Affirm launches in UK on Amazon and unveils ARC, an attention-based underwriting model that outperformed tree-based models by 2x

Speaker 2: So dropping that down I think is gonna be pretty key.

Speaker 1: Well, Nikita Beer is oh, do we have our interest? It's red. So It was green for

Speaker 8: a second.

Speaker 1: Let's wait. Nikita Beer has oh, I I guess we do have Max. Is that correct? Okay. Great. Let's bring in Max Levchin, the founder and CEO of Affirm. Welcome back to the show, Max. How are you doing?

Speaker 2: It's been too long.

Speaker 1: It has been too long. Let's start with the latest and greatest in in your world in Affirm. And then as always, I have so many I have so many questions about how you're running the company, what you're seeing, what's working managerially on the new technology on adoption side. But first, what's the biggest news in Affirm world?

Speaker 11: Today's news, we are available in The UK on Amazon. That's a massive Huge.

Speaker 1: Not everyone can get an Amazon deal done these days. Congratulations. You have to go straight to the top. Are you negotiating with Jesse?

Speaker 11: No. What what what Amazon

Speaker 1: does feel like a company that might build this themselves. Like, what is your pitch to a big company with a lot of engineers and they have AI tooling? What do you bring to the table when you're partnering with another company at that scale? It's a huge company.

Speaker 11: But it doesn't hurt that we've been partners for quite some time in The US, and and so we're we're definitely no strangers to working with the Amazon engineering team. They're they're excellent. They're very capable of building things. We're a specialist. We know what we're doing in things like underwriting. We have extraordinarily diverse capital markets program that allows us to fund the loans that we we do for for them and for all our other merchants. And so I think every company that's not a financial service specialist at some point or another flirts with the idea of, hey. Maybe we should do this ourselves. If they're not serious about it, then sometimes stick with it. If they're very serious about it and they're of a certain scale, they usually say, wait a second. We should partner with the very best. And, you know, I'm obviously biased, but I think we we've demonstrated that we're we're pretty great. So this is a a great continuation of the relationship we've we've built with them over the years here, and UK is certainly a super important market for us. We're Yeah. Very excited to be there. Also, you know, we we came there a little while ago with Shopify, but been needing to expand the relationship and are excited to be applied with Costco and now with Amazon and many others.

Speaker 1: Having already worked with Amazon for so many years, I imagine that the hurdle to rolling this out is not technical. It's not the actual integration. They They probably have a great team. You have a great team in place. Where are you seeing technical challenges emerge? Where are you seeing acceleration in your ability to deliver a better product? Is it on the underwriting side? Is AI helping there? Or is it on prioritization, conversion, all the downstream customer service? Like, there's so much in the business. What's really moving the needle for you?

Speaker 4: It's like you read our press

Speaker 11: releases. So I'll answer a bunch of them. There's actually a lot of really cool stuff in the question you just posed. The thing that I was referring to, firstly, so we just announced we launched an entire new family of underwriting models.

Speaker 1: Yeah.

Speaker 11: And this has been a long, long time coming. So we are a specialist, specialist specialist. We've been underwriting building underwriting models for fifteen years with, you know, umpteen petabytes of data that we train on. So we've we've we've been a MLIS specialist for a very long time. But up until recently, we primarily stuck to tree based models. They're deterministic. They're easier to audit. They're easier to explain to regulators. We have to do every year. And so all of that has been kind of the the the stronghold of a firm. And about three ish years ago, we said this attention idea that you see in LMs and the transformer architecture is really compelling because it just opens up new ways of capturing complex patterns in a way that humans actually cannot, and fundamentally, improving underwriting models for things like underwriting is expressing patterns you see in behaviors over and over again in a way that can be reused across multiple humans. And so we started an internal research project into using attention based modeling to understand behaviors to surface these patterns all in the service of underwriting people that are figuring out a little bit more about them. And so about a year ago, we had something we thought was really compelling, and we've been testing it quite obsessively. We're finally live as of a few days ago with a full suite of these attention driven models that outperform our own gradient boosted tree based models. And the way I mean, just to give you a sense of just how compelling this this this breakthrough is. So every quarter, we launch a minor addition of the model. Every year or so, we launch a brand new approach to the the core model, all using these three based architectures. We measure the improvement, and that's what we report to ourselves and, you know, our shareholders on. The improvement for this new we call it ARC. So you need the the code name for the architecture. The ARC based model outperformed the next planned improvement by a factor of two. I don't remember the last time I've seen a factor of two implementation improvement. Yeah. So it's it's just very hard to over state how how compelling this is. And so this is I'm very, very proud of the team, and this was a very, very large scale project that was just unbelievably successful.

Speaker 1: That's awesome.

Speaker 2: Talk about what you've learned about the the the timelines it takes for x basically, like, the the difference in execution between two companies to become obvious to the market. And when I say that right now, there's a bunch of new, like, AI companies, for example. Let's say, two vertical AI companies. They both have 500,000,000 in funding. Right now, it seems like like, you know, may maybe there's two ish years where where where it's sort of unclear, like, just how much better is one company versus the other. But over time, you know like one will will surface to the top. And then I would say we've also you can basically see that in every category where there's like there's a category like prediction markets last year. There was like two heavily funded companies and then you saw like difference in execution and they sort of like bifurcated over time. But I'm just wondering from your view how you how you work with your team. Like, when I talking to you, you just get this sense that like competing with you would be like living hell. And it's of the just like the the experience level and then the approach to all these different layers of the stack and the understanding of the category in your business. And like it just it feels like, you know, a firm is just pulling away very very strongly from other players in the market, whereas it looked like it was a pretty even race in many ways, like, you know, five years ago.

Speaker 11: Thank you. First of all, that's a I mean, I I happen to agree, but I'm obviously biased. I do agree that these things take a while to play out, and who knows which inning we're in and sort of how many more sort of ups and downs we're gonna see in the kind of the superficial judgment, you know, aka the stock price.

Speaker 2: But, like, private markets seem to, like, muddy this a lot. It's, like, almost like companies have to get companies have to get public and then have to to actually

Speaker 11: I'm not even sure public markets make it that much better. Public markets force you to be very transparent about a quarterly check-in. Like, you know, one of the things that I think we did really well as a public company, if I do say so myself, is we put a timeline of getting profitable out on the map and said we're gonna get there. And we did it quite far out. So twenty four months before we were profitable, we said, we're gonna be profitable twenty four months from now. And we just printed quarter after quarter after quarter saying, look. We are that much closer. And then on the dot, actually, a couple of couple of months prior, said, yep. Profitable now. Here it is. And I think it was a big credibility thing that private companies don't get to to have because the only people who know their internal metrics intimately are their venture capitalists. And even if they publish metrics, they wouldn't be helped to the sort of a standard GAAP accounting SEC regulated way of communicating. So in that sense, public companies have it probably a little bit easier, but you also get flapped around if you miss on a metric or the market thinks that you you messed up one of your metrics. But I think the way you know who's pulling away kind of early if if if I were, you know, putting on my occasional investor hat, I think,

Speaker 7: you know,

Speaker 11: I I'm feeding your compliment to me back to myself, but I I kinda

Speaker 8: happen to agree with it. I

Speaker 11: think you can tell operator to operator, people who are, quote unquote, for lack of a better term, serious people, you can tell. Like, people who know their metrics, people that understand the entirety of their stack that don't just say, well, you know, I have a great team, and so my AI engineers told me to say these words, and those are the words I'm I'm going to repeat now obsessively. Like, maybe the short hand is, like, companies run by engineers. We we are skilled in not BS ing. And so if that that may be, like, a good one eight predictor. Like, how likely are they just to do what they say they will?

Speaker 1: Like, if

Speaker 11: the person running it has an engineering degree in whatever engineering, probably gonna be pretty truthful.

Speaker 2: Evergreen. Evergreen. I had one more one more follow-up question. I I'm very curious to get your point of view. We've heard a million pitches on this show about how agents are gonna need to pay agents and we need all of this new financial infrastructure. And I've been consistently incredibly bearish on that just because we have a bunch of really robust financial infrastructure that's regulated. We even have companies that like, you know, think about Stripe for an example. They've built for developers at the core from the very beginning, which means that they're inherently well set up to work with agents. And using example, we see new personal agents like Instinct and Muse and there's a bunch of others coming. And there's no point where I'm sure somebody's pitched like a firm for agents, you know, or some some silly pitch like that. But at at with with all these new, you know, sort of applications, the agent will just go and tell the user, do you wanna pay cash or do you wanna use a firm? And like a firm and and they'll just get to Select like they would as a consumer. Mhmm. And so there's no new financial infrastructure needed. And I feel like Agentic payments, like net new Agentic payments might be an entire mirage and we may have gotten a bunch of like, you know, posts online and blog posts and all this stuff and then really nothing new happens. But what do you think?

Speaker 11: I'm gonna make a a bold claim.

Speaker 2: There we

Speaker 11: go. Affirm for agents will be the firm. I'm put it out there. I know it's risky. I I know what I'm saying. No. I I I think I happen to agree with you. I think there are definitely many really cool exciting developments in AgenTig. I am trying out all the same agents myself. Some of them are surprisingly good. Some of them are still lumbering through the same problems you see with some of the earlier attempts, but it's very clear that we will get we will all have agents doing our chores for us. I happen to believe that quite a lot of shopping isn't actually a chore. In fact, it's a form of entertainment. And so human in the loop will not just be a requirement. It will be a loss to humanity if we are not allowed or if we're not participating in some of the shopping choices, which includes, by the way, the way you pay. But some of these things will go to the agent. The underlying plumbing, and by that, I mean everything from deciding the best way to pay all the way down to figuring out the smartest choice of a plan, most rewarding transaction, best 0% loan, etcetera, I think that's going to primarily accrete to people who know what they're doing. We're specialists in the space, and that's why we have to continuously work on improving underwriting. We want to be more inclusive as in say yes to more people while maintaining the same level of credit performance. And so all of that is still, like, the work we have to do, and we have to do it faster, and we have to pull away from the competition as aggressively as we can. But I don't think there's an opportunity to dislodge a firm by showing up and saying, are just like a firm but smaller, less profitable with less credibility in the market and the capital markets in particular, but we are agentic. Are agentic too. We're pretty pretty agentic ourselves.

Speaker 1: Yeah. I know. I love it. That's a great take.

Speaker 2: How how have you been approaching leveraging open source models in various sort of like employee use cases and workflows? I think it's been you guys are such an you know, incredible engineering culture. I'm sure a lot of your team has been using open models in a bunch of different ways. Yet at the same time, if you were focused maybe five months ago about, you know, building your own harness or or using these harnesses and open source models, and then the cost of the frontiers drops like so dramatically to the point where you now have like frontier ish models that are cheaper than open source in in some cases. Maybe that wasn't the best use of time. So like, how are you thinking about allocating time to getting the most out of open models where it makes sense versus trying to avoid just wasting time when the cost of intelligence will continue to fall?

Speaker 11: So we actually did something pretty smart, if I do say so myself, pretty early on. So I'd sort of predicted that we're going to go through these moments where, like, oh my god. The best harness, the best model, the best the the combination of harness model, user interface is going to change. And there's so much money. There's so much innovation. There's so many really brilliant people who are working all day every day and making AI useful specifically for software engineers. It is foolish to commit to a configuration today. You know someone else is looking at it and saying, wait a second. That is the best way of writing software except I have a better idea. And writing software just became the best it's ever been by the hands of the company I'm about to compete with. So, like, the whole like, the self recursive self improvement that everybody's sort of either excited or terrified about, It hasn't come to the models yet, but it's certainly come to the development industry. Like, we are living through recursive self improvement of software engineering for humans and agents together. And so sometime around January of this year, we split off a team of about 12 people and basically said, your job is to make our development experience the absolute best for the current state of the art in a way that is easy to take advantage of now, but switch out to the next best thing later with a thoughtful continuous matter. So we don't want to have this disruptive moment where everybody stop. We're all gonna switch to product x. Oh, wait a second. Product y is available. So we have this team, and it's it's run really, really well by a bunch of very, very smart engineers who love their craft and know what they're doing as practitioners, but also great thinkers when it comes to developer experience. They have been keeping us at the almost the cutting edge of both the commercial frontier models as well as open weight models, harnesses, etcetera, where we organize the entire process through with our hands. And whatever it is they offer to the entire company is usually within a hot second of whatever is considered cutting edge, but it's thoughtful enough where if you yesterday, you were on harness a and today we really believe harness b is better, they will need the transition really simple. So just having a dedicated team that gives us the best possible developer experience without having to do a handbrake turn every three months has been unbelievably good investment. Like, when when we walked off this team and said, we're gonna have this big group of people whose only job is to make us more productive at the meta level, I think some people were doubting the validity of the idea.

Speaker 2: Yeah. It's interesting because the alternative is similarly sized companies, you have hundreds of people that are experimenting in real time and be like, well, I think found the best way to do it. And the other person's like, well, I'm using this thing,

Speaker 11: and then it's Start it there. Whiplash. So I'll I'll give you the real stats on this one since I'm a I'm a fan of numbers. So we were in the experiment away mode until we have this developer experience team, developer productivity team, and we were probably I think the percentage of code written by machines and humans together versus prior to this team's arrival increased by a factor of 10 when we organized the team and said, look. Here are the prescriptive approach we're gonna take. And there's always a menu. Like, you can use cursor. You can also use Clodder. We support all sorts of different harnesses and models, but we have a menu versus go figure out what works for you best. The tyranny of choice is a terrible thing. And telling a software engineer, go explore over the weekend your favorite way of writing code with an agent, it's not gonna be a weekend project. It's gonna be a six months long project. So lopping that off into a separate area where you have a rigorous approach, and then we constantly produce, here's the best way according to this team, and here's some of the choices you have in there has been really, really useful. We know it's doing well for us. Our so we measure productivity long before AI in PRs, pull requests per engineer per unit time, the cost per PR fully loaded everything from salaries all the way down to AWS costs has come down 30% since we created this developer productivity team. And so not only are we increasing the amount of code we're writing because we're able to leverage all the agents, the true cost per pull request is coming down quite steadily and has been for a while. And so

Speaker 2: I'm Very cool.

Speaker 11: Very excited about what's to come there. But I love the fact that we have this really well constrained approach.

Speaker 1: I love it.

Speaker 8: Makes a

Speaker 2: lot of sense. One, one word answer for the next one, since you like numbers. What's your p do?

Speaker 8: No answer.

Speaker 2: No answer. Alright. We'll get to it next time. Next.