–3% across very different systems.
Central banks are now reinventing payment systems from scratch: PIX in Brazil (launched 2020, now used weekly by a majority of Brazilian adults), UPI in India, Swish in Sweden, and others across East Asia and Switzerland.
He has not seen evidence that the ~2% level is massively inefficient, though it may not be optimal.
On global tax as a bigger change than transaction economics:
Jurisdictions are increasingly imposing sales taxes on businesses that have no physical presence there, creating a combinatorial problem of buyer jurisdictions, product types, and tax rates.
These taxes (often 5–10%) are a much bigger deal than transaction fees for internet businesses.
On Stripe’s growth and complements:
Stripe’s customers in aggregate are outgrowing the internet economy as a whole, and there is significant headroom before Stripe’s volume converges with global economic growth.
Growth comes from both the internet economy expanding and Stripe enabling activity that would not otherwise exist.
Example: Stripe was counterfactually responsible for some of the podcast host’s monetization — the host would not have charged for a newsletter without Substack’s ease of use, and would not have known how to receive ad payments without the LLC and bank account set up through Stripe Atlas.
Stripe’s strategy is to plug into every existing system and rail rather than building self-contained financial islands — a “financial air network” rather than a financial island.
As Stripe expands to more poorly served markets (e.g., Albania vs. the US), the marginal impact of providing financial infrastructure increases.
Stripe culture and operations
On writing culture at Stripe:
Writing serves both the reader (efficient communication, intertemporal understanding of thought processes) and the writer (organizing one’s own thoughts). Collison says the two are not separable.
He draws an analogy to Bruno Latour’s argument that the printing revolution partially caused the scientific revolution by making knowledge more rigid and thus easier to falsify. Textual cultures enable different kinds of collaboration and consistency than oral cultures.
On AI at Stripe:
Stripe built an internal tool for integrating LLMs into production services and workflows, including shared prompts across employees.
Example: an employee created a prompt for optimizing SQL queries that is cheap to invoke and sometimes produces good suggestions.
Stripe routes all LLM access through a central bus for observability and experimentation across models.
They are making millions of LLM invocations per day across dozens of production use cases.
Collison expects organizations with strong writing cultures to be among the first to experience AI productivity gains, since the model has a rich corpus of context to draw from.
On reliability and deployment:
Stripe processes roughly $1 trillion per year (~1% of global GDP), making outages extremely costly.
They deploy production services in the core charge flow around 1,000 times per day, with automated canary deployments (small traffic sliver incrementally expanded).
They maintain approximately 99.9995% reliability (~2.5 minutes of unavailability per year).
This requires enormous investment in automated measuring systems, secondary controls that detect deviations before they cause production problems, and multi-year tenacious adoption of operational excellence practices — most of which were well understood by production engineers in the 1930s.
Collison notes that the tech industry culturally values the spontaneous, creative, and iconoclastic, and does not give enough credit to process and operational excellence.
Big business and Stripe’s growth trajectory
Stripe’s customers are outgrowing the internet economy, but mathematically this must eventually converge. Collison thinks it will take many decades because there is enormous low-hanging fruit in basic questions businesses have not optimized.
Examples: extending capital to businesses (which then grow faster persistently), helping businesses decide which countries to sell in (many have not thought deeply about this).
On big business vs. startups:
Collison pushes back on the intuition that only startups drive innovation. A significant fraction of important inventions in any sector comes from established businesses.
Big businesses tend to pay better, be more efficient, produce more consumer surplus, and are responsible for more aggregate innovation than small businesses — yet public sentiment favors small business.
Large enterprises typically come to Stripe not to migrate existing operations but to launch new business lines or sell existing products to new markets — both of which represent genuine economic value creation.
Tyler Cowen wrote a book on this theme: Big Business: A Love Letter to an American Anti-Hero.
Personal reflections
On working with close partners:
Collison has co-founded or co-run every significant venture in his life with people he is very close to: Stripe with his brother John, Arc with Patrick Hsu and Silvana Konermann, Fast Grants with Tyler Cowen and Silvana.
He finds working with close partners underrated and expects to work with John for decades. Stripe would be less effective without either of them.
On Charlie Munger:
Collison and John recently published Poor Charlie’s Almanac. He does not know whether Munger commented on their relationship or whether it reminded Munger of his partnership with Buffett.
