Diving Deep with Max Levchin

Thank you for having me.

I think that's actually a fairly common thing in the world of startups and innovations and I think it's maybe more than anything a difference between being a futurist and a product designer. And a lot of us product designers think that seeing the future is hard and that's our job. And that's certainly what I was doing in the PayPal days. And yeah, it's actually the opposite. You can pretty readily see the future just by reading enough science fiction books. Like if you crack open anything from turn of the 20th century like Jules Verne, you promptly see like, oh well, he thought of submarines and all these amazing technologies that we now take totally for granted. And I think that trains us as we grow up to be product designers to think all I need to do is look into my mind's eye and the time domain doesn't matter. Like, you know, if Jules could see the submarines because they were just being talked about then and poof, we now have nuclear power submarines. What I need to do is see what the future looks like. And you know, we saw Palm Pilots. I saw the first Palm Pilot was the first real handheld device that would fit into your pocket versus all the things that came before. And Newton was supposed to be a handheld device, but it was more of a brick, not a thing that would fit into your pocket. And so, as I saw that, I was like, 'Oh goodness, that's a computer that fits into the palm of your hand.' So, you know, what's the future going to look like? It's going to need to have real security because it's going to have my most sacred and secret documents and my money and my contracts and all these things. So like what we need is to go build encryption libraries that work on really low power devices. That was the impetus behind the original Confinity. And like that's totally true. It all came to be. But if you are a thoughtful product designer, your answer is actually there are zero contracts and zero anything of value that sits on this. Fewer than 1 million devices sold mostly to nerds like me who just want to play with how to write code on this thing. And so the comeuppance that I got back then was it was very clear in the moment that we were too early. And the conversations internally were a little bit more like all right so some of these ideas are great but they're clearly 5 years out of time. How do we make sense with what we have today? And the honest answer then and now is it has nothing to do with handheld devices. Nothing to do with Palm Pilots. It has everything to do with the fact that there's absolutely no way to move money online. And we invented an extremely elaborate way to move money using two Palm Pilots and that's cool, but you just needed to move money. And the product market fit there came by like hitting us in the face and saying your web demo is more useful than your product because the web doesn't have this functionality at all. And so, it's a long way of answering it, but I'm often on the record with this notion of hard, valuable, and fun. And it stems from the right around the same time actually, Peter and I had this conversation that I sort of still remember vividly where he said something along the lines of the thing you built is really hard, but I'm just not so sure it's valuable. And most things that are really valuable are hard, but not everything that's hard is really valuable. And so you should always understand which way the arrow of causality is pointing because then you can decide whether it's something that you really want to spend your next 105 years or in my case back then four years working on. And so as we saw the market pull, my real lesson back then was hey from now on when I'm going to invent things I'm going to try to make sure I have a problem in mind versus a puzzle. You know, solving puzzles is great and I love solving puzzles as always, but real problems are more important.

Yeah, it's funny. I have exactly the same outlook on this problem. I don't call it the 15% problem, but I think it is uniquely afflicting the fintech universe. So I call it the Mint problem, which is where I came up with this was personal financial management. Yes. So every 3 years someone says, 'You know what I need is to connect all my accounts and get a perfect dashboard of my finances and really know what my cash flow looks like and I want to color code it and with AI and in 3D and the people who are like, 'Oh my god, where have you been my entire life? I got to have this thing. Of course, I'll subscribe for this thing for $30 a month or a year or whatever the cost.' Those are the same exact people who had that exact reaction to the last one of these and the one before and the one before and the one before going all the way back to Mint which was the first one of these like connect your bank account difficult as it is, Yodlee's half-broken, doesn't matter it's so much easier than doing this in Excel. And of course that very same group probably like their parents or the genetic precedence of these people were like all I want is to balance my checkbook I got this piece of paper and I'm never going to miss a dollar or a penny. And that group of people is like maybe I think 15 is very generous. I think it's like 3% of the population. And so that group is your total target market. And for some of them $30 a month is too much and they're not going to subscribe. But most of them are very good with money. So they actually have the money to burn on having a better tool. But you're sharpening an already very sharp pencil. You're not expanding your bag of pencils and pens. They're just saying for this particular consumer like the power user of personal finance, I got you your next Mint. Let's go. And there's a handful of these apps right now and they're absolutely stunning. They're gorgeous and yes, they're just that much better and yet I know what their growth curve looks like. It tops out. Anyway, so that's we agree on that front. The thing that I try to do and I'm again back to sort of over the years I spent a fair amount of my time rewiring my brain from being that of a software engineer to being that of a software designer and software engineers which I still am one as much as you know try as I might love a puzzle and build for themselves and great software designers or product managers product designers the first thing they do is they say let's embody the mind of a normal someone who lives in middle America who is not in fact a well educated computer science aware design conscious you know someone who's deeply obsessed with balancing their checkbook. In fact, the canonical person I loved to imagine when we were designing Affirm was a mother with a baby in her arm pushing a cart in a multi-lane checkout environment trying to pay for something. You have one hand. You got a line behind you. You got a screaming baby and a cashier who's overworked and underpaid. And the cashier is asking you, 'Hey, are these organic or not organic bananas? I can't tell from the label.' Like that's what that experience feels like. And you have to deeply feel that to design a product that doesn't suck. And so the idea of like, oh, cool. I'm just going to go into my app and dial me up a digital payment option is like, sure, or I'll go for the thing I know how to use, which is what's top of my wallet, right? And so I think just having that mindset and reminding yourself you're not building for yourself. If you are, you're going to run into that 1% problem again and again and again. And so that's where you know and over time you become better and better at it. I remember giving a talk at like startup school or something years ago where I said something along the lines of you have to embody the mind of a 17-year-old girl when you are trying to build a social media product. And the room looks at me like what exactly are you talking about? Why does this sound vaguely sus? I don't think you think what I think it means. You need to be so one with your customer that you understand what motivates them. And like the baby on your arm is a real problem. I used to go to Whole Foods and watch people check out specifically to understand how this actually works. And it works pretty simply. You put down the wallet, you pull it out with one hand, and that's the card that you're going to use. And so plastic is not going away anytime soon because of that specific modality.

Yeah, very much so. I mean part of why I do my own customer support on Twitter, right? We met through trying to offer customer. It's not just because I care about the product. Although I certainly do. I want to have that first touch moment when someone is upset or confused or unhappy, you can read your own copy a thousand times. Like the number of times I'd written what I thought was a great piece of product copy, you know, totally use this in the following way and then show it to someone like what exactly do you mean by this word? Like does the word sync or phasetron mean nothing to you idiot? Like no, it turns out to be not a commonly used word when it comes to, you know, day-to-day consumer physics. And so that I absolutely do. One of my favorite things to do when I travel. In fact, every city I've ever been to more than once, I can name my favorite grocery store to go to to people watch primarily to figure out how they shop, what they buy. Like you learn a lot. There's a totally I haven't been to Shanghai in forever, but there is a Japanese themed grocery store underground right next to my favorite hotel in Shanghai. And I would go there like right after getting off the plane because I wanted to see if they changed the way they place things on shelves. What's eye level? What's below eye level? Like they added self-checkout way before anybody else did. I was dying to understand how the self-checkout thing works. Interesting. They had a person help you self-checkout because it turns out to actually be harder to self-checkout than not. So anyway certainly been my experience.

That's a good idea. I'll take a note to do it.

We really kicked it off in 11 and 12 I think is like the official incorporation date. Affirm came out of a speaking of hard valuable funds. So I had a Yeah. So this is sort of a brief historical detour. But please Affirm was birthed in kind of a very slow nine-month process or so where I was hanging around with a couple of friends who eventually became my co-founders but who I mostly knew from PayPal years and we talked a lot about underwriting and how we build these super elaborate models. Like we when we were at PayPal random forests were already a thing as a modeling technique. But we absolutely pioneered it in an industry that at least I was aware of was doing random forests to do fraud fighting and we built these like ensemble models that attacked each other and you know voted on outcomes and did all like all the sort of under oversampling techniques and a lot of these things were just coming out of our own research. We like we were not I mean we're reading all the academic papers obviously Berkeley is where a lot of this stuff came out. And yeah, we were like definitely being told by a lot of people like what are you doing? Like there's some very basic ways of doing this. And at the time my answer was but like if you imagine this thing as a curve of events like you are like literally living in these like lumpy sums like you're barely approximating the curve like you don't understand integration and like these things are infinitely differentiable. Like you should only be using integration. You should not be trying to build up a bunch of, you know, rectangular blocks to sort of use. Your mom, your mom would appreciate this analogy.

Yeah. And so, and so we're like I just like people who compete with us like they are stuck in middle school math and we like we can't wait to get to grad school. Totally. And so like that that was always and it's a little bit of a hubris like obviously at the time we were we probably thought we were the smartest tools in the shed and we certainly weren't. But we definitely were a lot smarter than some of the players that we had to compete with. And so the same cast of characters but like 15 years older and a little bit more grounded and less self-obsessed. So like you know on the one hand like we learn a bunch of things. There are a lot of other smart people doing a lot of other smart things. Maybe we didn't really know about them. But this notion of a lot of competitors, a lot of people who have to solve these problems just choose the easy way out. They take a score that someone built not for the purpose like FICO is a brilliant company and an amazing collection of people and research but the tools are primarily written that FICO creates they were all published for the world that was largely offline and so the decision to say yes or no was around the person their ability and willingness to repay at the time of underwriting a line of credit or a mortgage or an insurance policy or something that you know it's a moment in time a big thing yeah that they're doing at a moment in time and so even though the event chain or the event curve is infinitely differentiable they just didn't have the luxury of saying and by the way we reserve the right to change our mind next time you swipe your card because that just wasn't a thing you could do to fight fraud a little bit but like no you can't change the risk assessment and certainly couldn't reprice it and so you had to reprice it on the spot once for every 5 years on average or whatever. And so we had all these conversations and basically like if we applied the same kind of if you're willing to do calculus or slightly more complicated math surely we're going to find some real savings like the area under the curve we're going to fit with our approximations. It's just going to be much smoother. We're not going to have these big lumpy boxes that are adding up to something roughly resembling the area under the curve. And so that was the foundational kind of logic behind Affirm where we can do better through math and the better living through mathematics is like I mean obviously it's a reference to better living through chemistry or whatever the exact line of that album but it's in our DNA this idea that you can and should not be afraid of applying progressively more sophisticated math to these problems if you have the opportunity to benefit from it. And we thought that point of sale lending in particular, which by the way was a later idea, we really settled on point of sale lending because it's the sort of a self-distributing loan type. Like we weren't particularly married to the idea of, you know, this or that loan type. We just thought we would do better underwriting because we could do it in real time. Interesting. We then have four computer scientists starting a lending company. What do we know about distributing loans? Like well the one place where you could have a loan is when somebody's already saying yes I'll buy this thing and I would like to use a credit product and well like literally we're competing with a thing that was underwritten on average three and a half years ago and by the way a merchant price that was set by some independent Salesforce salesperson who sold you a box that is like literally from 1980s with a integrated chip on like eight generations ago like that thing is like it's not computing anything it's just collecting an MDR like so if we float both the MDR and the EPR. We could like really really find the perfect approximation of the risk curve. So that's the origin story that's where it came from. The thing that actually doesn't get a lot of airplay and it's worth highlighting in terms of what turned out to be important. So there's a hypothesis like so we were having lunch in our favorite Middle Eastern place and sort of bantering around wouldn't it be cool if we like we found out what so by the time we were having this conversation random forest were like not just a thing there was everyone's thing and you know the modeling with that was like okay everybody's doing it now like well so what's like the latest hotness what should we do and so you looked at gradient boosting and like all the next things that are now of course well understood and well accepted so we could find some interesting area under the curve from Yes. And as we were poking around all this stuff we basically asked ourselves like what you know given that we know that at PayPal we were like the tallest you know most intelligent species because we were willing to read these papers and apply all these interesting bits of math and of course now everyone caught up. So it's not a permanent advantage. So how do you build a more permanent advantage? And so the idea we had which has turned out to be completely true is you build a factory to produce model after model after model. So you're not beholden to a single technique. You're not beholden to a single new paper. You don't just read about random forests first and then suddenly everybody caught up. You have to have a machinery to fully industrialize the production of the next generation of model. Both in terms of we have new data and so we need to process it differently and learn from it. But when and if there is a new paper or a new technique that sort of kicks you out of your chair like oh my god like this is a completely different way to model things you're ready to just pick that up and go which applies to today. So we had that moment in 2017 2018 when attention is all you need came out and so every one of us who was in the industry said wait a second I thought like first of all when we were doing this at PayPal I remember trying to build a neural network driven way to detect fraud and fight it and it totally flopped and like just not functional and it comported well with my experience in college doing this stuff or like yeah neural networks are like they were supposed to be a thing in the 70s and you know I went to college in the '90s and they still science fiction, right? Like these things sometimes take some time. Yeah. I remember reading the age of thinking machines and going like, 'Oh my god, like it's here now. I just need to build a bigger computer.' And we built a lot of big computers in the '90s and nothing happened. And then in 2017 it did and it was this amazing moment where suddenly like wait a second you can actually build a simpler thing a lot simpler than some of the earlier networks that people were playing with. And it really seems to work. And so the thing that we were able to do at Affirm is say all right so like obviously we're not in the business of chatbots or LLMs but we are in the business of building new models from the latest scientific ideas and observations and so we have been building models based on a lot of the ideas from attention is all you need and sort of surrounding papers before and after and we have some really stunning results to show for it and so you know I don't want to front on anything too exciting here but we've been consistent in our ability to stay ahead of the research with the most cutting edge research because we invested in building out this model building factory which has just been a really really fun thing to be a part of to be honest.

It's definitely I mean one of the reasons why AI is not taking any of our jobs or not not not many of our jobs in my foreseeable future is because ultimately the question you're asking and like many in that class of questions are very human like it is you can program an AI to do X or Y and you can tell it to try to be creative and it'll sort of span the between X and Y but to get out of the box you need creativity which is primarily curiosity, intuition, and taste. And we're pretty far from encoding taste. We're quite far from encoding intuition, and curiosity can be programmed, but curiosity in and of itself is not enough. You kind of have to know which lines you're going to paint outside of. So, the way we try to encode those three behaviors inside our very human team is combination of just challenging them to ask, are we plateauing or not? And you need metrics to tell you yes you are or no you're not. And so there are some great well sort of agreed upon metrics in statistical learning and AI and sort of all the I mean this is a well-studied area of math and so you have that very same area under the curve where you can look at how the model ability to rank is progressing and ask the question so how much better has it gotten from version n to version n plus one and if your area under the curve improvements are gently asymptoting towards nothing, you're not winning. You're doing fine. You're not losing, but everyone else who is taking these leaps forward is catching up definition unless there's sleep on the switch. And you're kind of thinking in the back of your head like if we're like kind of flattening out, there's someone behind me who's going up a hill like this. They're just earlier in the curve. Like you got to be very aware of the fact that you're not the only one competing. And so every year, every 6 months, we ask ourselves, what have we not tried that at the downside would actually bring the area into the curve down? Which is okay because we already have this performance and it's great and we love it. But what's the high beta thing to do? What crazy thing can we pull out of a bag now? For example, can we try something that's transformer-based? Yeah. There's all sorts of risks and not just like what if it doesn't work but also what if it run you know it's a sort of classic thing to worry about with transformers is hallucinations and so like what if it hallucinates a result that isn't real what if it hallucinates a result that isn't compliant. So there's all the very serious things that you want to worry about as an underwriter and certainly somebody who is subject to all the relevant underwriting related regulations. But if you don't try it, you're not going to find out that it's working. And that's how we keep ourselves honest. We ask the question like, you know, given what we know about the industry, if for this particular family of models, the curve is bending towards flat. And we know everybody else is chasing us. What new curve are we doing? And by the way, this logic, it has to apply not just to underwriting, but also to product. We have a very similar approach to product building where every time we look at our growth curve, we ask ourselves, all right, so if we are just the first in a pack, who's behind us and how far below the elbow of this particular curve are they? Because if they're not that far, we better be inventing a new curve.

Lots of good threads to pull on there. So I use the word explainability all the time. Over the last sort of couple of years as I've been looking at all the latest greatness in AI and ML, I've become more and more convinced that we'll probably solve explainability really really well. Specifically with LLMs and things that are sort of hang off the LLM tree, if you will. The most interesting problem that I think needs to be solved for this new class of models to be really sort of to come into its own is replayability, which is a sort of a code word for getting a different result while running a model with the same exact inputs. You know why FICO is so valuable? It's stable. Like it's actually too stable in my opinion for a variety of viable occasions like but it has to work like a few years at a time without costing lenders too much money. Affirm's internal transaction score internal user level scoring system that we've built like half a decade after we built the transactional one. They're significantly more sensitive but they rely on real-time data and so and and but they're 100% replayable. The same exact real-time data runs through that model will give you exactly same decision, same score.

Same same everything. And so I think that's the bar you have to meet with these transformer-like, again, like we're throwing around terminology and I'm deliberately being vague just because I don't want to give anything away to my...

And so I think this explainability thing is actually almost like a downstream problem from replayability, and replayability is a navigable part of these models. You can go back to... there's some really great papers out there, very recent, like last 30 days, that I just looked at, that were sort of dismantling this notion that in the older versions of GPTs, which are now just one of several contenders here.

If you set temperature to zero, you would get the same exact response every time, right? The answer is wrong. If you actually set your randomness factor from the possible candidates for the next token, you just say always pick the very top one, don't mess around.

There's some really fun... if you remember your college days, numerical methods is probably the name of the class you would have to take to know what I'm talking about. You will immediately recall things like fixed-length precision. So you have a 64-bit float, and you decide that maybe 48 bits is going to be representing your floating point, and that's 48 bits of precision. So if you need a 49th bit of precision to define a float, you don't have it. It's going to get rounded down to 48, and that's it. And by the way, a lot of the modern models actually use like 8-bit precision, so they're much, much narrower. So because of that, you have compounding of errors and all that good stuff. So you actually do end up with some fairly easy-to-understand ways for why hallucinations and lack of predictability, lack of replayability were a thing.

But then there's more, and I'll stop there because I was like, whoa, this is blowing my mind. I thought I understood it, and I forgot which lab put it out, but it was something that I thought I understood better than the average bear. I was like, all right, I came in here thinking that I got this, and I'm walking out knowing that I don't. I know more now that I read the paper.

Yeah. I thought I left my arrogance at the door at the founding of a firm, but it turns out I still have arrogance to give up. Anyway, back to the more tangible, maybe easier to reason about, regulation side of things.

So I actually think a lot of this will sound like totally upside down from what maybe as a Silicon Valley entrepreneur I'm expected to say, but first of all, I think regulation is more good than bad in a lot of cases. Some regulation is dumb. Some regulation is outdated, profoundly outdated. And it's mostly when the regulator or the writer of the regulation, which is frequently not the same person charged with enforcing it, writes things down almost too literally. Like, thou shalt not use an 8-bit integer, only 16-bit integer will do, is total lunacy. If you ever see that in regulation, that is a dumb regulation, because by the time it's written and published and enforceable, 16-bit integers will be a thing that no one cares about because you have 128-bit integers now. So anytime it gets super specific as to the implementation of the regulation, it always ends up falling short, in my opinion.

On the flip side, things like you can't discriminate against people based on who they are. You should be allowed to do so financially based on what their financial state is, really sound. It made sense when they wrote the first words of fair lending, and it's still true today. And so the standard of don't discriminate based on who the person is, where they're from, what they believe, who they marry, all the obvious things we take for granted now but did not in the 1950s, I think they're still true today. And that serves us really well. And the thing that is not in the law, which is great, it doesn't say things like, 'And by the way, if you're doing this with a computer, different measuring stick.' It's like, I don't care if you're doing this with an abacus or in your head or with a scorecard or a FICO score, if you end up discriminating based on race, that is not okay. Full stop. And so I think that regulation is fantastically great and doesn't need any words changed. I think some of the regulatory attempts to corral technology into the technology understanding of the moment frequently backfires because technology moves faster than the regulators can, and that is how it should be. You don't want to change regulations too often because it's too hard to enforce. So that's the basics. And the reason, by the way, for my philosophy on regulation writ large, not just in financial services, I grew up in the Soviet Union, so I saw Eastern European flavor of socialism up close and personal, and it sucked. Anytime someone tells you it worked in the Nordics, it's okay, the Soviet Union almost got it right, but no, it doesn't work. It's inherently corrupt. The idea that everyone's equal, but those distributing the money get to keep a little bit more than others, is abhorrent, and it is exactly what happens. The classic joke about the Soviet Union was that people who ran grocery stores were always fat.

And so socialism does not work. And yet free market economies, though they self-correct, they sometimes take too long. The classic example of the pizza shop that doesn't sell to gays is like, 'Wow, that's really offensive.' Fortunately, there'll be another pizza shop next door that will sell to gays and have a big rainbow flag. It'll be great. But in practice, how long before the second pizza shop opens and kills the one that is prejudiced? I think in practice, a regulation that says thou shalt not discriminate based on creed or sexual orientation is a pretty good regulatory thing to do, because it just runs the fact that most of us understand it's not cool in the 21st century to discriminate based on something completely unrelated to what you do as a business. And so I think that's the sort of speed-up towards a normal, sane, not stupid way of the world, perhaps pushing the market along versus waiting for the market to get there. That is, in my mind, the primary purpose of regulation. So as a consequence, generally speaking, if it's thoughtful, if it protects consumers, if it's there to speed towards the natural equilibrium that most reasonable people can agree should be there, big fan. Let's have more of that.

Yeah. And there's definitely no one-size-fits-all. And if you want to be truly risk-free, stay home and don't start companies, because that's one way to do it.

So we did a couple of things, and they're not exhaustive. We did even more than I'm about to describe, and as I go, I'm sure I'll think of more, and I'll go until you stop me. So one of the very first things we did, we basically said we should not do late fees and compounding interest. We just won't do that. And you can actually mount a reasonably good argument why that's a dumb idea. Late fees are there to slap your wrist and tell you, 'Hey, stop being sloppy. You should pay on time because when you're not, everyone in the system is paying a tiny amount for you to have the freedom to pay whenever you want instead of sticking to a schedule.' And I think that's actually true. Everything I just said, I happen to believe. As a libertarian, or at least I was probably more ardent libertarian 25 years ago when I was also a lot more arrogant. Hopefully I've become slightly more humble and aware of other points of view, but that argument stands the test of time reasonably well. The reason not to charge late fees, more than anything else, is twofold. One, it profoundly aligns you with your customers as far as financial well-being. If you charge late fees and the borrower is late, you'll make more money. So every contract is a little bit of an agreement and a little bit of a bet. I bet you'll be good for the money, I'm going to give you a loan, and you bet that I won't screw you when you are late or something bad happens. In the normal world of late fees and compounding interest, the bet I'm making is like, you think I'm on your side, I hope you don't notice that I'm not, because I want you to over time pay me everything back, but I kind of want you to be late because then I'll make more money. And the marginal cost of that gain is nothing. $38 for you to skip a day of payment is an amazing business opportunity for a lender.

It's basically infinite. And so the argument we made was if we get to make money only when consumers behave in the way that we ask them to behave, versus ask them but wink wink don't do this too frequently, we would build a better underwriting engine. And so as an incentive not to screw up, to build the best approximation of the risk curve, we should keep ourselves as pure as possible relative to the revenue-making opportunity of a lender. That is, if we say, 'We're going to loan you $1,000 and you'll pay us back $50 in interest,' that is the most we should ever have. No fees, no late fees, no compounding, no deferred interest, none of the sort of cool bets where I bet you'll be on time, but I bet you're going to be less on time than you think. And so that product innovation came entirely from this: let's just be purist because it'll force us to be good.

Part two: because we knew we were building a completely new class of products. You're totally right. As we were doing this, I would sit down with a partner bank or a law firm to hire and explain what we're trying to do. The reaction was inevitably the same. 'Okay, so part one, you're not charging late fees. You're going to lose your shirt. That's half the profits in the industry. Don't be an idiot.' But we made a decision that that's what I want to do. But more importantly, they would say things like, 'I don't even understand if this is a loan. Is this a retail installment contract? Is that what it is? No. Okay. It's not a line. So you have to show Truth in Lending disclosure every time. But you don't want to do that, do you? A Truth in Lending disclosure for every purchase in a store is just going to slow everything down.' And in that moment, I had a choice to make. I said, 'No, actually, we are going to show a TIL every single time. You're totally right. It is going to be a bump in the road. But I think if we're claiming that we are completely on your side, there are no late fees, no compounding, etc., an extra screen showing you that you will owe nothing extra will maybe slow you down a tiny bit, but it will make you feel like you're making a more responsible decision, give you more sense of control, but more importantly give you clarity that we're not just marketing ourselves as a no-late-fee company. We really will commit to you in a formal way that there are no late fees.' I convinced myself in the moment that this is going to work, it's going to be okay. Conversion is not going to suffer so badly. And the lawyers on the other side of the table are losing their minds, and I'm like, 'No, no, I think this actually makes sense.' And as we did this, I met with a bunch of regulators, quite some time ago now. And they all asked me, 'Hey, so how are you cheating? Are you doing any of these things?' And I said, 'No, I really am not. We're building a product that's just going to not be a cheat code.' And the more I talked to them, the more I understood that most of these people became regulators because they genuinely cared. They're not like, 'I'm a lawyer and I need to wield a stick and find bad guys in the industry.' They all came to it from the place of there are people who are just too busy, too tired, too overworked to read the fine print, and my job as a lawyer is to protect them, to make sure the fine print doesn't say anything too horribly nasty. And that actually influenced one of the core values we have, which is no fine print. I literally had this conversation enough times: 'You know what, we're just not going to have fine print, so that people who regulate us don't have to read it, and the consumers more importantly don't have to read it anyway.' So my regulatory experience or the relationship was really shaped by these early conversations: I am not trying to find shortcuts. I really want a product where people feel great about using it. So as I was doing this and we're trying to build a team, part three, and I really will stop because all the memories are flooding back. I remember starting to talk to early engineers, and of course as we were building our AI team, I really wanted the very best people. At PayPal, because ML and AI were so underutilized, we had this total pick of the very best people in academia and the industry, and it was great. By the time we were starting Affirm, Palantir was a thing. Everyone wanted to hire really great ML engineers. And so I was trying to ask, 'Where will I find a gusher of people who are just really good in machine learning and artificial intelligence and are interested in what we're doing?' And I realized that probably a lot of people who would normally work in a quant fund would go to Wall Street and build these major scale models and train them on very proprietary data. But we knew we'd have a ton of data because we're embedded at the point of sale. It'd be really interesting to train these things. So where do you get the best people? The thought that came to me, which turned out to be exactly right, is if you build a product that they're not embarrassed to talk about to their parents and friends, they wouldn't go for the money. Because if I sit on Wall Street and I squeeze pennies out of the market, my job is efficiency of capital. That's a cool story, but it's shorthand for 'I make a lot of money pushing bits around trying to find the next transaction that the other guys haven't caught on to yet.' What we're building is actually helping real people buy real things at states that are not party to this Wall Street bonanza. So if you have that connection, if you didn't come from New York or LA and you're not just naturally drawn to the Wall Street money, you're like, 'I love the idea of helping people like my parents or my cousins come into whatever it is they need to buy. I feel great about building machine learning models. I just don't love the lending world.' It's like, 'Well, this particular lending world is unlike anything you've ever seen. It's no fees, no screwing the customer. We have this extremely high brow mission that we really believe in and care about. And we're not afraid of the regulators. We love talking to them and explaining who we are. We like responding to information requests because when they say, 'Please disclose all your fees,' it's like, 'Cool, a sheet of zeros, please.' And so the early team foundations, especially on the engineering and ML side, was really this: 'Hey, you will be proud of what you do. You will not be embarrassed by who you work for or how we do our business. You will never have to say, 'Oh yeah, I'm sorry, I'm the one who gives you the 0% with an asterisk clone.' We don't do those. And it's exciting to be in that team.' So those were the three foundational trends that fed our choices on the product construction.

So that used to be a major worry of mine. Because when we were tiny, I was in everything. I would... back in the day, we would hire product managers, and I would tell them, 'You get one question to ask me about late fees. I know you came from financial services. People who didn't come from financial services just didn't have the affliction. But if you spent a lot of time in classic financial services, you'd be like, 'So you guys don't charge late fees. You know there's a lot of money in there, right?' Yes. Let me explain. Here's why. We want to align ourselves with our customers. Okay, cool. That was your first one. You can ask me again. And if it's the same question, you'll be fired. If it's the same question but in an interesting new thoughtful way and you had an insight, like if you want to make the argument for stretching the cost of time value of money across the entire user base, I'll humor that. But if it's 'I thought about it and we should charge late fees because it's more money,' you should not work here. You didn't get it. And so that was year one or two. And now it's almost year 15, and it is true that I'm not really involved in hiring every single product manager. I still interview quite a number of people, but mostly not. And so it turns out that if you are completely obsessive about the cultural DNA of the company in the early days and you just never ever compromise early, it actually becomes self-sustaining. This is an experiment you can only run once, so I don't know if this is true for everybody. But if you looked at the average tenure of senior people at Affirm, Libor, our president, responsible for both product and engineering these days, he started as CTO and kind of just picked up operations and risk over the last 10 years he's been here. He was also my college friend, so we've known each other for a very long time. And Michael, our chief operating officer, is now in his seventh and a half year. And Pat, our head of sales, is like 10 and a half years here. So you have a lot of these figures of importance and authority and key cultural significance. They all bought into this, and they weren't just interviewed by me. I was there with them incubating the product ideas, making mistakes, realizing we built something nobody cared about or something that really took off, and handling internal objections to why don't we do X or Y. 10 or 12 years ago, somebody asking me about late fees was obvious from the very beginning, but there were plenty of other much more subtle things like, 'Why don't we do this or that?' And at some point, I'd have to think about it pretty hard because this is an important cultural setting moment. Sometimes I thought yes, this makes sense, and other times I didn't. But in the early days, you embed all these decisions in the popular consciousness of the company, the shared DNA. And as you get to 2,500 people, which is where we are now, you have a lot of people who have been there through the years of this crucible of what is Affirm, how does it do when this kind of moral problem or moral hazard floats up in front of us. So it's more than just me. In fact, I used to make this joke, and now that I'm in my 50s, I've become maybe slightly too sensitive to it. But if lightning strikes me, not to make it too morbid, I don't think a single thing would change. I think the company would just self-enforce the DNA with or without its founder. I'm probably one of the more passionate people about the way we do business, but I think I'm far from the only one. I don't think it's this one Max holding up the dam of late fees. Quite the opposite. I think if anything, if lightning strikes me or I become senile, which was the 50-year-old joke, and started asking my own questions like, 'Why don't we charge late fees? Seems like a real problem,' they would rebel against me and literally toss me off the cliff. 'The old man's lost it. We got to get rid of him. We are the company that does not charge late fees under any circumstances.' So I think it's so enshrined in who we are and how we do what we do that it's entirely its own life. I used to take great pride in the idea that we never changed a word of our mission. It's fairly awkwardly written down. I wrote it kind of off the cuff on a napkin: 'Honest financial products that improve lives.' Cool. I like it. Let's go with that. Two years later, I was like, 'Oh, we should probably change it. It could be a little bit more clever, maybe pithier.' And they said, 'No, no, you can't. It was written down when you were starting the company, and 15 years in, I can't change a word. If I did, somebody would be like, 'Oh my god, he is modifying the mission. A comma has been introduced. Stop it. He's compromised. Can't do that.'

That part is actually I'm okay with this part being calcified. You cannot allow calcification of model building techniques or willingness to chase the latest hotness in machine learning, but calcifying a great mission? Totally cool.

So it's actually refreshing to hear your analysis because it's wrong. But it's wrong and it's right and it's wrong.

From the very beginning, the north star, when you're trying to build a thing that doesn't exist, you're always leaning into analogies like 'We're like Uber but for songs' or something. I was trying to explain to our earliest backers what we're going to be when we grow up. The shorthand was always we're going to be like American Express but in 21st century technology, where you would see us both as your source of credit, but much more importantly as a network where more and more consumers would find us and more and more merchants, and assuming this idea of honest financial products, no fees, no screwing around, no gimmicks, you would prefer us. And so as your preference for Affirm would grow, so would the number of merchants saying, 'Well, I can't ignore it. It's just too many people who like this thing so much.' It would create a flywheel of more consumers that like us because we are honest would naturally grow into more merchants. More merchants would then essentially act as an advertising engine for us, with more consumers saying, 'I've seen this logo before. It's on a bunch of places now. Gosh, I've seen it on Amazon.' So it would just be a flywheel. So we've always had this point of view that we are building a network, and one side has to reinforce the other. We are a much larger company now than we've been through the years, but we're still inevitably engineering-constrained. So as we prioritize the next thing to build, the first question I ask is, 'Does this help one side of the network or both?' And if it's both, that is the highest value thing we could be doing. Building something like we could offer our consumer something really neat, like everybody gets a free hat. How is the merchant going to care? They're not. 'Okay, it's nice that everybody gets a hat, or access to airport lounges, that's great, but that's not really helping the merchant side of the equation.' It probably helps consumer retention, makes them feel like we care for them even more, which we certainly do. But a better version would be something like create a platform where all the merchants can promote their discounted APR deals. So we had this really cool thing called Big Nothing Day, where the entire management team, but also the company, is really big into silly puns and strange funny movies. There's a fair amount of goofiness going on. We're a very serious company, so if you're as serious as we are about math, you have to take it out.

A lot of my shareholder letters make references to The Big Lebowski and things like that. My pronouns are El Duderino, and so on. Anyway, as we built our app, we thought about what it was used for. Initially, the use of the app was of course to pay your bills, set up your auto pay, all that stuff. And then it's like, well, we might as well make a catalog of all the merchants that allow Affirm, because that would be great. Consumers would be able to find, if I'm going to go to store X online or offline, is Affirm available? Then I'm like, well, some of these merchants offer 0% APRs, and that's really great, and we should let them highlight that in our app. About a couple of months ago, we had this idea: why don't we have a big festival where we have a day where any merchant that wants to offer zero will do so, and we will make it our business to tell all of our consumers that these merchants are offering 0% loans. Zero interest is the best price. And of course, this being Affirm, it's not like zero asterisk secret message. It's like zero means zero. You're paying for a $1,000 thing over 12 months, and it's going to cost you $1,000, which is pretty flipping good. It's better than any credit card can possibly offer you. So that's what we did. It's a great example of a thing we built whole cloth from nothing over the course of a few weeks. And then over many months, we actually signed up a bunch of merchants who wanted to participate in this extravaganza of zeros. It was huge. People loved it. It was a very good explanation of who we are to our consumers and to our merchants at the same time. And so this singular moment, singular set of features, singular set of productization, is a good example of how we think about building products and why the card and the app and the point of sale and all these things all play together to further the spin of this flywheel of more merchants care about accepting Affirm and using our 0% and other promotional capabilities to inspire more consumers that will come back and shop at more of these merchants and feel good about how they're buying it and make sure it's affordable and there are no late fees. So that's the crux of our product strategy, our network building strategy. That's why I was sort of chuckling at 'yes we are a merchant company' because we wanted merchants to believe that we will bring them more sales, and then we did. We had a lot of consumers, 24 million of them last quarter, 12 months trailing. And that's enough now for merchants to be like, 'I should offer Affirm.' Having just Visa and Mastercard logos, that's not enough. I got to have the Affirm logo too now on my door. As that flies on, it becomes like, well, we're now much more of a consumer company, aren't we? We have this app and the card. And of course, the more of that we have, the more merchants are like, 'Whoa, I definitely need to make sure Affirm is available.'

And we have two hands, and we're pushing both in the same direction.

So that particular one, and a class of these problems, is actually kind of very familiar territory. Long before we had this, the internal name for this product is AC/DC, which I did not name, but it's one of the better names I think we have. And it's Affirm card on other debit cards. So it's AC/DC.

I thought it was great.

I'm a t-shirt. Yeah.

I usually sport at least two, more often three, Affirm logos. You will not have a photo that I just took in front of 10 Downing Street in London where I was fortunate enough to be invited to speak with some very senior people there. I had to wear a suit and tie, and I was like, 'My god, I feel naked. There's not an Affirm logo on my body.' So I snuck in Affirm socks. I have a photo of me in front of 10 Downing Street pulling out my leg cheekily showing, just so you know, Affirm socks are there.