Patrick Collison on Stripe’s Early Choices, Smalltalk, and What...
a16z PodcastFull Title
Patrick Collison on Stripe’s Early Choices, Smalltalk, and What Comes After Coding
Summary
This episode features a conversation between Patrick Collison (Stripe CEO) and Michael Truel (CEO of Cursor), exploring early technology decisions at Stripe, the evolution of programming environments, and the potential impact of AI on software development and other fields.
The discussion highlights the long-term consequences of foundational technology choices, the importance of integrated development environments, and the ongoing evolution of AI and its role in scientific discovery and complex problem-solving.
Key Points
- Patrick Collison's first startup utilized Smalltalk due to its superior development environment, which allowed for interactive debugging and error correction, a feature he valued over mainstream language adoption.
- Stripe was founded using Ruby and MongoDB, choices that, despite the company's success and uptime, have required significant infrastructure development and ongoing maintenance, demonstrating the lasting impact of early tech decisions.
- The conversation emphasizes the stagnation in programming paradigms over the last 20 years, contrasting it with the potential for more integrated and interactive development environments inspired by Lisp machines, Smalltalk, and Mathematica.
- Collison touches upon his early AI experiments, including a Bayesian next-word predictor, and reflects on how AI has not yet demonstrably improved productivity numbers despite its advancements.
- The evolution of Stripe's V2 APIs is discussed as a complex "instruction set migration" rather than a simple product launch, underscoring the difficulty of updating core abstractions while maintaining backward compatibility.
- Collison expresses a belief that the future of software development may involve a shift from traditional coding to more declarative, higher-level interactions with AI, akin to advanced compiler technology.
- The discussion touches on the importance of API and data model design, drawing parallels to the success of iOS frameworks over Android's in influencing developer preference and ecosystem vibrancy.
- Regarding progress studies, Collison argues that the increasing uncertainty and the impact of AI make such fields more pressing, as the future becomes less predictable.
- A significant portion of the conversation delves into the potential of AI to revolutionize biology, enabling humans to "program" biological systems by combining reading (sequencing), thinking (deep learning), and writing (CRISPR) technologies.
- Collison shares his personal use of AI, primarily for factual queries and research, expressing dissatisfaction with current AI-generated writing capabilities.
- He highlights the difficulty in observing broad productivity gains from LLMs, suggesting that diffusion and integration into the economy are complex and time-consuming processes.
- The conversation concludes with Collison suggesting specific areas for improvement in AI-powered development tools, focusing on runtime integration, code refactoring, and maintaining software beauty and quality.
Conclusion
Early technology choices, particularly in programming languages and data stores, have profound and lasting impacts on a company's trajectory, requiring significant engineering effort to manage over time.
The evolution of development environments towards more integrated and interactive experiences, as seen in older languages like Smalltalk and Lisp machines, holds significant potential for improving programmer productivity and creativity.
AI is poised to revolutionize not only software development but also complex scientific fields like biology, enabling a more systematic and powerful approach to problem-solving and discovery.
Discussion Topics
- What are the most significant underrated ideas from older programming languages that could revolutionize modern software development?
- How will the increasing integration of AI into development workflows change the definition and practice of "programming" in the next decade?
- Given the complex interplay of technology, economics, and societal factors, what are the most critical questions we need to address to ensure AI's development leads to positive, rather than negative, societal outcomes?
Key Terms
- Turing loop
- A concept in computer science that refers to a computational model with theoretical completeness, similar to a Turing machine.
- Instruction set migration
- The process of updating or changing the set of commands a processor understands, often requiring significant software rewrites.
- Bayesian next word predictor
- An AI model that predicts the next word in a sequence based on probability, often used in language generation tasks.
- Lisp machines
- Specialized computers designed to efficiently run the Lisp programming language, known for their interactive development environments.
- Conway's Law
- A principle stating that organizations design systems that mirror their own communication structure.
- Epistemic technology
- Technology that aids in the acquisition of knowledge and understanding.
- Pleiotropy
- The phenomenon where one gene affects multiple, seemingly unrelated phenotypic traits.
- Monogenic diseases
- Diseases caused by a mutation in a single gene.
- Etiology
- The cause, set of causes, or manner of causation of a disease or condition.
- RLHF (Reinforcement Learning from Human Feedback)
- A technique used to fine-tune AI models by incorporating human preferences and feedback.
- LLM (Large Language Model)
- A type of AI model trained on vast amounts of text data, capable of understanding and generating human-like text.
- API (Application Programming Interface)
- A set of rules and specifications that allows different software applications to communicate with each other.
- V2 APIs
- Refers to the second version of Stripe's APIs, indicating a significant update or redesign.
- Non-mainstream language
- A programming language that is not widely adopted or used in the general industry.
- Heterodox
- Departing from established beliefs or doctrines.
- Prolixity
- The use of too many words; wordiness.
Timeline
Patrick Collison recounts his first startup being written in Smalltalk due to its powerful development environment.
Collison explains the decision to use Ruby and MongoDB for Stripe and the implications of these early choices.
Collison discusses how ideas from older programming languages have influenced modern ecosystems like JavaScript and Python.
The conversation shifts to the concept of integrated development environments, drawing parallels with Lisp machines and Mathematica.
Collison speculates on the future of software development, suggesting a move towards AI as an advanced compiler/interpreter rather than just a coding assistant.
The limited experimentation with programming paradigms over the past 20 years is noted.
The concept of lock-in with programming languages due to human learning and existing codebases is discussed.
The difficulty of managing large codebases and the potential for AI to alleviate this burden is explored.
The importance of creating codebases and environments that are easy to modify is highlighted, referencing a CS class.
The idea of AI assisting in the beautification and refactoring of code is presented.
The application of programming concepts to inter-organizational software development is considered.
Collison stresses the critical importance of taking APIs and data models seriously, citing their impact on organizational structure and business outcomes.
The comparison of iOS and Android ecosystems is used to illustrate how API and abstraction design impacts developer success and business ramifications.
The long-term endurance of well-designed APIs and abstractions, using Stripe and early iOS classes as examples, is discussed.
The metaphor of a "big bang" moment in startup technology decisions and their lasting consequences is explored.
Stripe's impressive API availability statistics are shared.
The debate around migrating Stripe's core language from Ruby is mentioned, with some services being rewritten in Java.
Collison explains the initial decision to use MongoDB over SQL for Stripe, driven by a dislike for SQL's translational mismatch with application domains.
The development and eventual shipping of Stripe's V2 APIs, a long and complex undertaking, is detailed.
The process of designing and migrating to new APIs, including managing customer upgrade paths, is described as akin to an instruction set migration.
Key lessons learned from the V2 API development include unifying elements and establishing a singular responsible designer.
Collison discusses his personal use of AI, primarily for factual queries and research, with limited satisfaction for writing tasks.
The generic nature of current AI writing output is noted as a point of disappointment.
Collison mentions using LLMs for running code through Cursor and his passive use of voice mode AI while reading.
Collison is characterized as an archetype of a "software industrialist" due to his roles at Stripe.
The relevance of progress studies in the face of increasing technological acceleration and uncertainty is argued.
The current political landscape in the U.S. and its unpredictability are touched upon.
The rapid technological advancements, including AI, stablecoins, and China's role in manufacturing, are identified as drivers of future uncertainty.
The concept of the "Schwartz window" is introduced to describe the broadening scope of contemplatable futures.
A recent paper claiming no observed productivity improvements from LLM usage is mentioned.
The current GDP growth in the U.S. is contrasted with less encouraging growth elsewhere, suggesting slow diffusion of new technologies.
Anthropic's founder's projection of AI increasing GDP growth by half a percent annually is noted.
Collison expresses excitement about the potential to "program human biology" through advancements in sequencing, deep learning, and gene editing.
The historical lack of cured complex diseases is attributed to limitations in epistemic technology.
The significant progress in reading (sequencing), thinking (deep learning), and writing (CRISPR) technologies in biology is highlighted.
The potential for AI to enable higher-level software building, benefiting designers and those less skilled in traditional coding, is discussed.
Potential beneficiaries of AI advancements and their impact on the economy are explored, with a focus on unpredictability and contingency.
Collison provides three personal suggestions for Cursor: runtime integration, refactoring/beautification, and improving the quality and maintainability of Stripe's architecture.
The concern that AI might lead to more "slop" and less of "the best things" is raised.
Episode Details
- Podcast
- a16z Podcast
- Episode
- Patrick Collison on Stripe’s Early Choices, Smalltalk, and What Comes After Coding
- Official Link
- https://a16z.com/podcasts/a16z-podcast/
- Published
- February 20, 2026