Will Machine Learning Revolutionize Loans? (Fireside Chat with Paul Gu)

0:00 uh Paul is the co-founder and head of product at upstart which issues personal loans um powered by Machine learning um I met Paul at a different conference Heron um and his talk on like why startup Equity is like worth pursuing do you remember that yeah um that was like great I think it was the best talk of the entire conference uh I had a great time chatting with Paul afterwards and we found out we had a lot in common uh

0:20 like in interest in prediction markets uh namely um he actually offered I think the angel best like even like way back when before we even were thinking about raising a seed round so uh Paul is one of the earliest investors in man um he is also a partner at methos which is an early stage VC firm um investing in pro-social AI um he's extremely knowledgeable uh and I'm so glad to have you here at Manifest this year so um

0:40 give it up for PA thanks happy to be here yeah so I've got a list of questions um they're kind of like um random go over a bunch of different topics that I think Paul and I have like talked about a little bit before and I also want to open it up to you all to like ask jump in feel free to jump in this is a really small crowd so it can be like a little bit more

0:58 intimate um starting with prediction markets um so Paul um I think like the one thing you told me like a long time ago which I still remember is that like at one point in time you were like half the liquidity on predict it or something like that um so can you like tell the audience these guys a little bit more about that yeah sure we we were just talking about that earlier and it seems like you know you you were maybe around

1:15 back then uh too but yeah I mean I think uh it was we were trying to figure out the timing but I think maybe like 2015 or so 2016 there was there started being these like markets that were fairly like fairly decent markets that they started out as these polling markets and they would be like like you know things that like seemed like they wouldn't be like that interesting of questions but it would just be like what's going to be

1:36 the real clear politics like polling average on some particular question and it would be like every week there would be a number or the 538 average or later there were these like how many times will Trump tweet this week and then how many times will these various other accounts tweet this week and um and they were basically like you know a sort of family of strategies that um all were like pretty profitable because the market just wasn't that efficient back

1:58 then so you know you had like you could there literally sometimes you could just make money on straight Arbitrage where it's like you have these questions where you know there's like yeses and NOS or like the sum of the things add up to way more than a 100 or whatever it is and some of those things you know you just you just write code and if if the Gap some you know because of the transaction fees it didn't always work but um but if

2:16 the gaps got big enough you could you could trade those and then you know you could you of course from real Arbitrage you move on to like maybe like stat ARB and so you're like okay well maybe you can't do that but you can kind of like see okay well in these like polling averages there's always these like nine polls this particular poll polling agency releases a poll on average every like three weeks and so that poll tends

2:37 to have this particular skew it's going to come in and do this why this or that and so those things weren't like well baked in sometimes you you just sort of you have family strategies doing that um some of it was literally like sometimes you just have like sort of speed type arbs and sometimes you have the kind of like the Citadel Market making thing where you're basically like basically I I I tried to be like because a lot of

2:56 the markets had spreads of at least a few cents and so it's like if you didn't think I mean the risk was always that like the pole was about to drop and if you're Market making when there's a pole that drops and you're not the first to move then you just get screwed and so you sort of had to be like okay I'm pretty confident that during these hours no polls get released so during those hours I'm always the market maker and so

3:14 you basically just like always have this bot that's like filling the bid ask all the time and then um but then you know you just have to be like you have to also have this other stupid scraper that's checking the the poll releases so you don't get totally fried if that happens and anyway so you had this family of you know regular ARB stat ARB um Market making um and uh and and sort of pure speed um as sort of your four

3:36 kinds of main strategies and you're running all these strategies and then it ends up you just do a ton of Trades um and uh that was sort of my guesstimate is that some large double digit percentage of all all the trades going on on there were um were being done and a lot of that just because you know you're you're looking to make like you know a tiny amount on the trades yeah my average was two cents per trade really

3:58 small another strategy is at least tweet Market in particular were really volatile like I don't you know so there would be multiple different brackets might be like Trump will tweet 70 times a fewer this week all the way up to 1 you know Inc of 10 and let's say normal week like half the time three or four of these brackets are going to hit 70 cents or more people are paying because they really think you know that whatever reason she's not been

4:26 tweeting a lot Market's about to run out and you know people up the one that's currently active and then he you know goes on a tweet St at lasts you know wild yeah yeah yeah the the Tweet storm nature of it made those tweet markets really interesting because especially because like the most interesting moments are like once he does one tweet because then it's like you know conditional and he has tweeted one time in the last one minute there's

4:50 like an extremely high chance there's going to be more and it's like this crazy distribution where you know there's like a chance it's going to be like 20 tweets and so um yeah those those that that was fun A lot of fun um so this is like a few years ago and you haven't really engaged much with um prediction markets at least like in this capacity since um I'm curious uh then that was like trading in public markets

5:11 uh have you considered using prediction markets like internally at upstart um or betting markets or bets uh for running your business yeah um so so I mean the the nature of what we do at upstart I mean is it there's there's a bunch of predicting of of different things that that goes on and um and some of those things are like kind of like the the the right sorts of questions to I mean there's there's things that are common

5:32 to every company so like everyone talks about like oh maybe use prediction Market to figure out like how long a project will take or something like that and um and those seem like sort of interesting but I think they're not that interesting because the number of like people with information is usually not that large so I think it's kind of like well actually only like four of us really know what's going on here so do you really need a market to like clear

5:52 the information on four people I don't think so um so it's like maybe if you're at like Google and there's like 800 people involved in a project there there's something to it but um even you know we're about a thousand people um but you know most projects don't have more than five meaningful participants and it's just like internally I think there's not that much juice to squeeze out of it but then there are all these questions that um that actually are are

6:14 hard sort of like external questions that we've thought we've determined our really like not good machine learning problems so and we we're in this business of we tried to get um risk prediction right on various types of Consumer loans and we try to you know get a lot of edge over kind of what a Traditional Bank lender would do and um and there's there's parts of that that are like the what we think of is the micro problem which is like how do

6:34 you separate and RIS rank risk between like a bunch of different individuals and we we think we do that as well as anybody but then there's like this other part of the problem which is like macro risk which is kind of like okay well there's things that are kind of like highly correlated across like all consumers or at least many consumers that go on they're typically influenced by um things that are um are not highly repeated over history it's like there

6:56 have been whatever like 10 measured recession in like recorded economic history um there's been maybe only like four major bouts of inflation like the number of times that sort of um you know the FED dramatically raises their lower rates is also like on you know you can count them on like two hands and this is like very poor machine learning problem right you can't train any kind of model on this and um and each each event is um

7:20 almost like each event has something unique about it so it's like this most recent sort of like crazy um roller coaster we went through was basically the results of like the the stimulus in 2020 turned out to be like the greatest event ever for um for Consumer loan repayment because basically the sorts of people that are in debt were like the people got the most stimulus and those people suddenly had like an incredible ability to pay back all loans but then

7:45 you know two years later the stimulus was withdrawn it was like the worst sort of Hangover ever because it was like these people had basically gotten used to spending at a certain like rate because they had you know stimmy money and then stimmy money just abruptly disappeared years over the course of a couple months and like of course your like lifestyle habits don't change that fast and then it's like repayment ability just fell off fell off a cliff

8:07 and you could see that in things like the the um the sort of measured personal savings rate and um and that question is is like these questions are like macro questions bad for ML but like there's obviously a lot of information um held by a lot of people out there and so like I think those are the sorts of questions we think there is potential for this and so of course you know we've been talking to you guys about like how we might

8:26 might be able to um to sort of make some public markets on questions that we are specially interested in for which we are not really the special holders of information I see um yeah that totally makes a lot of sense I think uh like you said many uh companies come to us being like Oh I want to forecast on okrs yeah um but that's not necessarily there's not too many participants like more interested in that um so uh any point

8:49 you guys also feel free to jump in if you have like a point on this topic yeah is the this probably basic question what is that default risk is it hedgeable like can you hedge that through swaps or yeah uh it's it's a good question um we've um we've we've obviously thought about this question like constantly and in short like I think unfortunately and I've kind of been disappointed by this I think earlier in my career I sort of

9:12 thought that financial markets is sort of solved hedging problems and it just turns out that actually like for hedging to be really efficient and not like prohibitively expensive it needs to be like an ultra Ultra Ultra Liquid market like you know you're talking about like things like you know there's there's decent hedging ability for like um for interest rate swaps like interest rate swaps are like pretty good a liquid Market even then they're not cheap um to

9:37 hedge um but uh but at least it's not prohibitively expensive but anytime you get into anything even like one step remov from like the thing that like is is most kind of broadly interesting there the the hedging markets become um so expensive that like the sort of fee you would pay to do that is basically eats like you know entire businesses wor of margins and so in short I think we think it's not possible in today's

10:04 markets to hedge this the kind of risk we're interested in because it's not precisely you know it's like if if we're just trying to like I mean actually even if you try to hedge something like um I don't know like uh S&P 500 goes down a lot that turns out to be a pretty expensive thing to hedge um and uh and then if you hedge some combination of that and interest rates and a few other things it ends up being like a pretty

10:25 expensive thing to to pull off and so what we've ended up doing is um thinking about ways that we could sort of restructure the business model to be um to to essentially like Smooth out the kind of like um uh the sort of ups and downs by um essentially um locking in sort of the entities that are taking on the end risk and Loans uh over the course of sort of a full economic cycle so in in the case of our business

10:51 essentially what happened was like the uh sort of Banks and uh and sort of like institutional investors that were exposed to loan risk produced from upstart in uh 2019 2018 19 2020 and 2021 like made uh probably something on the order of like uh twice as much return as they were like hoping to make and then uh from late 2021 through 2022 and 2023 they made like half as much return as they were they were hoping to make and

11:19 if you like invested sort of like a uniform amount across the whole period you would have done really well but of course like you know people like they were doing well ramped up their investing or like new people signed up and so end up having some and some people like stopped for various reasons so have some people that like did really well during a good time and then you have like half the sort of like a

11:36 partner institutions that got completely burned and like now don't want to do business with you anymore and uh and actually like I think the right thing for everybody would have been some sort of thing where um the sort of like uh upside performance essentially like used to hedge later downside and vice versa and so we've been structuring a bunch of contracts to to do this kind of thing it reminds me a bit of like has a

11:59 closed F like person squ and he basically has this Capital locked in so I'm wondering if that might be a mechanism to try get per yeah that's right that's right yeah yeah you you basically yeah you you need logins but then like you you need something beyond the lock in generally because people want like um well full credit Cycles are just a really long time and so you know in the case of um a lot of these funds there's usually some

12:26 way to redeem or maybe there's some liquidity you could trade it to somebody else or there there's there's something otherwise you know at some point the liquidity premium then gets very high and um so you have to pay that so there's um uh but but yes I mean you're definitely right that in in some way share or form like hedging this kind of macro risk is like pretty core to to the business but of course if we just get it

12:45 more right via something like prediction Market it's even better yeah traditionally like one thing that Ki for example tries to uh pitch as like why prediction markets are socially valuable is that they provide this like opportunity for hedging um have you um like thought about this or uh yeah I think it's true U of course like this is is where back get back to the scale R it's like well you know what what would it cost to you know like

13:08 where where can you find a counterparty that wants to like sort of take on oh I don't know like $500 million of counterparty risk on the trade right and it's like they're hedging like a fairly obscure thing it's like we specifically care about like the Financial Health of us consumers in their ability to repay unsecured credit right something like that and it's like well most of the time when you find these like hedging trades it's like um it's it's it's on a thing

13:33 where there is like a lot of parties that are interested in that core question naturally um and like something like interest rates affects all sorts of businesses but this is like it's like a little too specific and so obviously we can't go on Ki and be like where's $500 million of counterparty it's just like you know the the bid ask spread you'd have to take on that would be crazy yeah um bring it over to manifold a little

13:55 bit uh I wanted to ask why did you invest in manifold yeah um well let's see so I mean I've done I've done a moderate amount of like personal Angel Investing and then as you mentioned earlier now do some investing with with Mythos um and um really you know I guess I would say I invest in companies for for one of a few reasons um but I think in the case of uh of you guys it was kind of like uh I I

14:20 definitely like I mean at my core like I'm a I'm a deep believer in the idea of prediction markets like I want them to succeed in the world and I think especially when you're Angel Investing like it's like well one I mean in some sense you're trying to like predict which things will be successful but in another sense you're kind of just like saying like you're you're saying like I I want this thing to exist and and

14:40 succeed um and then you know the other half of the equation is really just I mean I it's it's like so overset it's like trit but I think it's just it's hard to like overstate the importance of just like uh just sort of Founders and the right people and it's just sort of being like hey this is and I think the thing I especially like is when you have like a strong kind of founder problem fit and um and most Founders even most

15:02 good Founders even many Founders I've invested in are like opportunistically working on the problem there's nothing wrong with that I think people should be opportunistic like you know be a good capitalist whatever but it's like you're just like it's just like hey if this year crypto is hot and next year AI is hot like the you know the year the thing is hot and the year you like start your business is going to be like there's

15:19 going to be a strong like determinant of what you work on but then there's like a small minority Founders where was like no regardless of which year it was they were going to work on something like related to this and um and when is a very strong founder problem fit um and and the problem happens to be one that you know you also like really want to see a solution in the world for then you know it's good reason to invest so um

15:40 thank you for that um I guess um then uh since your investment I I'm curious like uh is there any ways in which manifold has like positively surprised you or like disappointed you uh yeah I mean I would say man manifold like I mean it it it went in um in a lot of like directions I I didn't you know I I didn't um I didn't quite like I think the both the positive and negative I guess surprise

16:04 it comes sort of from the same place which is like sort of like this right which is like like you guys actually are like incredibly social people and like you like sort of have this like um large sprawling Network that allows you to try crazy things like the the dating thing or whatever whereas like that was not my mental model of like you know I was just like oh these are like these are like three nerds in a basement they know

16:23 nobody nobody knows them like prediction Market is like this Arcane thing that nobody cares about and like you know they're they're going to be like sort of slaving away trying to find some like weird niche of like a market where there's like a lot of people that naturally have an interest in trading the thing maybe they'll like find their way to some kind of obscure like Financial Security where there's a lot of interest or maybe it'll be something

16:43 in crypto um but it's going to be definitely like three guys in a basement and who never see the light of day and then then it was like then I read these investor updates and it was like there's like a dating show and then there's like all these crazy things and then this this conference and it's the first time I wasn't here last year but it's like it's like a real thing I'm like whoa like these guys like have friends like

17:02 this is crazy um yeah I I don't think we had any of these friends actually um before we started bifold a lot of it is because of prediction Market it turns out that um you can bring like people together when they like care a lot about one project I think basically like the 600 people who are coming here today like all like passionately think that prediction markers are at least interesting if not like are True Believers that like

17:23 they'll change the world so um awesome sometimes you like send out the bat signal and people show up yeah super cool um cool um then uh one story that you told me before but I think is like really cool is uh the way like upart started um originally on income share agreements this thing that like has fallen out of favor I think somewhat um but uh very early on upser I think went through a pivot from in agreements into

17:45 loans right now manifold we're going through a pivot trying to go from just like a play money system into um uh sweep Stakes based uh prize points like cash based rewards U and we're trying to think through like how do we message that how do we get like everyone on board for um mostly the community like our team is like sold on it anyways um can you talk about upstarts pivot and any advice you might have manifold for

18:05 our pivot yeah sure um well um yeah so we started the company back in 2012 um and uh our original idea was to build income share agreements um for kind of like broad any any kind of purpose and so the basic idea was like we will give you know is like people will apply for money just like they're applying for a loan then we're going to run some algorithms and then price the people and so the price would be expressed in the

18:29 form of like we will give you you know $20,000 in exchange for 1.8% of your income for the next 10 years or some something like that right and every person would have like some very specific price and offer um and we did about um maybe uh 300 or so of these things um so like a few million dollars worth of these contracts like was like you know we kind of like got off the ground for sure um but uh but then it

18:53 was also like obviously going too slowly and um I think in hindsight it's like more than obvious that we were like way too slow to Pivot where um we we had this um meeting with one of our um well it was a meeting of the board but one of our investors Josh cman at first round he was like I was just explaining something like oh um you know this quarter we have these like 10 AB tests

19:16 we're going to run to like boost like the sign up rate and it's going to be like I don't know we're going to change the C button from red to blue and then like you know instead of entering your email first in the form you're going to enter it last and like you know all the sort of growth marketing hacks that exist out there I probably read some stupid blog and it was like these 10 tricks always boost conversion by 50% or

19:34 something right and so it was like we're going to do all of the things that you could like find out there on on the internet and then and then you know he Josh looked confused didn't I don't think he heard anything I said and he just said something like um when I invested in this company I thought there were going to be hordes at the gate looking for this money are there hordes of people at the gate and then I was

19:56 like uh no and then um and then that was sort of like after that conversation like oh crap like you know yeah it's like we're not like we're not within like 10x of like the right answer here and um and and it just like sort of was way premature to try to like optimize the thing when you didn't obviously have any kind of product Market fit and you know it was like I think we were just a

20:18 little overly attached to the idea for I think some of the same kinds of like reasons that I think um you know people here at this conference are like really passionate about prediction markets and I think when you're really pass it was like income share agreements had like a similar thing you know it's like sort of this like almost like this religious idea in some sense because it just seems right like there's like you could just

20:36 write whole textbooks about like why it's better than like the Alternatives that exist and uh and back then I used to go to a lot more conferences because you know conference people liked hearing about things like income share agreements and like I was on like every sort of like major like you know TV thing like Bloomberg and CNBC and whatever they all wanted to hear about this but just no real customers wanted it but like you know like Bloomberg

20:55 loved to talk about this thing and um and I think I think that was as as much sort of a a negative signal on the whole thing as as anything and So eventually we sort of I think just in in the nick of time like got our act together and decided okay like okay what what what do people seem to really want and what we had observed was basically there were actually hordes of people at the gate

21:14 not for the income share agreements they were just there being like hey I heard that like you give like money to to young people who otherwise like don't get money like I'm that kind of person can you give me money and then we'd be like yeah yeah yeah like come in and we'll tell you like your what percent of your income we need and they' be like no no can you tell me what APR I can get

21:31 this money at and they would be like no no no see you're missing the point like theoretical benefits of like you know aprs are bad they like you know too much risk you can hedge your risk away if you like instead take the income share because then upside and downside and all these things and um and then people were just like I don't know what you're talking about can you tell me how much money I'll have to pay you back and like

21:49 no no it's a it's a percent of your income you just make the spreadsheet and comes a little higher then you can you can see make make yourself five tabs of you know your if you make a lot of money maybe you don't make a lot of money and then people like so wait how much do I have to pay you back and then it's just eventually like okay you know what just tell them the answer just give them an

22:05 answer they want an answer and listen to the people and um and then um but we were still so attached to the idea so we were like I was like okay here's what we're going to do we're going to do a Sprint to build a standard like loan product that will still be underwritten with sort of these like different algorithms or whatever but we're going to offer people side by side an income share agreement option and a loan option

22:27 because you know some people want each and over time like the market will learn to appreciate the values of the income share agreement and then it will you know the loan will just do some loans to like Bridge us for a couple years and then we'll be back to income share green that was the vision so then we we built this thing all the complexity of supporting this like side by-side product and then we launched it and 100%

22:46 of the people chose the loan and then like literally three days later after you know we had done this like two-month long build Sprint to manage this complexity um I was like um hey guys sorry I think we need to delete all of this stuff over here cuz no one's choosing it it's just in the way and then we deleted it and then and then you know the rest is history and and eventually became a public company

23:08 giving people aprs uh yeah so I think this kind of like thing is like really hard to realize like um because you start with like a mission yeah and then you're become attached to that mission so uh I guess the question is like how did you convince your team and I guess if you had like existing like Isa customers um like that moving on to this different was the right thing to do yeah I mean well we still have the ISA customers

23:31 because they are like you know long-dated income share contracts so there's still people you know 2% of their income like 10 years later so I think they're almost done though I think they should be done any day now because 10 years has elapsed but um yeah yeah you can imagine like when it was like you know by the time like the when we were like doing the IPO docs and like lawyers like what is this like weird like thing that you do you

23:56 are still collecting from like three 300 people like 2% of their income what is this and like like is there a way to get rid of this like no like it's just a there's no way out um so that was a an un so those people never never never migrated off but um but yeah I mean I think the hardest part was was was was migrating sort of ourselves off and there was a real like it was we we had

24:19 we had this offsite where we like debated this thing we laid into the night and it was like well like what are we even doing here like you know and um and and I think uh I think it was there were really like the problem we ended up having to decompose it into like a few sub problems of like well what is the real like problem that we're solving and it was like well you could view the

24:40 problem as like hey uh people are get you know people are getting all the money they want it's just coming in the wrong form which was maybe one interpretation of what we're doing and the other is like adjacent sort of different interpretation was like actually no people just aren't getting all the money they want and the thing they want more is the money and not the sort of like form of the money and um and uh and sort of I think originally we

25:03 sort of believed some combination of the two like um people aren't getting enough money when they need it and it's coming in the wrong form and maybe they're not getting the money because it's coming in the wrong form or something like that and I think we're just wrong about that like people just wanted the money and like they they they could care less what the form was they just wanted to be like the most understandable thing and

25:23 because of you know because that's what everybody's used to like everybody knows like how basic loans work you you know get X and you pay back Y and like that thing is like well understood they want actually the more clear you can make why even I don't think aprs are that great but the more clear you can make why like the better it is for people and um and so I think just sort of migrating from

25:43 like you know hypothesis 1 a to 1B like sort of made made the whole thing work and now like in hindsight it seems very clear was like originally we had sort of you know three three ideas and like two of the three were right and it turned out that was enough to build a good business so out of those first 300 income share agreement customers um like that's a small enough number that I assume there was some human in the loop

26:06 involvement in the early stages so who out of those was like the riskiest one that you were like it wasn't like integrated into your algorithm but you thought that like maybe I should second guess my algorithm here and how are they doing on income share agreement um let's see their name whatever yeah uh um uh you know I I don't I don't know I I think we actually tried pretty hard not

26:37 to like not to sort of like manually take flyers on people because it just sort of seemed like well you know we're not we're not like it's this is not like a VC business where it like doesn't really help us if we like sort of non-s scalably like uh you know put money into some person and then they do really well like it doesn't help us if it's not a repeatable formula so the whole point was we were just like doing like

26:56 repeatable formula things and um and in some ways it did have like some resemblance to like kind of like the early versions of the teal Fellowship which was the program I did because it was kind of like well in some sense like the teal Fellowship was funny because it was like the whole mission of it was like this anti- col Crusade but like the application process at least in the early years for the teal Fellowship was

27:16 exactly the same as a college application it was literally like submit your SAT scores here write two application essays and do an interview that was that was the application process for the teal fellowship and identical I was like wow I'm going to CR like I'm good at College admissions like so this is the same thing and um and so not not surprisingly you know selected for some of the same a subset of the same kinds of people who could like get

27:38 into good colleges and whatever and um and I think um this kind of played out kind of similarly like we were like well okay you might have these people who are because they have this money might make slightly different choices um but in the end I think it it did end up mostly working out that like oh the person who quits their like management consulting job at like Bane like ends up still like doing about as well on average as like

27:59 the people who stayed management Consultants so so I've been a strong believer for many years that the mass affluent and mass Market credit Market is too small because most of the loans houses ORS else yeah and lo are not very large um but then I looked at well when I looked at your website well you manage uh you manage bad debt by using data in

28:31 a very different way than but it was strange to see that your prices not very different your prices at are with with banks if not sometimes higher why is that why did the additional data and kind of new way of doing this which I think is better not reduce the price of these loans given that assum have less bad yeah oh yeah I I think I think the um it's a good question the the simple answer is

29:04 really um simple answer is really just that like prices are individualized and so like there's sort of in some sense like a price for everyone for every type of credit so it's like um no matter who you are there's somebody out there who will lend to you under some set of terms the terms might get progressively more terrible as you become like harder to less legible or or or more objectively risky but there's like always a price

29:25 and so then like when you when you look at a particular ular lenders price the interesting combination of characteristics to look at is like who are the people that they're lending to and what is the price because there's no like it's not like you know there's there's depending if you only lend to like you know if you only lend to like really well off people but you could have like a really low average um average price for your loans if you only

29:45 lend to like really risky people you might have a really high average price for your loans now the interesting question is in like the relative to the characteristics of the pool of people you're lending to how uh how are your prices and um and so in our case like compared to probably like most of the most of the uh sort of like lending type businesses that you would generally hear about um those businesses T typically cater to call it like you know I I

30:10 always hate compressing things back to just credit scores but it's maybe easier for discussion purposes if you were just to sort of compress everything to one dimension of like credit scores like a typical bank for example their average customer is probably about a 780 on the credit score Spectrum I'm an average sort of like fintech lender that you you know you would have heard of would be averaging about 720 um and then um we're serving customers that are more like you

30:34 know sort of like 660 um type customers and so like if you're were to sort of like make a curve of like Market average prices at these different credit scores of course it's like as you get to worse scores the prices get higher um and then uh we're we basically are running something like uh call it up not quite maybe like 40% lower than the sort of line until you get to about uh 700 where we then intersect those line and that's

30:59 uh that's so um basically for someone who maybe has like fairly bad credit like something like a 620 credit score um our sort of like average rate differential will be quite large it would be something like U maybe you might average a I don't know like a 36% APR somewhere else and with upstart you might average like a 25% APR you could say 25% that's still like super high U but 25 is also like quite a bit lower

31:22 than 36 that's that's that's kind of um so I think that's that's uh that's where we add value so this reminded me of something that I was thinking about like two weeks ago with tools like int Credit Karma that like train you to uh bump your credit score yeah on paper but they advertise to people at bumping your credit score as a way to just get access to more money so you can increase your spending yeah like is this causing

31:48 inflation in credit scores like are they more or less meaningful than they were before this phenomenon happened or was like more widespread and what's going on there um yeah there's there's some credit score inflation it's not just the the the main factor is not that the main I don't think the training is like terribly effective I think the main factor is that there's like like when you know people talk about how like um insurance for lawsuits is always getting

32:12 more expensive because like juries are awarding larger and larger sort of like you know uh amounts of money to people in claims courts uh there's like a similar phenomenon in credit scoring where because there's so much like public attention on it it's like sort of regulatory issue a PR issue there's like always this pressure to be like well hey this person only defaulted on their debt because they have a medical emergency we shouldn't count that against them or

32:32 like you know people need a second chance so you got to delete all the sort of bad stuff after a certain number of years or these categories of things should be excluded and so there is like there's actually sort of methodological inflation in traditional credit scores that happens because of like literally every year something changes that like slightly boosts like credit scores and for some people significantly and of course it's just like you know in some

32:53 sense like once you recognize that it's it's all getting graded on a curve like it doesn't do anything to inflate all the scores but that is what's happening um uh for um for us interestingly it it doesn't really matter too much for us because we don't care that much about those credit scores and we've got sort of all the like kind of Base signals so you don't care too much about it but actually the score ends up um being

33:16 really still relevant in our models but often as a um almost like a an opposite adverse selection driven variable where it's like when people have high credit scores you know that like there are lots of other options available to them and then we observe this interesting pattern where High credit score people who take um take our loans are just like worse uh all else equal because it's like if you if if there are like a hundred banks

33:40 that would like offer you a loan why did you like come to us and the fact that you did like there's often something wrong it's like you're like super desperate you don't really care or like actually like some your Banker knows something about you that like is hard to observe and they just won't give you money anymore like and um and so we've got this weird like opposite signal because of this weird selection bias effect is it

34:06 correct your model are SC is there do you see more opportunity to work in that direction to be better at forward looking forecasting whether people would default or not uh and how much can you go beyond what the yeah is it more forward looking um uh I think in a general sense the answer is yes um I mean I I guess it a little bit

34:37 depends what you mean by forward-looking I mean obviously like we are literally trying to just make predictions about the future that's like the only thing that we only dat and is basally picture yeah well I mean I guess all models are are trained on and do inference based on signals that have already happened because otherwise you wouldn't have the signal and all models are trying to make predictions about the future so in that sense like you know I not sure that they

35:05 are like um structurally different uh as models like the credit score what what traditional credit scores literally are is like they just like take only signals of your past credit Behavior I think that's that's the sense in which it's most obviously different from what we do is like they're like we will only consider what you have done in the past with credit and nothing else and then they take those factors and say let's predict if we were to give you a credit

35:28 card what is the likelihood you would become 60 days late on your credit card within uh within 12 months that's that's like actually what the sort of like traditional credit score is like trained and targeted to do so then but then it gets used in all sorts of other applications like and like oh you know hey you're trying to buy a house that is going to be a 30-year mortgage you know that's like uh $600,000 and they just sort of use the

35:50 score that was like trained for this like 60-day credit card delinquency thing which you know doesn't doesn't translate one to one um and uh any so so I think the thing we do that's most different is we consider a bunch more factors and then you know the models themselves are much more um uh they're they're they're sort of able to catch much more Nuance because they're not linear models and um uh and then um and then to your second part of your

36:16 question yes I think there's an enormous amount of room for the models to keep getting better like essentially it's like if you think about how much unexplained variance there is in like who pays back loans and who doesn't like if um if sort of like pure Randomness is if you called that like a zero and like sort of perfect prediction would be like a 100 and then I would say like maybe traditional banks are somewhere at like

36:37 five and we're like somewhere at like 11 um on on this sort of scale so yeah question I questiona do you make the lamb schol yeah um Lambda school um so yeah so we we were really early to the sort of income share game we took an approach that was like we want this to be general purpose you could use it for anything um and then um and then you know I think as as our thing stopped working I think

37:07 then there was like this wave of companies was like okay actually like the use case is going to be for schooling and there were basically two branches there was like the normal schooling people like you know they were going to use it at like Purdue or something and then there was like the um uh the the sort of alternative school people like Lambda and they were going to use it for coding boot camps and um and there was something appealing about

37:27 we thought about this during the pivot uh a lot was like well should should we just do this for the the education use case and the education use case helped with one large problem which was in some sense it helped with the um the explainability and therefore the sort of customer acquisition problem because you know they've centralized it for you and it sounds sort of legitimate that this institution is offering to you um but it came with a major downside which is that

37:50 they sort of had to offer the same deal to everyone and um that was sort of a I think a major disadvantage relative to the sort of original thing because you have to offer the same deal to everyone and outcomes are not um you know completely random like there's there's some amount of like predictability to the outcomes like you know whatever the some some people are like more confident that they're going to like do well after

38:12 the coding boot camp other people L well you have pretty severe um adverse selection problem um and I think um uh you know ultimately I as I understand with Lambda school um I I I I know them like a little bit but as I understand basically what happened is in some sense like this problem forced them to um uh forced them to make more and more kind of like aggressive claims about like what kinds of outcomes you could expect

38:41 um uh from from Lambda school and um and then uh and you know ultimately like it was not possible to keep up with those claims at scale in some sense I think because you know just like like if you read some like Brian Kaplan stuff on like education right you're just like well actually it's just hard to cause make that many people like that much uh more successful uh by sort of pure education intervention strategy and so

39:05 then um so then you're stuck in this world where it's like well you can't actually scalably do it to that many people and then you know then then you start having all sorts of problems so uh yeah vinent is famously um [Laughter] the uh one of the people who wrote uh an expose on Lambda school I think um I don't know you did not know that but I I may have seen yeah okay um but yeah um

39:32 on isas this is also a thing I think I've talked with you about a little bit um I'm still pretty bullish on isas I think especially in the world of like the internet Creator there's a lot more of like a potential market for it where like you now have these individuals who are kind of like businesses on their own they like you know stream they write on substack they like um make Tik Tok videos and they generate a bunch of

39:49 Revenue but um possibly like a lot of Revenue like down the line um yeah do you see like a business there for income share agreements um maybe so I I don't I don't understand uh I don't I don't really know too much about like how the um how the the exact sort of businesses here work I guess maybe what I would say is the things I would worry about are um one is you have to worry about diligence

40:11 transaction costs a lot um which is like if it takes so like for any kind of like thing to work it has to be you know the cost of evaluating whether to like do the transaction has to be a small percentage of the value of the transaction otherwise like the tax is too high and then it doesn't make sense for anybody to do and um and so when when Dave my co-founder and I first met he actually had wanted to build upstart

40:33 as like an eBay like auction for the the ISA and he was like oh it's going to be great you know people are gonna post a little pretty picture of themselves everybody's going to bid and be like you know and and find the market clearing price that way I just convinced him there's no way it's going to work because like it might be like a fun like a fun toy but like you know at any kind

40:49 of as a serious Endeavor like the the the the human cost of evaluating like how good do I think this other human is going to be is like non-trivial it's not like you just like glance at them you have to like kind of like read a lot about them like kind of like figure out what they're trying to do and then like you know you probably put in like multiple hours of time into this like evaluation and um and then for the

41:09 market to work you need a bunch of people to do that now it's like imagine a 100 people have each spent five hours each valuating this person 500 human hours have been spent to do like a single $20,000 transaction it's doomed from the start right um and um and so I I I think generally this is also why like um like there's enduring mystery that people constantly going after I'm I'm an LP in this um this uh this fund

41:30 called Indie VC and um they sort of have this thesis but there's been this constant question like why do VC's why are they only interested in like billion doll outcomes like this is kind of weird right because it seems like you could build lots of good like $50 million businesses and those are good businesses like why is there no like substantial Market of people to invest in those kinds of businesses and basically there's not and and it's like of course

41:52 but now but of course from like a efficient markets sort of perspective it's like well there should be because these businesses have been around longer than these kind of like hyperscaling Tech businesses as like a business model they've been around since the dawn of human history and yet there's no like sort of decent uh sort of network of investors for those kinds of businesses and and why is that and I think at least one part of the problem is this like

42:15 diligence transaction cost problem which is like it takes almost as long to like understand what your like Plumbing business does and whether your customers are real am I going to do like you know Channel checks and call like your houses and be like hey is this guy a good plumber and then like you know it be like how many times more do you expect to use this plumber next year and then call like 10 houses and then I'm like oh

42:32 crap this only added up to like $10,000 of Revenue and I spent you know 10 hours calling these random people about how good their you know your toilet fixes are and um and that um makes uh makes transaction cost very high and um and so I fear a similar kind of thing where it's like if the business are too bespoke then it gets hard but maybe there's something like very recurring about it so like you know there there

42:52 have been some some moderate successes with people trying to do these like things for like factoring type businesses where it's like hey you you know you you just you you basically have like um you have already sold some stuff you just haven't gotten paid yet and like those you know those business there's like Factor financing that's fairly sort of scaled up and successful s what is the factor here that's just the name of this it's basically like you

43:16 um it's like uh you know I sell um I don't know like I sell widgets on Amazon and uh I've already sold these widgets but you know Amazon has this horrendous like six-month payment policy you don't get paid for like 6 months it's it's like I know Amazon's going to pay me 6 months from now but I'm totally broke between now and 6 months from now I essentially can take Amazon's debt to me that they will pay me in six months and

43:39 use that to uh get credit and that so there the whole network of fa factoring businesses that's factoring is a little more than that but um but that's the basic idea you have high confidence in doing a certain amount of business and you sell sort of the the rights to that to get financing um so that's that's a kind of thing like this that has worked um but in general the more close you get to equity the more you need to understand

44:00 the unique characteristics of the business then the diligence costs get high and so you have to overcome that so how is AI not a way to reduce the cost of doing due diligence totally these small BC F focus on the 00 million company yeah for sure yeah I mean I think I think that is like that's that's that's definitely I think like one of the right thoughts to have like I think it's it's it seems right like there is

44:26 there's at least some kinds of these businesses for which it will now be possible to efficiently underwrite them in a way that was not before and um and I think uh it's just like before you could only do this for businesses where you could you evaluate the numbers because the numbers you could stick into you know uh various frame Frameworks for evaluation but suddenly we have the ability to do that with words and um that's that's uh should should change

44:49 some stuff so so someone's going to get a win there uh yeah that was related to the question I had um upstart was pretty early to AI I think and as a result you're like uh upstart is a public company now I think at one point it's kind of like a meme stock almost for people who like wanted to bet on the future of AI they just like bought upstart um with the massive move over to

45:07 the llm style kind of uh uh machine learning which is I think somewhat different than what upstarts really know for like um how do you think your business holds up um what are your like plans uh with this like LM stuff yeah I mean I think I think everything you said is true it's not really like a thing we like think too much about because it's it's kind of like well you know we're all like in the hot meme Target now

45:26 we're not in the hot Target was like that's not exactly what we're optimizing for um so it's sort of fine um but I mean yeah it's it's definitely true that um I we I don't know we we weren't like we never got to like GameStop status but like we were like at least somewhere in the thread like you know if you made money on GameStop where do you go next like there was a thing there um that was

45:47 uh that that was you know it's not really like necessarily a good thing but uh but it was a thing now it's not a thing um uh I would say maybe the real parts of that that we think about are like the thing that really did change from 2021 to today is um is many things but I think one of the core ones is the interest rate environment is very different like I think we're um uh you

46:11 know not being at zero rate anymore is like super different and I think that mean has multiple levels of implications for our business uh and and for lots of hyperscale businesses if you look at like IPOs from 2021 like they're not good you know like the sort of like IP our IPO class like um we were probably like you know one of one of the better performers still in the class we went public at 20 I think we're maybe like 25

46:38 today and you know we had been trading as high as 400 during the sort of combo of like zero interest rates plus like sort of meme Target like was like okay I guess we're at 400 now um and uh but you know 25 is not bad relative to 20 um especially look at that IPO class cuz like everybody in some sense was priced in like a zero like discount rate world and suddenly the discount rate is

46:59 nonzero and it like busts a lot of spreadsheet models I go back to diligence and I'm curious you you're an investor yourself so what kind of diligence that you do and also like given your founding do you think things that you especially um well you know early stage investing uh is um I guess part of the reason people love to do it is cuz like the actually the mo the most important

47:30 thing and the sort of like lazy and fun thing happen to be the same thing which is like like the most important thing is actually just founder quality and like you know people it's like enjoyable to meet great people and so it's like well all you have to do is just like randomly bump around and like bump into cool people and you're like oh this person like this person has has it you know and then it's like I just I want to be

47:51 involved in that and um and I think actually that is the thing that matters in these um again this is a very specific like sort of like VC Circle vend diagram thing of like when you're talking about like very early stage businesses and businesses where it's kind of like the ambition is to get really big or or go nowhere at all and uh and then you know sort of hyper indexing of founder quality becomes so important because it's like all the

48:16 other early signals it's like are only an indirect signal on the founder quality so it's like you don't really care like how much revenue you have after working on this for six months it's not relevant unless you're trying to build your local plumbing business then maybe it's relevant but um but otherwise it's not relevant except that it tells me maybe about how fast the founder moves or um or maybe how resourceful they are something like that cool didet your d uh yeah so um so

48:43 Dave is quite a bit older than me in fact we often joke he's my dad's age and um uh he so basically I had I had dropped it to do the teal Fellowship I was sort of like um thinking thinking about um income share idea probably from like like marginal Revolution or something something like this right and um and then uh Dave had been uh at Google for a bunch of years and um got to thinking about a very similar idea

49:08 and so we like both started working on this thing like kind of independently and as I mentioned earlier he had this like eBay version of the idea and I wanted to like uh build models and price the contracts and um we I I started just trying to like meet people who knew something about this we both did we both networked our way both through like multiple connections to this one woman named no lner who is now an entrepreneur

49:28 in her own right but she had previously um been the CEO of this like sort of like nonprofit thing that um was trying to do income share agreements for education in the US and she was like the one person in the US market who had ever uh done an income share agreement so we both networked our way to her and we both tried to convince her to do the thing with us because she was the only

49:50 person who knew anything about it she had been she had done it for several years she knew lots of the operational messes and she was like I'm never doing this again um I don't recommend it but hey just the other day I met another crazy person who also wants to do this thing that I don't recommend you do and uh and then she introduced the two of us and then uh and then um you know and

50:10 then uh I think my sort of like initial value ad was I convinced Dave that like we can't run the eBay auction on people and then we decided to start working together okay so we only have two more minutes so uh I think just one last question um maybe I'll go down there um do you think the business model by is a bad idea in the but a good idea in practice um is a bad idea in theory but

50:38 a good idea in practice um yeah I mean I do think it's a good idea in practice for sure um do I think it's a bad idea in theory um um what what what makes you say that is there something you have in mind when you say it's bad in theory well I mean what does it do in theory just being alone like that you can't more standard offering oh uh yes I actually think it does answer is May marketing and that's

51:04 why it's practice no no I think I think I think um I think what it's doing actually is um it is so um Merchants are always looking for secret ways to offer discounts that is a core problem of being like a merchant to Consumers you have very high margins generally like if you sell like genes your margins are super high and so you're always secretly willing to offer discounts but you can't because doing so would dis you know it

51:28 basically a failure of price discrimination there's you know like Airlines have some ability to price discriminate but like most products have no ability to price discriminate right like jeans are sold to the same price like to everybody and this is a colossal failure of like of of you know being a genan business you're like I'm just not maximizing my profit extraction because all these people if I double the price of the jeans would still buy it but then

51:48 these other people wouldn't buy it at all so how do you secretly offer discounts well it turns out bnpl is like this super sneaky way to offer discounts why because essentially the bmpl provider is just saying hey I will offer a super low interest or no interest loan to this person of course money is not free there's like there's a real cost in there who's paying that cost it's the merchant the merchant is paying the bmpl

52:09 provider because the Merchant's like I don't care if I lose like 10% of the transaction that's just a 10% discount that's nothing I offer like you know 20% discounts like on holidays all the time right and they're like but I'm only going to now offer this to the sorts of people who otherwise would be sort of un unable to buy this product because they the sort of people who need to like borrow a hundred bucks to buy the pair

52:29 of jeans and I just gave them a targeted 10% discount without being unable to sell $100 jeans to like the people who can afford it so unfortunately we're out of time uh but you can follow Paul Outdoors uh there if you'd like to talk with him a little bit more if that's okay with you if you have more time yeah um let's give it up one more time for Paul goo um all right thank you so much for doing

52:55 this with me I need to go talk so for go for it yeah all right maybe this a terrible idea but