# API Design, AI Productivity, and Software Architecture Strategy

**Podcast:** AI + a16z
**Published:** 2026-03-24

## Transcript

It's interesting to me that we haven't experimented in some sense that much with the paradigm of programming over the past 20 years.
Yes.
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.
And this starts to really feel like a new kind of Turing loop and to have its own sort of completeness.
I think that's a case where the right API design, the right abstraction design uh ended up having just quite significant business um ramifications.
I think the basic idea of as development environment and not just text editor is really the right idea.
And that's the thing I want to see a return to.
Patrick Collison wrote his first startup in Smalltalk.
Its development environment let him fix errors mid-request, inspect stack frames, and resume execution.
And he wanted that more than he wanted a mainstream language.
He and his brother chose Ruby and MongoDB for Stripe instead.
Those decisions still define the company 15 years and 44 seconds of annual downtime later.
Now Stripe is shipping V2 APIs, rewriting core abstractions first designed in 2010.
It's taken years.
Defining the new APIs is the easy part.
Making them work alongside everything already built on the old ones is, as Callson put it, more like an instruction set migration than a product launch.
This conversation previously aired on Cursors Podcast also gets into why AI hasn't moved product-to-v numbers, what today's dev environment could steal from Lisp machines, and Collison's work at Arc on foundational models for biology.
Michael Truell, CEO of Cursor, sits down with Patrick Collison, CEO of Stripe.
Well, it's great to have you.
Thanks for being here.
Thanks for having me.
Great to be here.
Yes.
I've heard that your first startup was written in Smalltalk.
Please explain.
I don't know what there is to explain.
It's the best programming language.
Well, I had, I'd worked on Lisp and Lisp dialects before that.
And actually I worked on Lisp Web Frameworks.
And when we went to build our first startup, we we first wrote it in, we we first implemented it in Rails, and then I found compared to Lisp, that development process kind of frustrating.
And I mean, we don't need to get into full details, but I thought that continuation-based web frameworks were really the right way to implement web applications.
Uh in there's no continuation-based framework in Ruby.
Um, and uh searching around, I found that there was a good one that had just been written in Smalltalk.
Uh so I decided to play with it a little bit.
And then I found that Smalltalk is actually this extremely interesting development environment that had a lot a lot of the aspects of Lisp that I'd really appreciate it there, like um, you know, and uh a fully interactive environment with a proper debugger so that you can you can edit the code while um you know in the middle of some web request or you know in deep in some stack trace or something.
Uh and you know, you could, for example, encounter an error uh 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.
And so rather than this kind of annoying feedback loop of having to add some log statements and like do this binary search, like find the problem and eventually like deploy a fixed version, you know, a process that could take an hour.
You could just like literally inspect the stack frame, see which variable has the wrong value, fix it, like you know, jump back up, hit proceed, and have the whole thing work.
So anyway, the point is uh in the hunt for this continuation-based web framework, uh, realized that Smalltalk in general had just a much more powerful development environment as compared to Ruby slash as compared to basically every other mainstream programming language.
Um and so we we decided to use it for the company, which in hindsight was I mean, I don't know if it was a terrible decision or not.
Uh the reason I think one would think it would be terrible uh is that it would be you know hard to hire people and hard to scale and you know, whatever.
It wasn't hard to hire people, or rather, nobody knew it, but it was easy to teach them.
The company did they know before they joined?
No, no.
They learned really quickly.
Um and then you have smart people learn languages really quickly.
Uh so I don't I don't think that's really a reason not to use a non-mainstream language.
The company didn't work, I think for unrelated reasons.
I think just the idea wasn't that strong.
Um, but we also chose Ruby for Stripe.
So I don't know.
I think maybe the gains were not quite as large as I hoped.
And was your small talk enthusiasm shared by the acquirers of the startup?
And what was the dynamic?
You know, was there like this blissfully ignorant management that foisted this small talk code base on a bunch of unsuspecting developers that were then kind of like piling it over, you know, or yeah, what was the dynamic between the the programmer's management sort of what happened to that small talk code base?
Yeah, yeah, yeah.
Does it still live on somewhere?
I wish.
Um and um I'm I'm 99% sure the answer to that is no.
Uh it um it's the the company that acquired us was mainly a talent acquisition.
So um yeah, the the code base itself was was less relevant.
Okay.
And it was immediately sort of just gone.
Yeah.
Okay, gotcha.
I've also heard that one of your earliest programming projects was working on an uh an AI bot written in Lisp.
True.
And it was, it was something like a uh it was a client for MSN.
Uh-huh.
Um, I know where you found that, but that is true.
Um and um I heard that you got kind of nerd tonight by the idea of trying to get it to pass the Turing test.
Yes.
Um and I'm curious, what did you miss?
You know, why didn't you make ChatGBT?
Um and well, uh maybe a little bit more seriously.
How did it work?
And what was the state of neural networks at the time?
And did you did you consider using um any antecedents to the technology we use today?
Yeah, so so that was the project.
Uh it was um a little critter that used MSN Messenger, uh, which was you know all the rage uh at the time.
I guess that puts me that that's like maybe a specific kind of you know sedimentary layer in the you know chronology of different uh instant messaging solutions and probably dates me quite precisely.
Um it was a really simple Bayesian next word predictor.
Like there was nothing really that sophisticated there to the end there was anything sophisticated, it was maybe that it used like that the training data was the conversations issue had an MSN Messenger rather than kind of general text corpus and and you it it worked reasonably well.
And you know, it better versions looked a couple of words ahead and and you know what have you.
And a couple, I mean, it never really passed the Turing test where you know people have actual suspicion that they're trying to you know exercise this this uh discernment.
Um but it certainly passed some weaker version of the Turing test where you know they they were unsuspecting.
Uh and you know, people ended up having quite quite lengthy conversations with it.
It c and that was part of how I discovered Lisp.
And I remember Paradigms of AI Programming by Peter Norvig being a really formative book and had you know all sorts of uh interesting approaches there.
Um it didn't have anything on neural networks, I'm almost sure.
And I never, I mean, I'd read some Marvin Minsky stuff, Society of the Mind or whatever, on neural nets, but I never really seriously looked at them.
And I actually experimented a lot with uh genetic algorithms.
Um they were, I guess, more practical on you know your own computer.
Like it's it's it takes a lot of computer training neural net.
So I I experimented a lot with with genetic algorithms, and actually I use Vorjak at the keyboard layout because uh it's more comfortable to type on them QWERTY.
Um but uh but uh as does John and my um my brother, so no one can ever use our our computers, but um uh I wrote a genetic um, I don't know, optimizer uh to uh to figure out what the optimal keyboard layout was.
And it turns out it is in fact basically Vorjak using a genetic approach.
Um so I went deep down that rabbit hole, but I never really played with neural networks, and I guess that's why, you know, um that, but probably seven other 70 other reasons is why, you know, um uh I did not create Chat GPT.
There is an old video of you being interviewed, I think, after selling automatic um where you're asked you're asked about small talk.
That's that's where I found kind of that weird weird act.
I think at the time, people asked you why, and one of the things you said was, yeah, I mean, you liked some features about small talk, Lisp Lisp style languages, and you predicted, and I think that this was circa maybe 2008 or something like that, that the mainline C style programming languages would increasingly borrow ideas from these older programming languages.
And that kind of has been the case in the JavaScript Python ecosystems.
Yeah.
Um, do you think that there are any underrated ideas buried away in kind of older, more esoteric programming languages that should be borrowed by the mainline?
Yeah, it's been it's been interesting how a lot of the ideas have been kind of have been borrowed by the JavaScript ecosystem and in a strange way, like through the web inspector.
Yeah.
Um where you have this, I mean, that that's one of the richest runtimes in some sense that people have you know general general exposure to.
I don't think JavaScript has first class stack frames.
Um maybe there's some weird extension or something where you can get that, but it but you know um ECMA script uh doesn't have that, I'm pretty sure.
First class stack frames actually let you do a lot of other things for kind of obvious reasons.
So maybe that's very that's kind of too specific.
Um I mean, I think the idea of, and maybe this is what cursor becomes.
I think the basic idea of as development environment and not just text editor is really the right idea.
And that's the thing I want to see a return to.
That's the thing that the Lisp machines had and genera, that's the thing that to some extent Mathematica has, that's the thing that Smalltalk has.
And I think it's just such a mistake that we have ended up with development environments where there is such a separation between the runtime, um, the text editing, and the um and the environment in which the code I mean, yep.
Well, the the runtime and the place where the code runs can be the same or different, but but there are kind of three maybe slightly conceptually different things.
Uh and in those three environments, they can all coexist in the same place.
And I find like I mean, it's still to this day, I use Mathematica a lot, not because I'm doing some particularly arcane, you know, symbolic mathematics, but because it's just a more efficient development environment.
Now, that's maybe a bit less true with uh LLMs because the Mathematica, you know, the uh um Mathematica does not support support cursor-style prompted development.
Um but uh but that that I think is the core idea that that I wish others would borrow.
And VS Code has been a step to some extent, slightly in that direction, but I think we could take it way further and it would be really I mean what I'd love to see, for example, is uh when I hover over a line of code, um, I would like to see you know profiling information about you know just the runtime characteristics uh of that code or that function or whatever.
Uh I would like to see logging and error information overlaid.
When I hover over a variable, I would like to see um how uh like the most common values that it takes uh uh on in production.
These these kinds of like just rich deep integrations.
Are you a fan of inventing on Principle and the stocks?
Yes, yes.
Yes.
Um I think Brett leans too much.
I mean, I a huge fan of Brett.
Um just he is such an incentive.
Have you've been to Dynamic Land.
Yes, okay.
Yeah.
Um and and have supported it.
Uh so huge fan of Brett.
Um the the place that I've maybe differed, or at least that just resonates with me, you know, somewhat less is Brett is really into this idea of um obviously of uh graphical and visual representations for phenomena.
And I think that works very well in um certain domains, like the kinds of dynamical systems that you know he has demonstrated some of the ideas with.
I think it's often very hard to find such useful spatial, continuous representations for arbitrary systems, like for various parts of Stripe.
I'm not quite sure what that would be, and I'm not sure even if we could find it, you know, exactly how useful it would be.
Maybe it's just me.
I reason much more kind of symbolically and sort of lexically uh than I do visually and graphically.
It might just be personal preference, but I don't know.
The kind of uh paradigm breaking that he's been engaged in, I think is hugely admirable.
Are you gonna make a uh a true integrated development environment?
So we are playing with ideas around letting the AI increasingly take time into the background to uh run its code and react to the output.
And we think that this should all work well together.
Like, you know, we've focused a ton on inflow, speed, and control.
And uh we think that that's really, really important for AI is you know to give programmers the control over everything, have them understand everything the AI is producing, um, also to give them really, really fast iteration loops, programmers hate waiting for things.
Um but in some cases, we think it's now becoming possible to go tell the AI to think for a bit and then come back to you and have the API be a little bit more like the API with another human being.
And we think you want that all to work well together.
So the AI can come back to you with 70% of something, and then you can bring it into the foreground really quickly, work with it, and then spin it back off to the background.
And as part of having the AI spend a bunch of time thinking in the background, to make that thinking useful, you kind of need it to run the code and then react to it.
Or else it's just kind of staring at the thing that it wrote and thinking more.
Maybe I'm supposed to be the one answering the questions rather than asking them.
But do you think in five years the main thing that I'm looking at in cursor will be code or something else?
I think it might be something else.
I think that there are big, big, big simplification, but kind of when you're defining what a piece of software is, there's like the uh logic component, which is what engineers spend a lot of time on of designing exactly how the software works.
There's also for end user applications and things that have GUIs, there's like this visual component.
And I think that there is, you know, maybe it's gonna be us, maybe it's gonna be someone else.
There is a future version of the world where the way you interact with AI is a little bit less like, you know, it's a human helper that you're delegating work to or looking over your shoulder, predicting the next set of things you're gonna do.
And instead, it's a little bit more of an advance in compiler or interpreter technology.
And it can lead you to a world where programming languages actually change.
And they can start to get a little bit less formal, they can start to get a little bit higher level, uh, they can start to be a little bit more about uh what you want and a little bit less about how you do it.
Um I think that it won't look like a Google Doc necessarily.
I think that there are things you want to keep around from programming, like that you know, the naming of objects somewhere, and then using that in a bunch of other places.
I think that there's also this other element too of the visuals of what a piece of software looks like.
And I think, you know, maybe us, maybe some other tool, but I think there's a world where kind of direct manipulation of the UI starts to play a little bit more into it.
Yeah.
But these are kind of far-flung experimental ideas.
Um in general, I will say, and yeah, it's not terrible, but um I feel like they're just it's interesting to me that we haven't experimented in some sense that much with the paradigm of programming over the past 20 years.
Yes.
And many of the things we're discussing here are from the 80s or the 70s.
Uh and um, you know, there are way more developers, obviously, now than there ever have been in the past, but in some sense, the sort of the aperture of um of experimentation there feels like it's it's really not that wide.
Again, the JavaScript ecosystem and a couple of others have done some cool things.
And there's been a lot of even you know, experimentation at the language level with Rust and Go and everything else.
But but at the kind of the development environment level, I don't know, I don't know why it is, but I mean maybe it's just too hard and complicated now, but there's been less than I would have expected.
Yeah.
I agree.
And um, I think that's a good idea.
Maybe this is maybe this helps.
Something we're working on.
Maybe this explains cursor's uh success to some extent, where uh you know you guys are the first people to really take it seriously in quite a while.
Well, I I mean, yeah, I think we also benefit a lot from the why now of like there's now this you know great new color to paint with, um, or set of colors to paint with.
Um I think also there's just a ton of lock-in with programming languages around both the neurons in your head of like programming languages are kind of complex UI for programmers to define exactly how the computer should function.
And so you know, people learn languages, and there's you know, uh people don't like to learn that many things.
Um and then there's also the lock-in of you have a lot a lot of logic sitting around in one language and you need to maintain that.
And I actually think that that's a pretty interesting, or one of our hopes is that as AI programming gets better and better and better, one of the downsides of working on professional applications with hundreds of people dealing with many millions of lines of logic is the weight of the code base really starts to weigh on you.
Yeah.
Um and so the feeling of being in a net new code base where it's just everything feels effortless goes away, everything's you know, a chore, you have to um you know, change one thing here, breaks you know, something else here, and it becomes kind of this big ball of mud.
Um and um uh making that effortless, reducing the kind of weight of an existing set of logic, I think is you know one of the areas in which AI can you know make programming better.
So someone said on Twitter today, maybe it was Andre Carpathy, but maybe I'm misattributing that.
And you know, maybe too many things uh through with uh uh vibe coding get attributed to Andre, like you know, to Churchill or Einstein or something.
But I think that can.
Um but um uh this person, whoever was with making the observation that you know it's one thing to be prompting the creation of code, but another place where AI could conceivably do a lot to help is in the beautification and the refactoring of code bases.
And you can imagine that you know you're producing all this, you know, a little bit ungainly, not quite correctly factored, you know, detritus at the front, and you have this, and then nocturnally, this thing comes up behind you and makes it all you know beautifully factored.
And the only CS class I ever took was this from uh this class from Jerry Sussman uh on um on it was basically focused on, I mean, he he called it large-scale symbolic systems, but really uh what he was trying to focus on was the idea of creating code bases and uh environments and uh abstractions that were easy to modify.
And like there were no assignments in the class where you'd write something from scratch.
Every assignment was about modifying an existing system and thinking about how could you design things in such a way that those modifications become, and you know, they might be quite deep modifications, um, uh become uh straightforward.
And I think that's a lovely idea.
Obviously, in practice, it's often very difficult to do that given all the exigencies and pressures of the things you want to ship, you know, today and next week and so forth.
But if you could have an AI, like often you're writing this stuff, you realize, well, I really should be doing it the beautiful way, but I'm not.
Yes.
Um maybe we could have an AI coming up uh behind us too at Cleanlet.
Yes, yes.
Maybe soon.
One thing that happens to a lot of developers that care, or a lot of people come to development because they care about building things.
They want to make things happen on a computer screen.
And so then that leads them to coding.
And then something that happens to uh, you know, a big group of developers is they eventually realize the software they want to create is is too big that they can't write all of the code themselves, and they have to go to humans to help them write the code.
And so they maybe they then become an engineering manager or director or whatever it is.
Maybe they start a company, right?
And then most of the work becomes not typing code, it becomes coordinating amongst people.
Do you think that there are any ideas from programming that are helpful for the act of kind of programming amongst the organization to get a group of people to uh to build software together?
Interesting.
I think taking uh APIs and data models really seriously.
If I was to do everything at Stripe again, um, I mean, there's a million small things that you would do differently, and even some some kind of big things.
But the the thing that I think we could maybe foreseeably and beneficially done differently, um would be to have spent even more time than we did on APIs and data models.
And you know, part of the reason is um is the uh the um, I guess Conway's law effect of how you know both of those things end up shaping the the organization.
So I guess if you don't deeply internalize that, then uh maybe you've less control over the uh organizational dynamics than you might otherwise like to have.
Um, but also I I think it um it ends up shaping not only, I mean, the weak version of Conway's law is that it uh it shapes your organization.
I think the strong version is that it substantially shapes your strategy uh and and just your your business outcomes.
And this isn't exactly maybe a version of that that, but I often reflect on how the iOS software ecosystem um for a very long time and you know plausibly still today was so much more vibrant and kind of vital and um successful than the uh Android app ecosystem.
And you know, there's a lot of things that are different across those two um ecosystems.
Um there are now way more Android devices in use, I believe, than uh than iOS devices.
Um I think much of the um the fact that app developers tended to prefer um building their apps on iOS and releasing their apps first on iOS, and maybe the iOS version being better than the Android version or you know, whatever, um, is because the the frameworks uh and the abstractions for iOS were just originally better than the than the Android ones.
Uh and so but I think that's a case where the right API design, the right abstraction design uh ended up having just quite significant business um ramifications.
And you know, I think there's kind of a sense that maybe it's not worth dwelling on these things because you know everything in technology changes so rapidly, and uh, you know, whatever assumptions you make, they'll be you know obsolete in two years or something.
I think in practice that's not true.
Uh and that like the right API design and the right abstraction to the right data models can really endure.
And you know, for the first versions of iOS, many of the classes that one used were prefixed with NS.
Uh and S, of course, standing for next step, right?
And so that's a case where the API design survived for you know two decades or more.
And in the case of Stripe, you know, Stripe is now 15 years old, and you know, there are lots of things that we designed 15 years ago that are still in use today, which is you know, kind of good and bad in the sense that they they endured, but also we are still still uh under the uh the living with their faults.
Exactly.
Uh and so anyway, uh that that that's maybe the thing that I would um that's the first thing that comes to mind.
In fact, on that final note, I was talking with uh an engineering leader at you know, kind of a preeminent successful Silicon Valley private company, and um they were talking about how their code base is largely in Scala.
And uh they said that they like to think of kind of the the beginnings of the startup as this big bang moment where this these you know tired overworked, um maybe over-caffeinated founding team members are willy-nilly making these initial technical decisions that then dictate the lives of hundreds of professional engineers in the future.
And that scholar choice was one of them.
Uh, and they sort of live with live with the faults of that um now.
But um, what what are those kind of uh what were the consequential, it could be good or bad, uh initial conditions of the stripe big bang that you guys still live with right now?
Um I mean, I think that metaphor is is well, it it sounds true to me, um, uh is is the first thing I'd say.
I mean, maybe it's a little bit of kind of survivorship bias, uh, where like the actual statement is the uh the early decisions that we made that we never changed are decisions that we lived with, but you know, there's a kind of uh tautology there or something.
Um and there are certainly design decisions we made pretty early on that you know are not true today.
So, you know, early versions of um the Stripe dashboard or something were built extraordinary differently to you know the dashboard of today.
Uh and the converse is also true.
So uh you know, initially we decided to use MongoDB at Stripe, and we decided to use Ruby at Stripe.
Um and those are still quite foundational technologies uh uh uh at Stripe.
And uh you know, we had to build a lot of you know infrastructure in order to make uh MongoDB as fault tolerant and as distributed and as um durable and as reliable and everything as we needed it to be and as it now is, uh like we had uh Stripe's critical API availability last year was 99.99986%, uh which is 44 seconds of unavailable unavailability through the whole year.
Um which is we we fit and others don't publish statistics that are kind of as granular, but we believe that is the best in the industry.
Um so uh so you know the everything that our storage team has built and many other teams, uh, you know, it ended up really working there.
But that that th that was a quite important critical decision.
Um initial decision.
Uh and you know, Ruby Ruby similarly.
I I guess companies sometimes change languages uh you know along the way.
But I but I feel like the initial language chosen tends to have.
I heard there were there were debates in Stripe about, or one of actually one of our co-founders, um uh intern at Stripe early on.
Um or not early on in Stripe's history, early on in kind of our our collective personal history.
Um and uh he remembers there being documents upon documents about a potential Java migration.
Yeah, so we um that partly happened.
Um as in we have rewritten a bunch of key services on Java.
Um so some services for which I don't know, throughput, throughput in particular is uh is really important.
And if you torture Ruby enough and you know, maybe rewrite uh parts of it uh, you know, parts of some um some sort of hot paths uh in C or something, you know, you can get it to be pretty fast.
Um but you're often fighting against the allocator and uh you know various parts of even just like Ruby strings are not that efficient and stuff.
So we we we've rewritten certain services in Java and um now we use both.
Why did you or did you consider anything other than Mongo and uh why did you pick Mongo early on?
And what was the what was the RFC process, RFP process, decision making process for that?
It was just me and John.
So you know we were sitting on the couch.
Um did they did they get through to you with like a a blog?
Was it um the just the reputation of Mongo at the time and open source communities, something else?
I think it was so uh I I wrote a data store for our um for our prior company, uh, an um object-based data store.
And I didn't really like SQL.
I I thought it was there was too much of a translational kind of mismatch between the domain of the application and you know that which SQL natively um you know makes expressible.
And so, you know, with SQL, obviously you have to collapse down into a relatively restricted set of primitive forms.
Whereas in your application, you know, you might have a concept of, I don't know, of uh, let's say in the case of Stripe, of money that doesn't like exactly comport with how the particular SQL database you're using happens to represent money or you know, whatever the case might be.
Um I just had this like principled objection to SQL.
And I'm not endorsing this or saying it was good, but um uh I um as as this interview shows, I suppose, uh I had all sorts of you know strange notions uh about technology.
Um with Stripe, we wanted to be more mainstream uh than uh and a little bit more um, you know, a little bit uh less heterodox uh in our technology choices than our prior company.
And so instead of using Smalltalk, you know, okay, we weren't gonna go to Java, but we went to Ruby, which at least on a relative basis seems more mainstream.
And similarly, rather than write our own object database, we went relatively more mainstream and used Mongo, which still gave a lot of flexibility, you know, by virtue of being a kind of a kind of object data store.
Um so that was fine.
Um everything I've said might disqualify me from you know ever making technology choices for uh for another company, but um Would you do anything differently about Stripe V2?
We haven't talked that much about it publicly yet.
Uh and the answer might be a bit like um, you know, this is the uh the Zhu and Lai uh uh quote about um the um Xiaoping um about the uh the French Revolution, you know, it's too soon to judge.
Um and so back in 2022, I believe, um we, I mean, to this discussion about data models and abstractions, we realized that a couple of the core abstractions in Stripe were just not the right long-term abstractions, and we uh we had to fix that.
Um and so we designed a bunch of V2 APIs.
Fortunately, we had contemplated the possibility of this earlier at Stripe.
So you know, most of the uh the you know REST URIs that uh people uh are familiar with in Stripe are prefixed with slash v1.
Uh they've been prefixed with slash v1 from you know 2010.
Uh and uh and so then in 2022 we decided, okay, we might use the uh the uh you know, we might increment the uh the namespace.
Um so we designed those new APIs, they've started to ship uh this year.
Congratulations.
Thank you.
Um and we're extremely excited about the functionality that it's going to enable.
And you know, without getting into the arcana of it, you know, they will enable things like historically uh we have um we have drawn distinctions and represented separately things like end customers, things like sub-accounts, things like recipients for different kinds of payments, and we're unifying all of those into being um you know into into the same kind of um entity representation, which is on some level clearly the right answer and you know makes a lot of um it will and uh is already changing the businesses of some of our customers because uh they can you know enable their users to do various things without having to re-enter details or maybe to bring the same account across different countries or whatever the case might be.
So anyway, it's been a long journey.
Um and the reason it was a long journey is um I guess the it's it's not that useful to just define these APIs in isolation.
If we just wanted to define them in isolation, that's a pretty easy thing to do.
Uh the thing that's difficult is to make them interoperable with all the existing things at Stripe and to build translation layers uh and so forth.
Um and then to figure out with our customers what a sensible upgrade path you know might look like, because we control our code base, we don't control theirs.
Uh and so it's maybe a I don't want to exaggerate it, but in certain respects at least, it feels a bit more like an instruction set migration for you know a chip architecture or something where the instruction set by itself is easy, but it's all the kind of coexistence questions that that become hard.
But hey, it's hard to ship this year, and we're we're excited about it.
I mean, I I guess your question was maybe what what lessons we've learned from it and maybe do you think there's anything bigger to draw out of that on either projects that are rewrites or thinking about these kind of decades-long abstractions and how to do that well.
My trite answer to that is to unify everything you can plausibly unify.
And how do you test design ideas for V2?
Well, the people designing it, um well, I'll give you one other lesson and then I'll answer that question.
So I think just the other lesson, just maybe a bit to you.
And also, is there some chief API designer who's the the mastermind of the and it's one person, it's not some sort of working group.
There is a working group.
There are working groups, but there is also a singular person who understands and is more than anyone else responsible for the whole.
And I and I think that's necessary.
Yeah.
Um my other kind of trite exhortation would be to make anything that plausibly could be an N by M relationship to support that, because you know if you only support one to N or N to one or you know, whatever, you'll in uh and even if it's non-obvious how it could possibly be N to M, just inevitably end up needing that.
And you know, you'll think, well, you could never have a company that's owned by two different companies or something, but you know, it turns out that sort of every permutation the space uh is uh is in fact uh eventually, eventually explored.
Um as to, you know, how to how to do that well, I I really feel like it's um I like the these new APIs, we we think they're the well, you asked the question of how do we know they're the right APIs.
Um partly from showing early versions of them to customers, partly because the people who designed them had spent many, many years in the um, you know, uh witnessing and living with the shortcomings of the prior versions.
So you know, we were kind of coming with uh strong opinions, but but it but even the strong opinions, I mean one can sometimes you know predict wrongly uh or extrapolate wrongly or overengineer something or whatever.
So I think the cycles of customer validation, customer feedback are extremely important.
I think it's also very important, we did a lot of this, to literally write the integrations that would exist in the new world, because I mean, you really, I mean, I think Java is maybe an example of yes, it fixes a bunch of problems with memory management or whatever that existed with C or C and antecedents, but at the cost of a lot of prolixity and uh and uh overhead.
Um, and in order to kind of safeguard ourselves against you know inadvertently over-engineering things, we force ourselves to uh to write a lot of API code specifically describing uh how we would implement various business models and flows and so forth, just to make sure that when you when you look at it, it feels right.
But I don't want to endorse our approaches too strongly just yet.
I mean, I'm feeling very optimistic, but uh, you know, it's we're we're I don't know what fraction, but 60, 70% done or something, but not like 100%.
Uh and so I don't want to, you know, prematurely declare any victory.
How do you, Patrick Collison, use AI?
Well, I the main ways are the predictable ones, where I use LLM chat tools um uh a lot.
Uh and I use them for.
Mainly for answering qu for kind of factual or empirical questions that I'm curious about.
Uh so for deep research style questions.
I don't always use deep research, and now that the LNs are getting better at tool use and just navigating the web themselves, you know, you don't need deep research as much.
Um but for answering empirical or or factual questions.
Um I wish they were useful for writing, uh, but I I usually end up dissatisfied with the writing that they produce.
Yes.
Um so I I don't really use them very much for that.
And even for editing or or or grading my own writing, I I mean Have you seen any improvements as the models have progressed on the writing?
I know, agree.
Also, yeah, it's that it's surprisingly generic.
Yes.
And um, I'm trying to pump it to like not generic and inserting names of people who say um and it just doesn't work.
And so I have been disappointed at the times when I've given it to the code.
People tell me that the base models um are better at this, um, and it's the sort of normification of RLHF that you know puts it in um in some kind of you know attractor basin.
Um yeah, I I I I have not succeeded in uh in using them uh effectively there.
Uh people say that Plaud is better and O3 is better than earlier OpenAI models.
And on a relative basis, that might be true, but I and I don't want to sound um you know um you know self-lauditory here and suggesting that I'm you know some particularly talented writer.
I don't think I am.
It's just like my personal style um uh differs from the personal style, so to speak, of the models.
And um, you know, in some self-centered way, I when I write, I want to use my personal style.
Um so I use them for for the the factual stuff a lot, and and I find them like terrific for that.
And uh you know, even when I'm reading a book, I'll sometimes activate uh I've been recently using Grox voice mode, um, and I'll just passively ask questions while I'm reading, and you know, Grock is just listening in the background, and you know the answers are are are very helpful.
Um and then I obviously use LMs for writing code and um you know typically mediated through cursor.
So we are interviewing you, Patrick Collison, as kind of uh the most if you had to pick the architect archetype of a software industrialist, I feel like you would be kind of straight out of central casting um uh for a number of reasons.
One is that you are running a large software company, um, a successful large software company.
Two is you started as a programmer and then you know moved to running the company.
Um and then three is the company also builds things for developers.
And so it's kind of the the intersection of you know, uh many circles in the Venn diagram.
Um and so it's helpful to hear about discussing kind of experiences with Stripe.
We are also interviewing Patrick Collison, the moonlighting economist and student of the world.
And so are progress studies doomed now that AI is here.
Is there any need for them?
Well, I was going to say I think the need for progress studies has increased, but again, I don't mean to suggest that proper noun progress studies sees increased need, but I think the kinds of questions that Progress Studies tries to answer are now more pressing and urgent because I think the degrees of freedom are increasing.
And I think there's some Pand Glassian view that AI will just magically solve all the problems and predictions of the future are hard, but one, I don't think that's true, and two, I inasmuch as we have evidence to date, I don't think that's been the track record.
So I think that how we use these things, what kinds of decisions we make, what kind of um you know considerations and you know uh margins of human welfare we seek to further, you know, I I think all those those judgments are are going to really matter.
And maybe a critique you could have leveled um at Progress Studies or Progress Studies styled thinking five years ago is these are all nice questions, but the world is on a kind of um foreordained escalator path to you know some kind of teleological outcome.
And I don't think the world feels that way today, or certainly it feels much less that way today than it did.
And so global affairs or something else.
No, I um I mean maybe somewhat global affairs, but uh the trifecta of um yeah, global affairs writ large.
Second, I think that aspirations and ideals are becoming contested more actively, and you know, there's an ambiguity these days in the US as to what the left and the right even stand for.
And you know, I guess we currently have one party endorsing tariffs and another party opposing them, but with the valences kind of shift uh flipped from what what one might have expected uh historically.
And then third, yeah, obviously, technology and first and foremost AI, but you know, in our industry, stable coins, the rise of China as the preeminent manufacturing power in many technologies of the future, like drones and robots and batteries and solar, you know, et cetera.
So yeah, in in many different ways, I feel like the future is is um, you know, uh Peter Schwartz um has this concept of you know the Schwartz window as the um the window of you know contemplatable futures uh in you know whatever number of years hence.
And I feel like that Schwartz window uh uh, you know, as of say 2005, as we contemplate the world of 2015, was fairly narrow and was correctly fairly narrow.
I think the world of 2015 did in fact unfold largely the way we would have expected in 2005.
And I feel like today in 2025, that window for 2035, like it feels extremely broad.
Um so um, yeah, I think the progress studies style questions are are more pressing.
So you were on the record um in saying that people should focus more on the question of why we don't see improvements in productivity numbers as information technology increases, and also as uh more people have started working on science and technology and more money has gone into it.
Um and what do the numbers look like now?
Do we see AI in the numbers?
There was a new paper published on this very recently, like the past couple of days.
Um that I've not had a chance to, I just cued it today to read.
Um I have at this moment only read the abstract.
Its claim uh is that one does not, in fact, observe uh productivity improvements stemming from um from uh from use of language models.
Um, I certainly can't uh what they're looking at.
They appear to be undertaking some kind of natural experiment looking at the individual level uh based on um based on uh intensity of LLN usage, but I certainly cannot endorse their methodological rigor, and upon understanding it better, you know, I might be, I might be either really impressed and find it very credible or horrified.
I don't know.
Um but that was just the finding I happen to stumble upon today.
I mean, look, overall, GDP growth in the US looks well, over the last two years, it's been somewhat better than we expected.
Um obviously it you know, we're we're speaking right now at a kind of volatile time.
We certainly don't see any evidence for exponential takeoff.
Um and if we you know, in as much as we thought that the encouraging GDP figures we have seen in the US for the last two years are attributable to some of these new technologies, I think you would also expect to see them in other countries, right?
Because these technologies are quasi-public good.
Anyone can, you know, um can uh can use these LLMs.
GDP growth outside of the US has has not been that encouraging.
We're we're not living in some you know massively accelerated period of economic growth for the world writ large.
Uh and so, you know, obviously it's it's early days, uh, but uh I think um, you know, I I think we're seeing that the diffusion of these technologies through the economy uh really takes time and uh in involves substantial complexity.
And maybe just last point on that is um I believe Jack Clark said in an interview with Tyler Cohen, uh one of the co-founders of Anthropic, and you know, Anthropic to some extent has well and Anthropic has always taken the concept of AGI and even ASI, I feel extremely seriously.
And you know, Dario speaks about this publicly, he um he he's written about it, you know, et cetera.
And again, Jack Clark, one of the co-founders, and he said that he expects uh AI to uh uh increase GDP growth by half percent a year uh and I thought that I mean I I interpret Jack as as really an optimist and you know half a point a year is in fact a lot of incremental GDP when compounded so I'm not saying that that's small but uh I think uh I think it's interesting that that was his figure.
Yes.
Do you think that uh with the form factor that AI is taking in the economy right now if we just kind of uh stretch the line forward do you think we're going to need new measures in economic productivity than we have right now.
So like assume real productivity goes up assume that AI keeps getting better it gets kind of deployed in the ways you would expect you think we'll need new measures or it should show up in the numbers.
No I don't think so like I think I don't think that GDP is perfect.
I think GDP can be improved but um but I in in any world where what we generally think of um as the economy is massively enhanced it'll show up in GDP I believe.
Yep.
When will we be able to program human biology?
I'm very excited about this.
At Arc, which is this biomedical research organization, which I was uh involved in uh founding we're working on uh training foundation models for biology using DNA and uh things like that.
Um we're working on a virtual cell um and uh and generally we're trying, I mean, a thing that I think is that I didn't appreciate until really spending more time in biology is we've never, like we humanity have never cured a complex disease.
So, you know, one ontology or schema or something of diseases would be, you know, you have infectious diseases, the flu, the cold, COVID, whatever, and you know, tuberculosis and you know, uh very uh diseases with high mortality rates.
Uh, then uh you have monogenic diseases where you know there's just sort of one genet mutation that is um responsible for the disease like Huntington's.
And then you've complex diseases.
Uh and the complex diseases are kind of the residual that you know are now left after we've cured most of the problematic infectious diseases, at least in the Western world.
Um, most cardiovascular disease, most cancers, most autoimmune disease, most neurodegenerative disease, et cetera.
For certain of these conditions, we have maybe treatments that help, like statins with cardiovascular disease, but for none of them can we really say that we've cured it, that we like understand the causal pathways and you know, meaningful detail, and that you know, it's it's just you know we can we can vaccinate against it or something.
Um I think this is our hypothesis, you know, could be wrong, is that this is in part because we don't have uh experimental and kind of maybe epistemic is too grandiose a word, but kind of epistemic technology that's up to the task, like the um the pliotropy of the genes in terms of all the different parts of the body and the systems and the mechanisms inside the cell, what they affect is so you know, there's so much combinatoric complexity there.
And then you know, the environment is such a vast and you know difficult to quantify thing that just it's it's it's really hard to understand for any of these conditions, you know, the etiology and the dynamics uh and so forth.
Okay, then over the last 10 ish years, I mean a bit longer, but a lot of the development has happened the last 10 years.
We've gotten um we've gotten three new classes of technology in biology.
Um for reading, we've gotten much better sequencing technology, single cell sequencing, the ability to sequence uh you know, single cell sequencing of RNA, um, and those improvements at the kind of think level, uh, we um you know, we've gotten neural networks and deep learning and transformers and you know everything, everything there, or I mean, they've existed for a long time, but we've gotten the recent improvements uh uh in them, uh, and the transformer in particular.
Uh, and then on the right side, we've seen obviously huge improvements in functional genomics and CRISPR and uh you know, bridge editing, which is a technology that kind of arc, but the ability to kind of make very specific directed perturbations in cells.
But if you put those together, you now have the ability to, again, at the kind of level of the individual cell, to uh read, think, and to write.
And this starts to really feel like a new kind of Turing loop and to have its own sort of completeness.
Um, and you know, we will see how much uh this can do against uh these complex diseases and whether sort of this systematic approach is up to um up to the task of um of shedding new light on their dynamics, but we are hopeful and excited.
If we here at Kersher and also others in the industry are successful in automating lots of programming as we know it today, and you know, replacing it with a form of software building that's much higher level and more productive, and it's much more just focused on defining what you would like the software to look like.
If we you know succeed in that, um who are you long?
You know, people talk about the designers and how you know this will be like a renaissance for them, but um, you know, are you long the the grad students?
I mean, there are lots of really, really amazing grad students who uh who are awesome and then uh maybe are less less skilled at making things happen on computers.
Um but who do you think you know is the most unexpected beneficiary of a world where both many more people can make things on computers, and then also, you know, especially if it's an evolution away from programming, um, you know, the people who are already making things on computers are much, much, much more productive.
I don't have a high confidence answer to that.
You know, there's all sorts of trite stock answers like real assets, especially constrained real assets.
Um, you know, maybe we should be allowing SF real estate uh or something because um, you know, it is uh it is one of the most beautiful cities in the world and will be enduringly so.
Maybe we should be allowing the inputs and the ingredients to these systems because you know demand uh for for them will go parabolic, and so maybe we should be long copper.
Um maybe we should be long um positional goods and celebrities and you know Taylor Swift's music catalog.
Yeah, there's a lot of I think compelling theories here, but part of what I think is interesting about this economic moment uh is the unpredictability and the contingency and kind of sensitivity to the precise assumptions in the technology trajectory itself.
Um and the shape that it takes in five or ten years or whatever, I think is going to do a lot to determine the answer to that.
And as I look backwards the last couple of years, I'm struck by how many predictions have held up you know reasonably poorly.
Uh uh, you know, even from people who are on the face of it, you know, extremely well informed.
Uh and so um I've asked a lot of people this question, and I have not heard any answers that are so compelling that I feel like I have conviction.
So we uh are very happy to be serving Stripe and your guys' mission.
Uh what would you like us to build?
Um how can we make Cursure better for you, either you, Patrick Halson, or you Stripe?
Well, you guys are already making Stripe better.
So um, you know, keep doing what you're doing is uh would would not be a bad outcome from our vantage point.
Um Cursor has uh today hundreds and soon thousands of extremely enthusiastic Stripe employees who are daily users uh of uh of Cursor, and they report that it's uh you know very significant productivity uh enhancement.
Um and so it's we'll wait for the economic numbers.
Uh well, you know, the economy is pretty big, um and these diffusions take time.
Yes, yes.
So you know, it it seems kind of greedy to uh to want more if you're already making, you know, Stripe spends more on RD and software creation than we spend on you know any single undertaking.
And so if you're making that process more efficient and more productive, then you know maybe it seems greedy to want anything more.
Um if I'm being selfish, okay, three things.
Correct.
The runtime characteristics and integration stuff that we just discussed, I think would be really valuable.
I think the the refactoring and the beautification stuff that again we also talked about, I think would be extremely helpful.
And and and I think really change our degrees of freedom.
As in if you could lower the cost of future changes to Stripe and improve the quality of the architecture.
And then third, we really care about um what we call at Stripe craft and beauty.
Uh, and we want our we want our software to be well designed and pleasant to use, and pleasant to use not only in the superficial, you know, um pixel sense, uh, but also in the deep it works very well sense, uh, and uh is something you can set up and you know largely forget about and just trust or forget about it in as much as you want to.
You know, there's obviously a concern with AI that it leads to the creation of more slop um and more kind of crappy things, but not more of the best things.
I don't know what it would be that Cursor would do to um ensure that the world is creating more of the best software and not just more software, but I think that's an interesting and important uh dimension.
So those would be my I mean besides all the obvious things to do, those would be three suggestions.
Amazing.
Thank you, Patrick.
All right, thank you for having me.
Yes.
Thanks for listening to this episode of the A16Z Podcast.
If you like this episode, be sure to like, comment, subscribe, leave us a rating or a review, and share it with your friends and family.
For more episodes, go to YouTube, Apple Podcasts, and Spotify.
Follow us on XAA16Z, and subscribe to our Substack at A16Z.substack.com.
Thanks again for listening, and I'll see you in the next episode.
This information is for educational purposes only and is not a recommendation to buy, hold, or sell any investment or financial product.
This podcast has been produced by a third party and may include pay promotional advertisements, other company references, and individuals unaffiliated with A16Z.
Such advertisements, companies and individuals are not endorsed by AH Capital Management LLC, A16Z, or any of its affiliates.
Information is from sources deemed reliable on the date of publication, but A16Z does not guarantee its accuracy.