Patrick Collison, CEO of Stripe, sits down with Michael Truel, CEO of Cursor, to discuss programming paradigms, development environments, and the intersection of AI with both software engineering and biology. Collison brings a unique perspective as both a software industrialist running a major payments company and a moonlighting researcher involved with ARC, a biomedical research organization working on foundation models for biology.

The conversation spans Collison's early programming experiences with Smalltalk and Lisp, including his formative work with Paradigms of AI Programming by Peter Norvig and Society of Mind by Marvin Minsky during his AI bot experiments. They explore how early technical decisions at Stripe continue to shape the company 15 years later, the challenges of shipping V2 APIs, and why AI hasn't yet moved productivity numbers despite widespread adoption.

Collison advocates for a return to integrated development environments that blur the lines between runtime and editing, drawing inspiration from Lisp machines and Smalltalk's interactive debugging capabilities. The discussion also covers his work at ARC on programming human biology, where new technologies for reading DNA, thinking with neural networks, and writing with CRISPR create what he calls "a new kind of Turing loop" for tackling complex diseases.

Smalltalk's Interactive Development Environment

Collison's first startup was built in Smalltalk after finding Ruby frustrating compared to Lisp, specifically seeking continuation-based web frameworks that Ruby lacked.

Smalltalk's development environment allowed real-time debugging: "You could encounter an error with some web request, edit the code to fix the error, and then resume higher up in the stack such that the entire web request would just complete" - Patrick.

The interactive debugging eliminated lengthy feedback loops, replacing "this binary search and find the problem and eventually deploy a fixed version, a process that could take an hour" with immediate fixes.

Hiring wasn't a problem despite the obscure language choice - "nobody knew it, but it was easy to teach them" and "smart people learn languages really quickly" - Patrick.

Early AI Experiments and Formative Books

Collison built an AI bot for MSN Messenger using "a really simple Bayesian next-word predictor" trained on conversation data rather than general text corpora.

Paradigms of AI Programming by Peter Norvig was "a really formative book" that introduced various AI approaches, though it notably didn't cover neural networks.

He experimented extensively with genetic algorithms instead of neural networks because "it takes a lot of computer training in neural nets" while genetic algorithms were "more practical on your own computer."

His genetic algorithm work included optimizing keyboard layouts, discovering that "it is, in fact, basically Dvorak using a genetic approach," leading both Collison brothers to use Dvorak keyboards.

The Case for Integrated Development Environments

Collison advocates for development environments that unify runtime, debugging, and editing: "I think the basic idea of as development environment and not just text editor is really the right idea."

He envisions rich runtime integration where hovering over code shows "profiling information about just the runtime characteristics of that code" and "logging and error information overlaid."

Variable inspection should show "the most common values that it takes on in production," creating deep integration between development and runtime environments.

While appreciating Bret Victor's "Inventing on Principle" work, Collison notes limitations: "I reason much more kind of symbolically and sort of lexically than I do visually and graphically."

Stripe's Foundational Technology Decisions

Early decisions to use Ruby and MongoDB still define Stripe 15 years later, with Collison comparing it to "this big bang moment where these tired, overworked, maybe overcaffeinated founding team members are willy-nilly making these initial technical decisions."

The MongoDB choice stemmed from Collison's "principled objection to SQL" due to "too much of a translational kind of mismatch between the domain of the application and that which SQL natively makes expressible."

Stripe achieved 99.99986% API availability in the past year, equivalent to "44 seconds of unavailability through the whole year," which Collison believes is "the best in the industry."

The technology choices required significant infrastructure investment: "we had to build a lot of infrastructure in order to make MongoDB as fault-tolerant and as distributed and as durable and as reliable and everything as we needed it to be."

Stripe V2 APIs and Long-term Abstractions

Stripe V2 APIs launched in 2024 after being designed in 2022, unifying previously separate entities like "end customers, things like sub-accounts, things like recipients for different kinds of payments" into the same entity representation.

The migration resembles "an instruction set migration for a chip architecture" rather than a typical product launch due to coexistence requirements with existing systems.

Collison's key design principles include "unify everything you can plausibly unify" and "make anything that plausibly could be an n-by-m relationship to support that."

The team validated designs by "literally write the integrations that would exist in the new world" to ensure the APIs "feel right" when implemented.

AI's Missing Productivity Impact

Despite widespread AI adoption, recent research suggests "one does not, in fact, observe productivity improvements stemming from use of language models" according to a paper Collison referenced.

Jack Clark from Anthropic, despite being "really an optimist," expects AI to increase GDP growth by only "half a percent a year" - significant when compounded but modest in absolute terms.

Global GDP patterns don't support AI-driven growth: "if we thought that the encouraging GDP figures we have seen in the US over the last two years are attributable to some of these new technologies, I think you would also expect to see them in other countries."

Collison finds LLMs useful for factual questions but disappointing for writing: "I wish they were useful for writing, but I usually end up dissatisfied with the writing that they produce."

Programming Biology with Foundation Models

At ARC, Collison is working on foundation models for biology, motivated by the fact that "humanity has never cured a complex disease" like most cardiovascular disease, cancers, or autoimmune conditions.

Three new technology classes enable biological programming: better sequencing for reading, neural networks for thinking, and CRISPR/base editing for writing genetic modifications.

These technologies create "the ability to, again, at the kind of level of the individual cell, to read, think, and to write" - forming "a new kind of Turing loop and to have its own sort of completeness."

The approach targets complex diseases where "the pleiotropy of the genes in terms of all the different parts of the body and the systems and the mechanisms inside the cell" creates "so much combinatoric complexity."