Hello, Acquired listeners. Since we did our episode on Shopify six years ago, they have turned into a giant public company. They went public in 2015 for less than $1.5 billion — very cute by today's standards. They're worth nearly $200 billion now and doing about $10 billion a year in revenue. We've gotten to know Toby and the team pretty well since then, and Toby is an incredibly unique thinker. We were originally thinking, let's do an ACQ2 episode on Shopify's recent developments, sort of everything that's happened since, which ended up turning more into a philosophical conversation about the state of computing, how AI changes everything, and what's possible to build in the future because of AI.
Yeah. We also got into some of the more insane moments for Toby as a CEO over the past few years, like when the company almost became a meme stock, going from trading around a 20x revenue multiple before COVID up to 70x, then crashing back down after COVID and ZIRP. Today, though, they're nearly back at an all-time high, but this time because the business and the revenues are actually doing great, not because of a COVID and ZIRP-fueled bet on the future.
Yes. So, please enjoy our conversation with the founder and CEO of Shopify, Toby Lütke. Where have you spent a disproportionate amount of your mental cycles in the last month?
It is a privilege of a lifetime to be part of another platform shift. And man, like, it's just every single time I think I have my head wrapped around it, it's remarkable how quickly one normalizes to completely futuristic new things appearing. We all kind of like, yeah, I guess I can just paste this into ChatGPT and ask it questions about it. I'm still hung up on like, there's software that no one wrote and I have to interview it to figure out its capability. So my time is like really just interviewing these new models as they come out for their capabilities and just trying to figure out how to make them idealized, non-judgmental teachers for people as they keep getting better, and how to create structure around judging new models for their capabilities, figuring out where the edges are of what they can do right now, but also make it so that I can very quickly retest any new model coming against those edges to figure out if we've crested another.
Do you have your own personal test harnesses built to try to understand capabilities?
That's exactly it. This is sort of what I'm describing, but you're nailing it. I have, like, a Toby eval — it's literally a folder of prompts with expected and judged results, and I run it against every model. At some point I'm like, yes, I have these chats and I have these standard questions, and then I sort of make decisions based on this. I can write code that does this for me and run this against every model. And it's funny how your thinking evolves while engaging with these tools, right? This is not how I would have approached this before, but speaking with a lot of machine learning experts, this is how they have been doing this. These ideas of building evals, they're called, right? That's not a term to most people, and it's sort of a niche product of the machine learning world because it's essentially just batch jobs which just try the same thing over and over again and then judge the results.
Yeah, I know them as unit tests.
Exactly. So yeah, exactly. Unit tests were kind of a revolution in the early 2000s, end of the '90s. I remember when people sort of proposed automated unit tests as a thing. And I talked with you before about my wonderful Meister where I did my apprenticeship at Siemens, and he always said, like, you have two years to develop software after you start, and afterwards it's like someone poured cement into it, right? Like you're never going to change it again. And that was true in the '90s. That's how people thought about software.
Right. Software engineering was like civil engineering. It's like you build the bridge, the bridge is done.
Exactly. There was an end date to it. And then if you wanted to evolve it, you rewrote it, right? Like, you know, Windows was a rewrite every major version. Well, not actually — they took some with it, but it's mostly start from scratch. Software evolves forever. I mean, I still can't really explain why bit rot exists. Like, why does a great piece of software get really, really bad like a decade later, right? It hasn't changed. We have changed, right? Our expectations have changed, maybe.
As soon as the new iOS is announced, the old iOS feels like garbage, right?
And it's not that your phone got slower, and it's not that — it's those UI paradigms sort of expired.
And I see this really vividly. This is a good example because, you know, I do all these product reviews, and I'm also like — I mean, it's such a bad thing to do, but I just, I physically can't restrain myself from hitting that update to latest beta button when it appears anywhere, right? Like it's just — I cannot — it's like the siren's call. There's no amount of...
Do you run beta software on your phone?
First version. Like I just can't not. And I have a single phone, and I'm like, I'm just going to — I find all the bugs that are waiting for me as interesting constraints that I can play with, you know, just like my app.
You have like Craig Federighi on speed dial and you're like sending him screenshots of...
Nope. No, no externalities. I'm just like, I own it. I own my own problems. I'm going to the first developer beta, and if five of my apps that I need every day don't work, I use that as a delightful opportunity to learn about new apps. Right. So it's just like...
You're a public company CEO and sometimes your phone is just randomly crashing.
But that actually is avoidable, but not for you.
Yeah. It's like, whatever the consequences on this.
This is why Toby is special.
So anyway, I just live in — you know, as much as iOS is just updating to a new UI paradigm called Liquid Glass, which was very much maligned right off the bat by people, and somewhat rightfully so in some instances. But to be fair, like, Apple has an uncanny ability to introduce things that afterwards seem that was clearly the right thing, and actually it's not even clear that they are better. It's just very clear that they make everything else worse. It's induced bit rot on everyone else in a way. So I've been living with this changed UI for a while, and I'm in these product reviews with my own teams, and...
you know, we're looking at designs which are clearly sort of good designs in the context of the world that we are now leaving in a way, and like, it's incredible what it does. Because I'm like, hey designers, I know you — I need you on the betas. You kind of need to design for the device as it will be experienced by people in the future, right? And so I find that is just like the biggest job that, you know, I personally can do for the company, and then my company can do for my customers, which is like live in everyone else's relative future. You don't even need to predict the future that well if you just live in everyone else's relative future, which really is just a couple of high-conviction, high-courage update-to-beta clicks.
It's not that hard to live in the future. You just do.
Yeah. Yeah, exactly. You just go experience, you know, develop taste for it and figure out how to make your own decisions. And this is, of course, extraordinary — making, you know, Sidekick inside of Shopify and like having the AI assistant, which is now like massive.
Yeah. Okay, so yeah, tell us about that. And like, back to bit rot. I mean, I was going to ask before — a lot of code is now not written by humans, and it's trivial to just be like, yeah, update this, make this better. What's going on in Shopify?
I mean, exactly, like what an incredible change. Code used to be by far the most expensive thing to produce, right? I think really, really good code for load-bearing things is still exactly like this, but around the margins there's a lot of the kind of code that glues together inspired infrastructure can now just be treated as like a first draft, right, for something. And I think that's actually extremely healthy. Like, this is a property of almost every great piece of software. There's a very clear split somewhere. There's a layer — like, the Linux kernel needs to be developed in one way, and then the Linux operating system just uses the kernel's infrastructure and pretends like every computer is the same, because you can — it creates a pretend world that's easier to reason about, so that then the software can be implemented.
Abstraction. It's a beautiful thing, yeah.
Abstraction, but inspired abstraction. Abstractions can also be bad because, like, abstraction is just another word for pretension. It pretends like — everything that abstracts something just pretends it's different from how it actually is. And so abstractions can be incredibly powerful, and there are some one could name which have really stood the test of time. But a lot of abstractions are lossy abstractions. They actually pretend the system is something — is a lot less powerful than what it actually is, which really, really matters if you evolve things on top. And then you end up on this like tower of abstractions. You sit on top, you want something for which you would actually have to go down here and connect. And then you have this sort of spiderweb of cables left and right. And it's just like, it's all spaghetti in the end.
This was like the world of — to make it concrete for listeners — the write once, run anywhere mobile idea, where, oh, we'll just cross-compile to all of the different phones back in 2012 or so. PhoneGap or...
Yeah. And the abstractions ended up being too lossy, where you didn't actually have access to the...
capability. The HTML5 debacle of like, oh great, we'll just write for HTML5 and it'll run like a native app, and you know, like, no.
Yeah. And, you know, very, very good software companies believed that to be the way, right? Like, hybrid applications were all the rage in the beginning, and remember the phones were really, really slow back then. The phones are insanely quick, and even today hybrid apps are not that great. So, you know, Facebook famously was all in on this idea, and Xamarin is a good example. So there's a lot of these kind of picking the wrong abstractions early constrains what you can do on top, and very often when the abstraction is wrong, it means you can't actually reach great software anymore. You actually pretended the world is simpler, but you don't actually allow all the most load-bearing pieces into this pretended world that are at some point required to making the software just really, really good. And so, you know, again, I'm a tool maker, right? Let's keep beating on this Xamarin thing, which no one's heard of, but like, just I think we've described it enough so that people might have an idea for what it might be. Just imagine you use very bad software, and the team that started the project decided early to Xamarin, and then they never had a chance of making this as good as all the other things you really appreciate on your phone due to this decision, no matter how good they are.
It's like building your house on top of your kitchen table instead of on the foundation of your house.
That's a very good analogy. There are early decisions that will constrain how good you can make something. In our parlance at Shopify, that means certain ideas, certain abstractions, certain choices bring the ceiling down. As in, like, if you have a scale from 1 to 10 of how good something can be — 10 is world-class masterpiece and 1 is like horrible — what you hope to do is that you use abstractions to make it easier to reach a good number. It brings the floor up. It's at least this good, right? That's what most abstractions do. And from a business value, a lot of businesses choose tools to make it at least this good because that is sort of a passing grade, right? Like, it clears the hurdle.
Why would any customer use Shopify? It's like, well, okay, I could go build my own e-commerce checkout payment system.
My store is going to be at least this good even if I do zero customization.
And well, if you set out to do your own from scratch, and this is not your core business, what you get back might actually be like completely underperforming what you wanted, right? Like you can't even use it.
Yes, very, very likely. And it becomes one of those ever — like death-match projects that you just pour resources in, and it takes incredibly good people to rescue a bad project, which is like hard to even convince good people to do this.
So you end up in this world where — you know, so bring the floor up is what people hope for. What a lot of people miss is very often the same systems bring the ceiling down. And so what you are trying to do as a tool maker, as an infrastructure builder, as a product maker, is you want to make a tool that brings the floor up significantly but doesn't constrain the ceiling. And that is extremely hard. I think it takes — sometimes it takes — it might not take decades to build it, but it does take decades to build the mental picture to really understand all the aspects of a problem domain in such a way that you know you have to look around five corners to know, am I going to restrict someone's potential with making this sort of choice that's here ahead of me? Modern operating systems are wonderful; they don't usually lower the ceiling. Like, you can build any amazing piece of software on top of, you know, Windows, Linux, and OS X. Now, we figured out how to do this because we have been building operating systems for a very long time, and there were clearly attempts that were wrong.
You know, there's many such spaces, e-commerce software being one of them. Like, we're just like, yes, we want to build something that works like this. However, it also matters how likely it is — like, what is the distribution inside of the users of the system of people who hit 10 out of 10 versus, like, if your floor is 7 out of 10, which is really good, is everyone clustered there and one or two people manage to take it to the starry heights? That's also bad. And again, this is why I think AI is so exciting, because I think its principal ability is really to help massively shift the sort of scatter plot of where people end up as compared to their own vision. I want people, if they have — like, whatever their vision is unencumbered is a 10 of 10 on the scale, I want them to hit it. And I just — like, if they don't, there needs to be good reasons, and ideally their own — like, the choices they made, maybe they made a mistake, maybe their product is not good.
Not that the platform constrained them into not being able to make something great.
Computers shouldn't have to constrain people. It's like, this is not the computer's role. The computer is to get people to build things that they could not possibly imagine would have been solved otherwise. It's like, this is the sorting — this is the ordering — like, computers are there to make humans better, never to constrain what humans can do.
What does that look like from a product? I could imagine, but what sort of products are you building that take advantage of that?
Usually when it is a 10 out of 10, very often people are working with, you know, experts, some people who really dedicate themselves to really, really understanding all the ins and outs of a really complex piece of software. In many cases, that's the only way we could get there, although I think the role of software is to minimize the need for this kind of thing. And we want to make it approachable so that people can sort of, again, accomplish the initial goal: get online, have a checkout, be able to sell the product that they spent a lot of time with. And if they have additional time to invest, be able to — that we call this hill climbing, right? So now with like, you have an AI, you can ask — in fact, we are going to increasingly ask you what your goals are, right? Or at least we either try to find ways to deduct that from what people are doing with the software, or just straight up ask. My theory is that people are very happy to tell us what their goals are, and a lot of software just doesn't do it because it can't action on it, right? Maybe it has a — we could put a to-do list there and help you with record-keeping, but it can't be — we can't do anything about a to-do list before AI. And now again with these agentic flows, we increasingly can. So yeah, like, I think the way to get to 10 out of 10 is just like really sit on the same side of the table. Like, we have an intake; you naturally tell us what you're trying to accomplish, and then we can show you, here's what a prototype would look like of this idea that you have, and just like also implement just these obscure best practices, right? I've spoken with one of our customers. They're selling rugs and um — rocks. Sorry, rugs.
The woolly stuff on the floor.
Yeah. Yeah. Like beautiful carpets, really.
I should use the word carpets because that works better with my accent. We were just not making progress with the Europe strategy and kind of about to pull out, and then they used some tools that we wrote out, not really knowing what it would be used for, to just change the pictures — like literally just...
Upgrade, update the images of all their products away from Malibu beach houses to Parisian apartments.
Like, this was not a new photo shoot. This was actually just like, do this for me.
And their sales tripled because four years ago, three years ago, this would actually technically not have been possible to do. Now, it's about the most boring thing you can imagine, right? Like, this is how much the world has changed.
I mean, at a minimum, it would have been like, okay, let's book a photo shoot in Paris and let's like, you know...
Like allocate some capital, organization — like it needs someone's attention, probably someone...
product there, let's lay out the...
And so now this is just done, and like this has proven out. And again, I'm sure they don't run these AI-changed product pictures now in production. I'm sure they did the shoot afterwards. But they also discovered something over their business which would have been entirely obscured to them. But this is an applicable lesson across the entire platform and something we can absolutely help small businesses that don't have the resources to go reshoot all their pictures.
Well, and shoot — like, the stuff is now so good that like, yeah, you probably don't need to do the shoot, right? Like, even if you are the biggest business in the world, like, the difference is unnoticeable.
Yeah. And, you know, like when you talk about they're the biggest business in the world, they already operate like this. Like, IKEA does not — that is like, the IKEA catalog has no photos in it. It's all renders. They just like, they create really, really good versions of...
Oh, car commercials are not videos of cars driving. It's renders of the cars.
Yeah. And you only know because the professional drives on closed roads has disappeared, which was required before, right? Like, so that just doesn't — they don't need to show this anymore because...
Like, there's no one actually driving. That's the giveaway.
So I think the way this really like hits for me the most is, again, I grew up on like all the sci-fi books that everyone my generation really grew up on, or many, many did anyway. The sort of like very, very optimistic but also far-future, like Asimov and so on.
What's your favorite, by the way?
Um, my favorite — that's actually, I think Snow Crash is my favorite. It's just — and Neuromancer is — I recently reread Neuromancer, it's a really, really good book.
In so many cases, like especially Heinlein did this, like there's usually like a catch-up chapter. Here's how we get to the starting point of a book. And one of the events is always in 20... 2110, we pass the Turing test, and like there are celebrations in the streets, and then something turns sour usually right after, right? Like to set the stage. And you know what no sci-fi author dared to predict is that the Turing test just passes by.
We just blow by and nobody noticed. There was no New York Times front-page article that we've passed the Turing test. There's no — it's not even like the tech community was exploding about it. We all were just like, oh yeah, we're in the generative AI era now.
If you just like — clearly GPT-3.5, which is like, it's significantly worse than something you can run on your phone right now locally, like, passed a Turing test with flying colors. And it's — this is crazy. Like, this is the craziest thing about the actual world if you think about it. It's interesting that — if I — maybe I'll take back my previous comment. What was actually happening when we passed the Turing test and no one cared was we were all so focused on, oh, is this the path to AGI.
I kind of wonder if AGI will be the same thing, where we do actually achieve AGI, but we'll all actually be talking about some other future milestone that we're wondering about.
I think humans are incredibly good at moving a goalpost.
Yeah. Yeah, it's like there's some kind of expectation-related version of a hedonistic treadmill that we sometimes talk about, which is like just like it moves forward.
It's that we adjust — we adjust to our circumstances incredibly quickly. And so, like, I mean, almost everyone listening lives their own dream life from many years ago, right? Like people accomplish their goals, and we kind of never stopped to think about it. And I think this is like, it's really, really hard for us to judge our own environment and what's going on around us. So I think on a societal level this even seems to replicate. And again, the Turing test was one of the best takes we've had, and people can like — now people criticize it because it's not good enough, because I suppose it wasn't. But like, it was still the best test that people generally referred to about figuring out if something is intelligent. And there was no parades, no front-page news, no ticker-tape parades, you know, just like, we should have...
Yeah. It does seem though that there are few CEOs, of which you are one, that noticed this, and like you're behaving — you started this whole conversation by saying you have the privilege now of living through another platform shift. I don't think the world has quite realized yet that it's a platform shift.
Yeah. I wish people would, and I wish people would like sort of seep in the remarkableness of it. Like it's just like, I don't like — I don't know if you've ever done this. Like, I mean, I love the VR goggles, Meta Quest and so on. I use it more than most because I use it for like racing simulator preparations and so on. It's an incredible device, which also like VR was another one of those things which we thought we would get at some point, and then it came, and like it's kind of like there — people use it sometimes for like their fitness workouts or whatever. It's just like it just folded into the ambient world, at least for some.
We take it wildly for granted that that technology is cheap and plentiful today.
Yeah, exactly. $299 or whatever for like a remarkable device, right? You know, augmented reality, like also something people talk about. You can buy them next week. Like it's just like an event next week after we film this, and then afterwards you can just purchase video glasses.
And so that's all around us. And, you know, one thing that really struck me is when you play with Meta Quest and the VR goggles, there's one game which I really highly recommend people. It's called Job Simulator. Have you ever encountered Job Simulator? Okay, you have to try it out. Job Simulator is like this sort of satirical take of like, it pretends to be in a far future, and there's museums where people go to figure out what a job was.
And it starts you on a cubicle, and you know, like you have to turn on the computer, and to turn it on you actually have to plug a power cable into the wall, right? And it's like otherwise it doesn't work, and you just think it's broken. And it's so wonderful because like it plays on the future historians and robots kind of getting details wrong. I actually think it's like profound and actually brilliant commentary on the world. But I think what idea this nails is that I think we will have VR simulations of these times. I think we do — like I think this is the golden age of humanity right now. It's the — I think it's the end of the beginning, and so I think we will — I think people in a thousand years will study these times the way we are studying Cicero's times of the end of the Roman Republic, and as a significant part of history, and because it's kind of where we figured ourselves out in a lot of ways. And yes, there's like people look at — you know, you look at the news, there's lots of — like there's a lot of current affairs and events. I think they will fade out behind simplistic terms. Like obviously there was like strife in any kind of time doing great change, but it's so unmitigatedly good. It's like there's so many things we are bringing toward — these gifts of incredible technology and progress that are coming rapidly, that are all kind of the types of things that are going to make — that give individuals with vision so much more agency, which is like the thing that like it's always at the top of a funnel of anything like that people truly care about. Like it's everything that people really love, all the products, all the experiences are other people — it's like are things that people fought for. It was a passion — the results and the residue in a museum of — it's a museum of other people's passion projects, as John Carmack puts it.
Such a good quote. I think that is probably the best tweet on the entire X platform. Like, it's so good.
We'll link to it for listeners who haven't seen it. I'm sitting here thinking if I could kind of describe your life: you started a company in the previous technology era, the web era, and then, you know, took it into the mobile era. You have a family. You have hobbies. You've talked about racing. And the entire world is going through what I think you think is the most transformative change in the last hundreds of years.
What should you do with your current life in reaction to that, knowing that you live at one of the most like crucible moments in history when we look back at it?
You specifically, or is my...
Maybe you specifically, and then like generalize it. What should people do if they believe this?
It sounds like you're interviewing models every day.
Yeah. Like the ones with floating points in them, just to be clear, like, because...
Oh, yeah. Yeah. Yeah. Yeah. You were interviewing foundational...
models. Exactly. Yeah. Yeah. It's an important distinction.
Good distinction. Yes.
So yeah, like I mean, here's what I'm doing. I am very much — this comes fairly natural to me, and I think it's a little bit a job for me to transmit some of the awe I feel about the times, right? Like I think that what we think of history currently is wrong. And again, the people in a thousand years from now who are also going to play 2020 Simulator...
I'm never going to get this image out of my head. I love it.
They will think about these decades and this time differently because what they will do in due time is reintegrate technology into our own history. Our own self-history is one of sometimes decline, right? Like we're talking about the Golden Gate Bridge being built in whatever, two, three years. And we are rightfully saying we can't do this anymore. We are looking at great works and infrastructure projects and like the great generation, the wars and these kind of things, and look at this and say, well, we can't experience a world right now where people build Bethlehem Steel or people build Henry Ford's factories or — they're just like physical manifestations of the incredible power of, you know, human beings coming together and creating things is sort of manifest for us to see in front of our eyes. And so therefore people judge that we no longer do this. However, we are actually absolutely doing this. It's just the great works are kernels of operating systems, the browser, lots of the software platforms, the phones, the incredible advances in silicon lithography, and the models that we create and train, and the data centers that support them. A lot of it is actually physical building — laying the energy infrastructure and the power generation, the power transmission, the buildings themselves, the heat management, the cooling.
Yeah, I mean, less than three years ago ChatGPT didn't exist.
But it is physically not as imposing. So like, a data center — I love the blinking lights down the aisle, but man, if you have been to what a refinery looks like or a natural gas liquefaction plant, it just hits you on a level that is — it requires zero analogy to understand how incredible of an achievement each of these things is. Like the bridges we drive over are like — they're just like, man, this is humans being their best selves, creating, you know, shaping things, not just like — and very often not stopping at utility, but also going for just beauty by itself. Like the ornamentation is there because we're not just creating things — so like there's no ornamentation on the outside of a data center that eventually will train superintelligence, if that were to...
In fact, we hide it. We don't put logos on it. We don't even — we try not to have it on maps.
Yes, exactly. It's going to be as obscured as you can make it. So I think this is one of those things in a world that sends people astray. So back to what I like — I want — I love to transmit this because I have cultivated, like every technologist has, an appreciation for the abstract in the same way that people naturally come pre-loaded with an appreciation for the concrete. Like I can look at a piece of software and appreciate it in the same way how I can look at the Pantheon. And unfortunately that is not something that comes easy. In fact, that requires — and is also probably not worth the time that it takes. You really have to rewire your brain in a very significant way. People who build technology...
Any human can look at the Pantheon and be like...
But not any human can look at a piece of software and say, wow.
Exactly. And we even have thousands of years of understanding about why the Pantheon is — like we can start with math to tell you why the ratios are correct. We are now — we're close in the abstract realm. We can't describe why software is good, truly. We like — it's taste. It's some people know it when they see it. It's something you cultivate. And so it's just abstract on top of abstract. If you build a tech company, you kind of have to do what I'm talking about to yourself because again, I see this with my customers. The reason why my customers love building physical stores, retail stores, is because they can stand in them. They can see them. They can judge every piece by touching it. They can watch their customers come in and look at things. They learn what the first thing is they look at, and how they handle it. Do they pick it up? Is it clothing? Does it feel right? And do they put it down? Is that something for them? And they will rearrange the store every day. They hill-climb inside of it because they can judge everything. It's concrete.
With an online store, it's not like — we try really hard to tell you sort of life, how many people there are, and tell you stats about it. It's always abstract. It's numbers. It's not real. It doesn't hit us. Even if you cultivate an appreciation for looking at the numbers and looking at stats and become data-driven, if you would do an analysis of how a day went in your mind while being in an fMRI machine, it would be a very small — it would be neocortex only that would light up. If you would think about your day in the shop talking with people, it's the entire brain lighting up. It's just we don't — we are predisposed to appreciate the concrete. Of course, of course we are.
We have this exact thing with Acquired. We release an episode, numbers instantly hit hundreds of thousands, and then over the course of months hit millions, and like I look at the dashboard and I feel nothing.
And then David and I do — then we do an in-person thing at a tiny, tiny fraction of those people, and it's like...
And it's like the whole brain lights up.
Like emotions I've never experienced before. Memory is created, nostalgia is created, relationship is created. You're on cloud nine for weeks, and you're like, that was nothing compared to what was in the dashboard. But...
Yes, it's so true. And it's like I don't know how to solve it, and I think it might be not solvable because it kind of depends on — like it might honestly depend on our hardware at some point. Like it's just like we can convince our incredibly malleable frontal cortex of anything, and it can learn to appreciate the abstract, but the lizard brain is not convinced.
Is the age of VR the answer? Like is the next stage after charts and graphs and tables to like create full abstract experiences — millions of people? Yeah. Like this is how many people, if you were standing in your store right now, this is what it would look like.
How busy your Shopify store would be.
I actually think it will get better, and then not nearly the same. I just don't think it's reachable. I don't — I do not think you can replace it, and I think that's a real constraint, and we just have to like...
That's a great point though that you were saying a minute ago about like part of your job and part of the industry's job is to convey that, yes, and describe that to the world.
Okay. So what do other people — what should other people do if they agree with you that this is this moment of great historical significance?
The actual thing I do is I found myself a thing I really, really care about and try to bridge from a future to people I like, I care about. It's like I build Shopify. I try to understand what the future looks like in any — with any trick in the book. And I want to — I want to help with what's going on. And having — finding a task that you care about inside of the change is, I think, the best position anyone can take. It doesn't have to be a software company. Like so much of music is going to change because we can generate quite extraordinary music...
from neurons directly. That's seen as a threat, and similar to how words can be generated with AI in an even more obvious way. It's an integration to help everyone figure out what is the clearly human part. Because here's a funny thing: there's a couple of worlds where AI has already massively... you know, obviously math, like the calculator has been there for a long time, and the best calculations and fastest calculations were clearly not done by humans ever since the calculator came around, right? So chess is such a good example of this. Like the best game of chess in every year of the last 20 years has been played by machine against machine.
And no one gives a... no one watches those matches. No one looks at them, no one studies them. Everyone knows who Magnus Carlsen is, right? Like we actually care not about chess. We care about humans playing chess. And this is the one thing people really do get very wrong here. We need to figure out how to treat the new capabilities as instruments that we can play, and then explore that space and teach others. And I think that's going to be the most fulfilling and fun position because it allows you to steep yourself in the things that were previously impossible while they are still hard, and figure them out, which is always one of the most fun things. And then teach others. I can highly recommend it. It is incredibly fulfilling.
All right, listeners. This is a great time to thank friend of the show, Plaid.
Yes, odds are you've used Plaid without even realizing it. If you've ever linked your bank account to apps like Robinhood, Venmo, or Chime, you're one of the millions of people, like one in every two Americans, who've already used Plaid. I feel like I've grown up in the tech industry alongside Plaid. So many modern experiences are powered by them. And at its core, Plaid isn't just about making it easier to connect to your bank. It ends up being the backbone of thousands of companies building faster, safer, and more seamless financial experiences. Whether it's reducing fraud, speeding up onboarding, or turning old-school banking processes into something that feels instant and effortless, Plaid is making it happen.
So last year, Plaid rolled out some powerful tools. Think cash flow data for better credit decisions, anti-fraud tech with AI, and analytics for bank payments. And this year, they're leveling up again with major updates across all three of those products. And they're even helping businesses manage things like direct billing for your subscription. So the bottom line is Plaid is making it easier for companies to build smarter, safer, and more personalized financial experiences that just work. If you're building financial tools or infrastructure, Plaid's data analytics can give you a serious edge. Whether it's fighting fraud, underwriting smarter, or managing payments more efficiently.
And if you want to learn more about how Plaid created one of the biggest networks in financial services today, listen to our recent ACQ2 episode with Plaid's founder and CEO, Zach Perret. Our thanks to Plaid.
It's funny. We were talking to another CEO yesterday who is a huge football fan, American football fan, and last year wrote an AI agent to do his fantasy draft for him and scour the whole internet, take into account all the likelihoods to be injured.
Yeah. Incorporate stuff that people generally... like first he was like, what do people not incorporate in their fantasy drafts?
Well, first it was what do people incorporate but they actually don't know all the information. Go great, go ingest all that information. Now, what are the factors that I think people aren't thinking about?
Yeah, yeah, yeah. And so like one of those was like propensity to be injured.
Yeah, that should be really important. Most people don't account for that. Great. Write an AI agent that does that.
Or if they do, they have incomplete information and you could go get more.
Yeah. So wrote it, crushed it, you know, like won the league, you know, all that. Like this year everybody in his league is using an AI agent, you know.
Yes, yes, yes. I mean, so then it like, you know, well, shouldn't NFL head coaches do that too? Like...
One thing on this, this like, well, great, now everybody's doing it. This thing I've been thinking about is when you really see the full capabilities of the latest thing in AI and you get a demo of something or you get to use something for the first time that you're like, 'This isn't how I was using it before and this is not how other people use AI.' You think, 'Whoa, I have an edge.' Yeah. I actually can do something that people didn't think was possible. It's become clear to me that even 5 years from now, people who are successful, it's not going to be people that are making things in an unassisted way. Everything is going to be assisted. And so the best people in the world at anything necessarily will be the people who best know how to use these tools.
Do you agree that like this is the new bar to be a human out in the world is understand how to best be a user of AI tools?
Yes. But like I think for a bigger reason than just the AI tool. I think in terms of stability, which even our industry had, like the things were just like not changing that much.
Exactly. So that time, like I think that entire 2010s decade in tech was, I mean obviously people will rattle off a couple of things that changed, but like we had the mobile phone, we had the platform, and we just sort of scaled the ideas that all come from the previous decade. There it's pretty obscure, but then when things change faster it becomes very obvious. And there's two groups of people: people who fundamentally have fallen in love with problems and people who fundamentally have fallen in love with solutions. And I think that sometimes, like of course the people who really fall in love with solutions obviously then grieve if those solutions become un-ideal. That makes perfect sense and I'm sure it sucks. It's happened to me, too. I love computer programming, like low-level stuff, for just like for no reason. I will never implement most algorithms again because I can just like when I type the function name, there's great text already and I can hit tab and then it's implemented. And sometimes better than probably my naive first implementation would have been anyway. So like there's some artistry lost and some skills that I no longer need. But again the people who love to solve problems are the ones who are just so enabled right now. And I think everyone starts there. No one falls in love with solutions at the beginning of their careers. It's like you can't. And so I think people need to remember how they got to where they are right now and realize that this is just a bit of a calcification of who they were and they need to take a step back and realize probably life was more fun when I was walking around looking for how I will do my craft, right? Like how I will do my thing. Like this even gets so far, like even in the corporate world, like job descriptions, like job titles, we've gotten incredibly... we are almost pushing people into this, like into falling in love with solutions. Like you know, you're a sales operations expert, right? Like what is that? That's like a title for a particular way how companies end up solving a particular operational problem they had at some point. What you actually are is like you're an amazing human being that got hired into a company and the company wants you to add as much of what you've got to the mission, right? And like clearly you can zoom out. Everyone does this when going out with friends and venting, right? Like just like you talk about how the place could be better. Well, forget about the title. Like let's figure out like how can you contribute. And everyone who's like really immersing themselves in these tools. I sent a memo to my company and we set the expectation that we require that people reflexively reach for AI now, and we require it because it's unfair not to, because the people who do are otherwise going to be people who are all the best careers to themselves, right? And I think it's fair to tell everyone that that's what it takes.
To be clear for everyone who didn't see or forgot the memo from like just a couple months ago, that feels like a few years ago. You're now required in Shopify to use AI as your first pass at everything, right?
Which was controversial when you sent it and now seems obvious.
Yeah, it's like, yeah, exactly. I didn't think I was sending a terribly controversial message to begin with, but then someone told me that it's like being leaked in bad faith, I think. So and then I just like posted it. So that I think worked quite well at least there. I was surprised how many people cared about that. And because it says a couple of things. Don't, if you at the beginning of a project use AI to prototype it, and then if AI is really bad at prototyping it, make fun of AI. I just want... it's not about what you make with it. It's that you did it because it might surprise you. And also, earlier in the conversation we talked about evals. If the AI is bad at the thing that you asked it to do, now you have an eval for the next AI that come coming out. So you're building your quiver of tests.
What are some of your evals? Like you mentioned you have this like folder of instructions. What are some of your tests?
A lot of this is like things where you know, like from coding for instance, where it just like misunderstands a certain prompt which it had enough context. I like the term context engineering because I think the fundamental skill of using AI well is to be able to state a problem with enough context in such a way that without any additional pieces of information the task is plausibly solvable. Right? That's actually a really tricky thing. And now with a lot of agents, the agents end up being pretty good at figuring out what's missing and go get it themselves. But I think that's... I mean the crazy thing about AI is like it can do this. I think the skill people should build is trying to not require it to do it. Why is this a useful skill to build? First of all, it makes it much cheaper to use AI, makes much higher likelihood of success, but this is really funny. I think it makes you better in so many other ways that has nothing to do with AI. I write much better emails since I became a good context engineer. I just like I now realize how much is unsaid in instructions and why so much of what people describe in companies as politics is actually bad context engineering. It's really funny that to like because I so often end up doing the what is the root note of a problem here that we all have, right? And you realize after digging one, two layers down, there's just a fundamental disagreement on an assumption about how to do like what does good look like for this area.
David and I have just finally started taking some things that we previously thought, oh that's part of the secret sauce I've acquired and who knows why my neurons fire that way. It just comes out and it's part of the magic. We started actually writing down instructions so that AI can do things that we previously thought were like part of our magic. And it's less magical than you think. We've just never had to actually codify how we do the job before. And I think that's probably a lot of people. It's a failure of introspection of what actually makes the work that you make good. And once you actually write down your process and all that and have an evaluation framework for what makes it good, it actually is quite executable by AI.
100%. And you're... I don't know how far you're down this journey now, but like there's an entire world that opens itself up to you when once you start doing this because like I have this too and we have this at Shopify. We use Anthropic's term for... we call these things constitutions, which is maybe a little bit too grand, but like figuring out a document that is actually like sort of all the things that are not platitudes, where like all the things where someone else, like some other company would plausibly take the other side of. But those are the real choices, right? Like if no one would take the other side it's a platitude. But teamwork is clearly something every company wants, that usually the company values end up being just platitudes. It's very funny. So you end up like creating a list of these kind of things that you uniquely would pick a side on.
The trade-offs that you're choosing.
Yeah, exactly. What makes... you know, basically like write down your uniqueness and then you get in an absolutely fascinating world because first of all you can use it. You can hold up to your book and just say, okay, well me here with this and these are things you need to know in your again context engineered window. But it gets way more interesting when you use your favorite episodes' transcripts. You add this and say what's missing from a document.
And you get suggestions and you're like, could we have made Costco better in some way?
Yeah. Right. It's scary, right? Like yeah. And so what you will realize is this sort of... this is another hill climbing exercise because this constitution will get much much longer and you actually will collaborate with AI on it. Because these things are incredibly valuable. They are, you know, just like... you know, anyway there's so many... this gets too much into the weeds but like inside of Shopify there's so many things that can be done like this. We have product principles which I wrote at some point and have since been hill climbing with all in a loop like this. And any project in our internal systems now like when it goes into review time is like we can run against the product principles and also the specific principles of this area. And it's not like some control mechanism. It's in fact it's honestly just like here is where interesting discussion should happen. It guides the conversation because often what we find is the product principles are wrong or imprecise or stated too broadly with not enough provisos. And because you want to write for a large audience in generalities, but like when you do work with AI you can actually say, okay, well actually let's write all statements next to the convictions, right? Because...
It's interesting, you need deterministic instructions for a non-deterministic machine.
Right, because yeah, like and they are of course non-deterministically interpreted in the end. But this is also the beauty of it because sometimes it tests you. Sometimes it says, 'Hey, this is contradictory and it isn't.' And then you have to have a conversation with it about it and like you realize, you know what, that mistake in the way you read that you made probably happens non-zero times with humans too, right? So finding a more precise version of stating the thing that you're trying to go for or better examples is also really valuable. All this to say increasingly I find that, and this is why I'm saying you have to interview these things. Like one thing that I'm really noticing with the latest series of models specifically, o3 thinking and Pro, is that, you know, I feel like I'm pretty good at asking questions and I can take these conversations with AI into very very very interesting places. I have started now asking the AI what my next question should be and I get better ones very often and that's pretty impressive because again it wants to predict and they have no memory very often with all your private conversations so they like... you know, I probably when I'm at my absolute best I ask better questions than when I'm at my average, and so it can execute against an idealized version of myself. And it's really really fascinating and really valuable. I have run a keyboard logger and sort of what is the active window and take a screenshot of the window like of my machine every 10 minutes and archive this for going on 15 years now. Like it's something I've always been running and just for my own purposes.
You have an archive of your daily digital activity of the last 15 years.
What a gift. Like nobody has that.
I would never imagine this ended up being as valuable as it is.
It's like how Toby became Toby. How Shopify became Shopify.
It's basically that. And now a lot of my weekend projects are actually just like, you know, running through all of this and putting it together because of course it's dispersed keyboards with lots of noise in it and...
Different formats over time.
Yeah. And like every once in a while I use Vim at which point there's a lot of what's in the keyboard buffer is complete nonsense. Like but the nice thing is with enough AI tools you can turn this into pretty clean timelines.
Right, without you having to write a script that's like, okay, every time I'm hitting, you know, Q, disregard.
Yeah. And of course the format changed a million times and has missing bits and pieces. And then I have my calendar which I can... which I'm now adding to it over the times and a lot of emails, like just headers of who I converse with and so on.
So you have that in a repository and then you can just like whenever you say like whatever AI model you're using, like hey I need you to make something like Toby, just look at this and then like do it like this.
Yeah exactly. So like it's... I mean the actual format was I just always turned it into text files which are like the most lasting bits of artificial... like it's the best format in the world because it lasts forever. And now I've been working on doing more and more work on just like analyzing it and cleaning it and adding more things because the value of this is getting incredibly high. I've also been like always doing like my notes, I kept them and so like their times and so on. So I have a lot of digital artifacts about myself which you know again...
What prompted you to start doing that?
