Max Levchin - Building Affirm, PayPal, and Why He Only Starts Network Businesses

The only kind of businesses I like to start anyway are networks. And once you have a product market fit, if it so happens that the product is a network, it becomes to at least some degree self-reinforcing. My career in computer science is a sort of a constant attempt to lean back into math. The goal of Affirm is this nuanced approach where credit is good. I firmly believe that access to credit is really, really important. Permanent debt is terrible. If you're running the kind of company where you're just absolutely sure that there's nothing left, like you've reached the end of the internet, there's nothing to build, you probably should look for someone else to run the company. Not to delve too deep into the Satoshi conspiracy theories. It's probably more likely a group than one person. What did you learn working with Elon Musk? How about Peter Thiel?

Yep.

And my grandfather and my uncle, like I have a lot of physicists. It runs in the family. I'm sure there's not entirely uncorrelated, but yes, physics is definitely a thing in the family.

Yes. Although the same grandmother spent a lot of time dissuading me from majoring in math because there is no Nobel Prize in mathematics. She told me over and over again. Yeah, she was actually, she did not speak with that accent to me. That was entirely artistic license. She spoke to me in Russian. So I probably wouldn't need to replicate that here. But it is true that there's no Nobel Prize in math. The legend has it it's because Nobel's daughter married a mathematician and he didn't approve the marriage. I don't actually know if that's true or not, but that was also a lore my grandmother told me. Sounds like a father-in-law thing to do, but I majored in computer science in part because I really wanted to do math and math was definitely frowned upon and my grandmother was very pro-physics, tolerant of computer science, not super excited about math as a major.

Definitely a very conscious effort and so I don't know if I have any profound techniques to impart. I was a reasonably well-trained musician by the time I came to the US and I'd played clarinet since the age of six or so. And if you are a woodwind player in particular, you really have a good sense of control of your embouchure, which is a fancy word for musculature of the lips. And so, and you presumably have enough of a musical ear to be able to say, 'Wait a second, that's not quite right. I need to work on that vowel or that phoneme.' And so I think it was just a combination of having the predisposition towards constantly refining a sound and knowing what it takes to refine. So it's funny enough the same grandmother was completely fluent in English. She wrote many papers that she didn't have to translate. She wrote them in English. And her own English accent was horrible. Like it really sounded like a Russian villain from a Hollywood movie. And I remember when I was a kid, she would show me these really thick books that she had on a shelf that's what she learned from. So she'd never had English instructions. You just learned from books. And to learn the accent, she had these books that showed sort of cross-section of a human face with like the placement of the tongue. She'd be like, 'You need to do that diphthong. You have to do the diphthong, right?' except in her case it actually sounded like something totally not like the word difficult sounds like. And so I remember thinking you can only go so far without direct imitation. So but anyway I definitely worked on the accent very hard and I'm still amusingly enough slightly self-conscious of the accent. My high school buddies that I still keep in touch with that heard me fresh off the boat sometimes mock me or troll me by telling me that accent is coming back and I immediately tense up and try to figure out where did I screw up, like what did I say? So it's definitely no longer conscious modification of the speech but I'm definitely conscious of the fact that I used to have one and I have to maintain the perfect accent. And the reason for it was actually pretty simple. I realized as soon as I got to my high school in Chicago that there were kids who spoke like the locals and the ones who didn't. The ones who didn't just didn't get the same treatment. So like it or not I had to blend in.

Yeah.

Sure. Actually goes right back to the grandma conversation. So years ago when she sort of hammered into my head that there's no Nobel Prize in mathematics, I sort of said, well surely some Nobel-like character must have introduced a thing to highlight successful mathematicians. And there's of course the Fields Medal which is a great way of being distinguished as or being recognized as an important contributor to the field of mathematics. And cryptography is in fact kind of, there's actually a really really interesting kind of tangential related tidbit. But my career in computer science is a sort of a constant attempt to lean back into math which every time touches this incredible vaguely accidental luck. So when I was in college all I wanted to do was math and I was majoring in computer science. So I ended up just kind of leaning into number theory which is a discrete mathematics part that is really useful for three things in practice. Cryptography which was my first love, computer graphics which is another sort of big part of my life for a while, and machine learning and now deep learning of course. And so sort of I keep on bumping into this one part of math that I just sort of fell in love with in college and it keeps on having these amazing commercial applications. And so sort of a fun fact to it, before the web, before payments, before PayPal, the natural progression of a number theory specialist, sort of college degree in computer science with some number theory inclinations would have been to work for AT&T or go work for the NSA, but I was not a citizen, so the NSA wouldn't have me. Back to the Levchin Prize. So, I think it's 10 years ago now, I realized that there's not a cryptography specific equivalent to the Fields Medal and by then I'd sort of acquired a bunch of friends in the field of cryptography that some of them were actually very helpful to me in the early days of PayPal, just friends that I met along the way and sort of had this dinner where I said there should be something to recognize people that really push our field forward and I'm an occasional participant, a little bit of a tourist these days because I keep on starting companies but I love the field. I feel like there's an opportunity to recognize contribution to practical valuable inventions and ideas in cryptography. Why hasn't there been one? And we sort of went around the room in this dinner and no one stepped forward and so why not, like maybe it should be me. I did not initially plan to name it after myself, but as prizes go, you sort of set up the funds and hand it off to some sort of an independent body to judge and award every year. So I have nothing to do with neither the nominations nor the final decisions but the committee that runs the prize had a free hand in naming it and so they chose to name it after me which for a time was embarrassing and now I sort of feel grateful for their recognition of my recognition.

Yeah so the latest I guess recipients, they've been active in the field for half a century.

It's not really a problem. We've been living with, we being the practitioners and people who are just interested in the field, for 20 years now with a full understanding that quantum computing is coming. It's a matter of when, not if. It is a fundamental promise of quantum computing that you will have the ability to brute force or quasi brute force these schemes that power things like RSA and several other cryptographic primitives. We have plenty of post-quantum, quantum-resistant, post-quantum crypto algorithms. So I don't think there's, plenty of teams that I think are scrambling to say, 'Wait a second. We thought this is coming 10 years from now. We were caught asleep at the switch.' But most sort of well-managed systems are long, at least planned for a thoughtful transition to a post-quantum future, post-quantum cryptographic futures. I think the world that was rocked is more of a newspaper headline, less serious system design.

That's my theory.

You know, for a long time I thought Adam Back was in fact Satoshi. I've met Dr. Back. He swears it's not him. I have no reason not to believe him. I therefore revise my views to one level of indirection. That is to say, I think he knows who it is, but it's probably not him. That's as far as I can go.

It is less likely Jack Dorsey in my mind than Adam Back, even after Adam says that it's not him. But it's also perhaps because when I was hanging around Cypherpunks, I did not meet Jack Dorsey. I met him later. So in my mind that person is already, I've already intersected with them, whoever they are. It's also not to sort of delve too deep into the Satoshi conspiracy theories. It's probably more likely a group than one person. It just feels more like a work of a couple of really thoughtful people because hashcash and stuff that Adam Back put together quite publicly without sort of being synonymous is a foundation of what's gone into the original Bitcoin paper and so it feels like a collaboration not just one man's genius.

Probably fill out a full interview at some point. He has an incredible sense of conviction about just about everything he says and does, which he is probably the first person to tell you he's not 100% right about. So there's sort of a typical truism in leadership and founder leadership in particular where you sort of, a wrong decision now is better than indecision and right decision later. I think he takes it to the absolute extreme where often wrong, never in doubt, and that is a superpower of his where he will train his aim on a problem, see a solution, declare it to be so, tell the troops to march forward. It's probably better than 50/50 right, but it doesn't matter. He doesn't mind being wrong and he will just retry and retry and retry again. And I think that's certainly something especially for someone with a kind of a protoscientific background/mind is a very difficult thing to get behind. Like 25 years ago, 30 years ago in the 90s I would have said that's a crazy thing to do, like let me lean back, think it through, plan out the game, I'll be right then I'll move. Watching someone who is willing to make the move with seemingly barest of thought because there's many moves, perhaps infinite opportunities for refinement later is a good counterweight to that approach. And sort of I'm not sure where I am these days on sort of shoot first, apologize later versus lean back, think it all through, but I'd certainly picked up quite a lot of the got-to-decide sort of attitude.

Elon and Peter actually share that sort of decisiveness as a superpower thing between the two of them. They're quite similar. Peter has another thing which, I mean Elon and I didn't work together that long so I haven't seen him in the sort of full gamut of leadership skills. Peter has one that is truly unique. He becomes so genuinely excited for the work you're doing. Like you can't fake it. Like he will become not just a willing participant or a supporter. Like he can be giddy with excitement when he sees you do great work and that motivates people like you will not believe. Like when he sort of leans into a professional relationship and becomes a supporter you sort of feel like you have to do your best work. Like this guy is so invested in what I'm doing he may be more invested than I am, like I can't let him down. And so if he is the CEO of a company you're running it's an incredible leadership skill because everyone around him knows that he is so invested in what we're doing together in your work specifically. Your work is really important to him right now. So that's a great skill and a great sort of, I'm not sure it's a skill versus an innate behavior of his, but it's always motivated me to do my best work at PayPal.

I think so. It obviously goes beyond no late fees. And no late fees is kind of the easiest one to explain because everybody has such a visceral 'I hate late fees. It's so annoying.' And we didn't stop there. We don't compound interest. So, it's always a simple interest loan. We pre-disclose interest. So, if you're going to be charged interest, you know upfront how many dollars it'll cost and it can't go up from there. We don't do deferred interest which is where we agree that it's zero for now but if some condition is met it flips, becomes not zero. And so it sort of eliminated all of these things and they were all invented by the industry for two reasons. The typical reason explained is well you know what you need a carrot and a stick. I'll give you a nice rate. Maybe I'll give you even a very low interest rate, but if you are late or if you're not following my exact schedule, you need a way to slap your wrist. And so that's when the late fees come in. It's not actually the real reason. The truth is a little bit darker. And I can very quickly prove that it is in fact the case that none of these are effective mechanisms. So there's kind of three groups of people that are late or flip into this different interest scheme, etc. Some of them are just sloppy and they miss a day because they didn't set the reminder because the calendar changed because their phone was off or whatever it is. They get slapped the late fee. They're like 'I hate late fees,' but they're too honest to say, you know what, I'm not paying. In fact, they know it'll cost them more and be problematic if they don't pay. So they just sort of eat it and like, okay, fine. I'll pay you and I'll pay your late fee and I hate you and maybe I'll never use you again. And that's like 95 plus percent of people who are late. Another group of people much smaller are people who had never intended to pay and you can charge them a trillion percent late fees. They will still not pay you anything. So the sort of story you tell yourself about well late fee will slap your wrist, like there's no risk to slap. They've long left the building. Like they don't care. And then the smallest group but very real are good people that mean to pay. They're not sloppy. They just had something change in their life and they cannot. And that group is actually like they're on their back, like they fell down and by charging them late fees you're telling them like yeah I'm going to kick you while you're down and they're the ones who are definitely going to remember and they're going to say you know what I'm not using you again. Like I sent you a sad letter that was true, I had an illness or I had a death in the family or I lost my job whatever it is and I asked for a hardship excuse or a delay and you wouldn't grant me one because you wanted a late fee. Like screw you. And so, but those two last groups, neither one can pay. So, there's no late fees there. With the latter, you have to wait and then give them a chance. With the former, there's nothing you can do because they don't care. The 95 plus percent of the total late fee content or revolving content with, you know, on these complex interest schemes, they're just sloppy. Like the best thing for them and for yourself you can do is say hey I think you missed a payment, like let me remind you, like let me remind you again. Vast majority of these people say oh my god I'm so sorry, so what's the late fee? Because they're so conditioned by the industry to expect it. And the answer is like Affirm doesn't charge any. And that is the brand-making moment that we hoped for which worked beautifully. People now know that we will not kick them when they're down or certainly when they were asleep at the switch briefly. So that's sort of the why. The whole idea of well you need a characteristic like that, that's all sort of industry telling stories primarily to itself. I think at this point no one remembers this analysis but it is certainly true. The darker truer reason is inside the fine print that explains all these fee schedules typically on the back of your credit card statement is the business model of majority of the issuers where they derive something like half of their profits from late fees, deferred interest, all that junk. And so when we started the company had sort of multiple reasons why we didn't want to do this. One, we wanted to build a brand and we wanted our story to be, hey, you can rely on us not to take advantage of you when you were sloppy or when you actually had something bad happen. Just as importantly, we wanted to attract the brightest minds. And a typical conversation you have with somebody who is in machine learning or data science, really good computer scientist who works on Wall Street. Yeah. So what made you go join a hedge fund? You say, 'Well, I really love financial mathematics, like the bond math, like sort of complicated compounding curves, the sort of really interesting distributions you get out of, like I love that stuff and I didn't want to work in banking because like I don't want to explain myself at cocktail parties like I make money by charging late fees, like yuck.' Like I'd rather go to a hedge fund and squeeze the last penny out of the market for efficiency purposes, which is also kind of a conversation at a cocktail party, but at least you know you're not riding the back of someone at the same party who is currently experiencing a late fee. And so I thought, you know, what if we could steal some of these best minds by telling them you could work in finance, you could build these really complex models and yet you could hold your head up high and never worry about having to explain yourself like, 'Oh, I make money by charging a lot of late fees.' And so that worked beautifully. Like our original team was some of the brightest people I'd met in already by then having been in the industry for 20 years. There are some really really great people and many of them are still here. The third one kind of the most important reason is we thought from the very beginning that our advantage would be underwriting and nothing makes you better at a thing than constraints. So if you look at any bank, anytime they talk about increasing delinquencies, the CFO of that bank typically on a public call will say something like, well, you know, delinquencies are rising a little bit, but it's okay because the late fee income will rise with them. And so we'll sort that out. It's okay like delinquency is rising fine. Late fee will rise accordingly. And at the time, and I to this day, I believe that that's basically shorthand for we can be sloppy. Like we don't have to be that good at underwriting because if we screw it up over here, we'll just make it up on the late fees. It's all good. And we really wanted to be the very best at underwriting. So we said we will charge nothing above what we said we would because then whatever happens to delinquencies will not get an offsetting income anywhere else which will make us scramble and improve the underwriting functions. And that's exactly what happened. Our underwriting is second to none and we've never had the 'oh don't worry about it we'll make it up on delinquencies as the fee income rises.' And so I'm not sure at the time I knew all of these would actually play out as I had expected them to but they did and you know more often than not in my career as an entrepreneur make some assumptions and you know if one out of three works out you're great. And this was sort of three for three. So that's the full story of why and why we're so committed to it.

Totally. It's actually sometimes difficult for me to explain as well as I'd like because for majority of the world it seems that the ChatGPT moments...

The introduction of LLM into the CS vernacular was like this binary thing where neural networks were written off to the scrap heap of history and suddenly they were back. LLMs were a shiny example of how neural networks actually make a difference and would be amazing. If you're in the industry, and I don't think it's unique to us but we were certainly very much in it for a long time, it's always been a continuum. You look for new modeling techniques, you ask yourself what academic papers have been shown to pick up any incremental alpha, where can you find another technique, another idea. So long before, I mean not long before ChatGPT moments but certainly around the same time, we were all reading the same papers asking ourselves what can we learn from these new applications of neural networks. So we're now deploying models that were built with the full appreciation for just how powerful the attention mechanism is. That kind of 'the attention is all you need' paper that came out in sort of 17-18 time frame really shone the light on a very simple architectural idea that allows you, with great computational expense incidentally, to model these fantastically complex systems that have no need for inherent labeling. And we've sort of taken that and ran with it and have all sorts of really interesting ideas that we try to keep close to our vest for a little while before the industry catches up. But we have sort of a whole new family of models that we've been deploying very successfully in fraud fighting in particular, but also applications in credit of course. But long before that, we sort of asked the question, can you find incremental information value simply by outsourcing feature finding to LLM models? And that's been the case, you know, and well understood by the industry at this point, but I think we were there either first or close to first. And so we've been sort of on this journey of whatever the latest hotness in machine learning research is, we want it, we want to be the first ones to try it. We want to run it against what we have today and make sure that we're staying ahead of the train and have been very successful at it.