Patrick Collison on Stripe’s Early Choices, Smalltalk, and What Comes After Coding

Episode Summary

Executive Summary: Patrick Collison and Cursor CEO Michael Truell discuss how programming may evolve from text editing into a richer development environment powered by AI, drawing lessons from Smalltalk, Lisp, and Mathematica. Collison argues that APIs, data models, and abstractions have durable business consequences, while Stripe’s V2 rewrite shows how hard it is to migrate core systems. The conversation also covers AI’s limited near-term productivity impact, progress studies, and ARC’s biology foundation models.

Main Topics: Programming as a development environment, not just a text editor (Priority: 5/5): Collison argues the industry has under-explored IDEs and runtime-integrated workflows. He wants hover-time profiling, logging, production values, and tighter connections between code, runtime, and debugging—more like Smalltalk, Lisp machines, or Mathematica than typical editors. AI-assisted software creation and refactoring (Priority: 5/5): The speakers discuss Cursor’s direction: letting AI think in the background, run code, react to outputs, and return partial solutions. Collison believes AI will matter not only for generating new code but also for beautifying, refactoring, and reducing the burden of large codebases. Stripe’s early technology choices and long-term consequences (Priority: 5/5): Collison reflects on choosing Smalltalk, Ruby, and MongoDB, emphasizing that early abstractions shape company strategy, org design, and product evolution for years. He frames Stripe’s history as evidence that initial architecture decisions can have durable business ramifications. Stripe V2 APIs and difficult migrations (Priority: 5/5): Stripe is rewriting core abstractions and shipping V2 APIs. Collison says the hard part is not defining new APIs but making them coexist with legacy systems and customer integrations, likening the effort to an instruction-set migration. AI, productivity, and macroeconomic impact (Priority: 4/5): Collison is skeptical that current AI tools have yet shown up clearly in productivity statistics. He notes mixed evidence in GDP and cites a new paper suggesting no observable productivity gains from language models so far, while acknowledging diffusion takes time. Biology as a new programmable system (Priority: 4/5): Collison describes ARC’s work on foundation models for biology and a 'virtual cell' concept. He argues that improvements in sequencing, machine learning, and gene-editing now make it possible to read, think, and write at the level of the cell, potentially transforming complex disease research. Progress studies and the uncertainty of the future (Priority: 4/5): Collison says progress studies is more relevant now because the future is more open-ended due to AI, geopolitics, shifting ideologies, and industrial competition. He argues that the 'Schwartz window' of plausible futures has widened significantly.

Key Arguments: APIs and abstractions are not just technical details; they can have long-lasting business and organizational consequences. AI will likely be most valuable when it augments the whole software lifecycle, especially debugging, runtime awareness, and refactoring—not just code generation. The best development environments collapse the gap between code, runtime, debugger, and data, enabling faster, more interactive iteration. Stripe’s V2 experience shows that rewriting abstractions is easy in isolation but extremely hard when old and new systems must coexist. Current AI advances have not yet clearly translated into broad productivity or GDP gains, likely because diffusion through real workflows is slow. Biology may become programmable through the convergence of better reading (sequencing), thinking (models), and writing (CRISPR/editing) tools. Early technology decisions made at startup formation can constrain or empower a company for decades. Progress studies remains important because the range of plausible futures is now wider and less predictable than it seemed a decade ago.

Data Points: Stripe API availability: 99.99986% - Collison says Stripe’s critical API availability last year amounted to 44 seconds of unavailability. Annual downtime: 44 seconds - Used by Collison to quantify Stripe’s critical API unavailability over a year. Stripe age: 15 years - Collison notes Stripe is about 15 years old while still living with early design decisions. V2 API timeline: Designed in 2022; started shipping this year - Stripe’s new API abstractions were planned in 2022 and began rolling out in the current year. Stripe employee usage of Cursor: Hundreds and soon thousands - Truel says Cursor is used daily by a large and growing number of Stripe employees. Improvement estimate from AI: Half a percent GDP growth per year - Collison references Jack Clark’s estimate for AI’s potential macroeconomic impact. Time horizon example: 2005 to 2015 vs. 2025 to 2035 - Collison contrasts a narrow historical future window with a much broader one today. Aired venue: Previously aired on Cursor's podcast / A16Z Podcast - The conversation is presented as a re-airing of an earlier Cursor podcast episode on A16Z.

Pivotal Quotes: "I think the basic idea of as development environment and not just text editor is really the right idea." — Patrick Collison: Collison explains what he wants modern programming tools to recover from Smalltalk, Lisp machines, and Mathematica. "If you put those together, you now have the ability to, again, at the kind of level of the individual cell, to read, think, and to write." — Patrick Collison: Collison describes the biological tooling stack of sequencing, models, and gene editing. "It's not that useful to just define these APIs in isolation... it feels a bit more like an instruction set migration for a chip architecture." — Patrick Collison: He explains why Stripe’s V2 migration is difficult despite the APIs themselves being conceptually straightforward.

Implications: The next wave of software tools may blend editor, debugger, runtime, and AI into one environment. For companies, API design and data models will matter even more. In biology, similar platform shifts could make complex disease more tractable.