In this episode, David Lykken sits down with Pavan Agarwal, Founder and CEO of SunWest Mortgage. He is also the visionary behind Angel AI. A pioneer in AI-driven mortgage solutions, they discuss the latest developments in artificial intelligence and its impact on the mortgage industry. They delve into a newly published article that highlights the inherent biases found in large language models (LLMs) like ChatGPT, particularly in lending decisions. Pavan provides an in-depth analysis of how Angel AI, built on a different architectural foundation, overcomes these challenges to ensure fair and unbiased mortgage decisions. Tune in as they explore the critical differences between various AI technologies and the future of fair lending in a rapidly evolving landscape.
[David] Welcome to another podcast on AI, there is so much developing and I want to talk about the latest article that just was published today. We’re recording this with my good friend Pavan of Angel AI. August 21st, at the same day the article came out. Welcome to the podcast Pavan. [Pavan] I’m so excited to be here, David. I give the credit to my daughter Alina. Found the news piece. It was published I would say four hours ago. [David] Wow. This is fresh off the spot and it really hits to something that is a real priority for you and it’s fair lending. Getting the bias out. Let’s first of all talk about what this article is setting forth that they’re saying in this article that there is in fact a bias because of LLMs, Large Language Models, and that’s not what you built Angel AI on. So, let’s first of all, lay a foundation down for our viewers, our listeners, who are listening to this, about what is an LLM and what are the perils that it is bringing to the AI, at least as it relates to underwriting. [Pavan] There’s a lot to unpack this. I’m going to start with what is an LLM and what the researchers did here, and it’s fascinating. It’s so amazing what they did. First of all, Large Language Models basically think of it this way. It’s like they trained it on chat GPT for the latest version of ChatGPT and that’s what they tested it with and it’s like a massive garbage cleaner. It’s like it picked up everything humanity has ever done and loaded it into its database that is freely accessible on the internet. All data on the internet. Basically, ChatGPT and other AI companies are saying, Hey, we’ve got a problem now because we have loaded every word that has ever been written by humanity which is available, which we can get to in the public domain that we can get to on the internet. We have loaded everything into our AI training database. Okay. So the number one, let’s start with that [David]. When you say we’re, you’re talking about what other systems are doing. [Pavan] Yes. I’m talking about ChatGPT and Google’s system and other, Elon Musk was working on one. [David] What you’re saying. I just want to slow it down a little bit to make sure it’s getting across. What you’re saying is they’re taking all the knowledge, which had a lot of bias in it and load that into their system and now they’re starting to roll out answers and guess what? It’s not producing good results. [Pavan] Because it’s garbage in, garbage out. You said it very well. They take everything that’s available, in this case about mortgages, that’s publicly available on the internet. They took everything about mortgages. All right. ChatGPT has everything that it can possibly find about mortgages and it’s in its database and that includes everything that’s publicly available about loans that were approved, loans that were declined and why they were approved and why they were declined, right? And remember, LLMs, Large Language Models are artificial ChatGPT, all these kinds of artificial intelligence. They cannot reason, they don’t understand the data. They just repeat patterns. They just take past patterns and repeat it to new requests. What this researcher did was said, okay, he took thousand sample loans. He’s created a simulation of thousand loans based on HMDA data Which is publicly available remember the same HMDA data as was already loaded into ChatGPT. It’s already trained on that. If ChatGPT is going to get trained on mortgages, the first thing is going to use is HMDA data, right? That’s the most important piece of mortgage data that’s out there. So this researcher took a thousand samples of HMDA data, scrambled it, making it unique, putting new ratios, new kind of information and then ran it back through ChatGPT and asked, would you approve or decline this loan? and in that data, the researcher included the race of the borrower and what it found was same thing that we found, our results are nearly identical, ChatGPT declined black borrowers about twice. Maybe more than twice that of white borrowers. [David] Was it because they were declined because they identified themselves as black or was it in the data and just the fact that they identify themselves as black didn’t matter? [Pavan] The data that the researcher fed into ChatGPT, when he asked the question, would you approve or decline this loan, the data he fed into it he included the race of the borrower and ChatGPT looked at all of the data, including the race of the borrower, applied it against previous patterns. This is the key. Remember, artificial intelligence is nothing more than taking a previous pattern and applying it to new data. It doesn’t reason, it doesn’t think, it just takes old patterns and applies it to new data and repeats the patterns. [David] So, the outcome here wasn’t terribly surprising to you. This new article that’s supposedly releasing this latest data. You’ve been a researcher and a developer in this space for decades. One of the forerunners on it. and so this article did come as a major surprise, but the way it’s being rolled out, I’m concerned that some in the industry, some of the trade rags might pick this up and say that AI is not going to get us the advantage that we had hoped. And that’s not true. [Pavan] That’s not true at all, the article does confirm is what I’ve been saying all along is that the industry is riddled with shenanigans and bias because remember, it’s not AI’s fault. It’s not ChatGPT’s fault. It’s that data, historical data. They just proved that historical industry data has a pattern of declining black borrowers more than twice the rate of white borrowers. [David] It’s just astounding to me that exists, but I grew up in Minnesota. There was not the level of racism, anything close to what’s out there, but it’s just really interesting. What are some other things that you gleaned from this article that you think it’s important for our listeners to know, and then I want to go into contrasting it to how Angel AI has been built differently, but what are some other salient points about this article that’s interesting that might get misconstrued when someone reads it? [Pavan] Two things. Number one, it focused on only using ChatGPT. ChatGPT is a class of artificial intelligence technology called Large Language Models. Meta has their own version of ChatGPT, Google has their own version, Amazon has their own version they bought a company called Anthropic with a system called Claude. And they’re all basically large language models, it’s a type of artificial intelligence. And it’s a type of artificial intelligence that takes patterns of words and applies it to new words that you give it and it looks like magic, but it isn’t. It’s just repeating patterns. Number one, the test was specifically on one type of artificial intelligence technology, which is Large Language Model. Unfortunately, this is the type of artificial intelligence technology that most people today understand. People assume that artificial intelligence is ChatGPT. When the answer is no, ChatGPT is a type of artificial intelligence. [David] Explain that a little bit more in depth if you wouldn’t mind, If you could start separating this out a little bit, it’d be great. [Pavan] Yeah, before I do that, what he did was he ran his test thousand loan test case, thousand sample loan test case against this engine. And when he gave it the race data for the thousand loans, then he saw that it declined the black community twice, maybe more than twice than a white community. But when he took the race out, then it was even. When he took the race identifier out, then the loan that it would have declined for a hypothetical black borrower, it ended up approving. [David] Wow. That is really disconcerting. [Pavan] That just tells you, the fault isn’t in ChatGPT. The fault isn’t in the AI. The fault is in humanity. It’s in the system that generated the historical data that the AI was trained on to the systemic bias. And all he did was prove that there is systemic bias. Artificial intelligence is a great tool because it puts a lens on humanity. It’s a great reflection mirror. Now we get to see ourselves. We’ve taken all of our information and put it there. Now we can look at ourselves and say, Hey, we like what we see. Exactly. I got a mole on my face. Where’d that come from? [David] Yeah, we’ve gotten accurate mirrors for the first time. We’re really getting accurate mirrors. And it’s just really interesting about this and so they took the thousand loans, ran it through as the 2022 MORA HMDA date is the data set that they use. And it’s just really interesting. This is based on experience results using Open AI’s, ChatGPT, Turbo, LLM, Explain what that is and how that is different from Angel AI. [Pavan] That’s a great question. First of all, Turbo LM, ChatGPT, they keep coming out new versions with new capabilities and they come up with new, just you have the Toyota Camry and the Toyota Camry SE and the LX and whatever. They have all these different extensions for the newer versions, which is good for them and the large language model, the way it works is it takes all the past data, all the text in the universe on the world, and it stores that information and it figures out all the patterns in them and then when you ask it a new question. It says, Oh, that’s a new pattern or that question is close to this pattern that I already have in my database, but you’re typing here it looks just like this thing that’s over here. I’m oversimplifying it but it says, I asked a question about war and peace and it’s going to say, Oh, there’s a novel called war and peace. I’m going to go grab that novel and start printing it out for you. That’s all it’s doing. In this case, because he’s using, remember, large language models cannot reason. They’re not alive. They can trick humans [David] Yeah, because they’re not reasoning. They’re just simulating data, processing data that they’ve been fed. [Pavan] They’re just repeating past patterns. It’s just that it looks like they’re reasoning because they’re loaded with trillions of patterns. More patterns than a human being could ever remember and they find a new pattern that you’ve never thought about, but it’s in their database and then they apply it. [David] Interesting. There’s so much interesting information about this. What I thought also was interesting in the article is that the models also exhibited a bias against Hispanic applicants, but generally to a lesser extent, I wonder why that is. I started wondering about it. They go, it’s must be in the data. It just must be inherently in there. [Pavan] In our own comparison, when we compared the results of Angel AI against our AI which is not LLM based and it cannot be, and we’ll go into that, why it cannot be. Our AI showed the same thing. We approved twice as many loans than the industry average for blacks and about 1.5 times as many loans for Hispanics and what’s even sad and I don’t think this researcher did it. What we found is that for veterans, if two black borrowers were declined, three veteran borrowers were declined, it’s another level for veterans. It’s insane. [David] Wow. Wow. That’s crazy. And not only did they use chat GPT, I’m reading in the article here that they used other AIs, Opus, Bettas. Lamas, I’m bringing stuff up, but I don’t even understand what it is that I’m trying to get, [Pavan] It’s different brands, right? You have a Toyota or a Honda or a Ford, they’re all automobiles. [David] They all still work in the same genre. [Pavan] Yeah, they still have an internal combustion engine, they still have wheels, they still have, suspension. They all work basically the same. They’re different brands of it, different flavors, but they’re basically the same. And so it’s not surprising. The results are the same. [David] Yeah. Yeah. Generally, they were the same. There was some higher grade discrimination with ChatGPT for whatever reason that is. Let’s get into how did you avoid this pitfall, Pavan? This is so significant. When people are starting to, in your name, it’s Angel AI and we have to be very careful that not all AI is created equally and this is where I want to go with this interview. [Pavan] Not all AI is Angel. And not all AI is created with God in mind and with the goal of doing God’s will, right? and which is ultimately that we’re all God’s creatures, and we need to respect and love one another. You got to start with that basic foundation. Okay that note, there’s a lot of companies right now that are rushing to build Some kind of AI capabilities. [David] And I applaud them. You can’t fault them because AI has become the center of the universe right now. Everything’s revolves around that conversation. What are you doing with AI? So you understand, but it’s the pitfalls in that gets people in trouble. And I get concerned about people starting to turn away from AI based on this article and that’s going to be a big mistake. If they don’t understand that. [Pavan] That’s why I said in the beginning, all AI is not ChatGPT, all AI is not large language models. But unfortunately, from what I’ve seen, all banks and lenders that are building AI, or some kind of AI solution, are building it on top of one of these LLMs. Because there’s this idea, because LLMs are so remarkable. Don’t get me wrong, they are an amazing piece of technology. It’s one of the greatest human inventions ever. [David] It’s amazing. The potential for it is tremendous. [Pavan] But it’s not the end all. it’s not the solution for everything. It isn’t the God particle, right? It isn’t the God program. LLMs cannot reason and anybody who’s ever written a loan file understands there’s a level of reasoning involved. Let me give you an example. What does that mean? You can ask ChatGPT or any of these LLMs, ask them this question. You can all try it out, it might work now because this has been around for a while. I didn’t make up this test. Because it’s been around, so it gets absorbed into this database and now it learns how to do it. If you said I have a clothesline and I’m drawing three shirts on my clothes line, and it takes one hour to dry three shirts, how long would it take to dry nine shirts if I hang nine shirts? So, the LLM would take nine divided by three and say, okay, three hours, right? But the LLM cannot reason. It doesn’t understand that when you hang clothes on a clothesline, It doesn’t matter how many clothes you have, it’s going to take the same amount of time. That’s reasoning, a human being can reason and LLM cannot reason. Now, if you prompted correctly and you said the amount of time, when you dry clothes in a clothesline, it doesn’t matter how long you hang it. It doesn’t matter how many clothes you hang, it takes the same amount of time. If you gave it all that information, then it would probably give the right answer. But that kind of missing information, that subtlety, it’s not really missing. You can infer that information based on what you got. That kind of inference, right? That comes from reasoning, right? Underwriters and all of you watching who’ve unwritten a loan file, have been involved in the loan process. There’s a level of inference that happens when you underwrite a loan file. [David] I haven’t had to do the underwriter for years, so I understand that. that’s true. [Pavan] If you have two paychecks with overtime and one paycheck without overtime, You can’t just assume that his overtime’s ended. Maybe he was sick. Maybe he got a promotion and now he’s not eligible for overtime. Maybe the base went up. There’s all of these different things that could be happening in there. Whereas, without that kind of intelligence loaded into the model, it’s going to say there’s no more overtime, no more overtime, right? Instead of maybe asking for why has the overtime stopped or something else, it would just say overtime, no overtime, so we can’t use overtime anymore. [David] So, the problem sounds like from this article, it’s the LLM and some of the things conclusions you could draw from the comments you’ve already made. You say the problems of the LLMs is the problems in the data, how the LLM, the large language model is being employed. I sound like I know what I’m talking about. This is your world. Please take us through this. [Pavan] LLMs are great for a lot of things. But making financial decisions, making absolutely like hard decisions. We’re going to turn right, turn left. We’re going to give this guy a million dollars, or we’re not going to give him a million dollars. These are big decisions. They’re not good at that. It’s the wrong tool to use for that job. [David] What is the right use of an LLM? [Pavan] If I want an LLM to draft an email, a thank you letter for David, thank you for having me on this podcast, it can write that letter and it can make it sound like as if I wrote it. Any kind of generative work, anything that’s relatively simple, that’s routine. LLMs will automate a lot of routine, simple routine work and when you see most people utilizing LLMs in their lives, they’re using it for that, a lot of routines and work and graphics designers love LLMs, like ChatGPT because it does so much heavy lifting that would take you hours in Photoshop that you do manually adjusting pixel by pixel. LLM would do it in a few seconds. Anyone who’s worked with any of these kinds of graphic systems or written emails and stuff knows that there’s so much redundancy in a lot of work that you do. LLMs eliminate redundancy. [David] When you gave us some examples of very simple projects, although we value those, it does help save a lot of time, but underwriting is not simple. It’s very complex. There’s so many aspects that go into that. Is it accurate to say that an LLM won’t work or will work? if it’s properly deployed [Pavan] To prove a negative is hard, I’m going to stay away from saying it won’t work. I won’t be able to prove that. Based on my understanding of how they work and I’ve looked through the code and I know how they’re structured and I don’t believe they’re going to work. We build angel AI before we had the first version running, I think it was ChatGPT 1.0 when it was barely functional. LLMs were just something of the geek squad messed with and no one knew about it. It was up in Nerdville and where I live in Nerdville, that’s where LLMs were back then. [David] By the way, for all you geek squads out there and nerds, we love you because we value you. I’m on my laptop right now because my computer is in with the geek squad at Apple right now getting some power issues addressed so we’re grateful for the geeks out there, but you’re really bringing out a really good point, Pavan, that we need to dive deeper into on this because you’ve approached this differently. You’re not going to say LLMs don’t work. I understand that. So, how is Angel AI working? [Pavan] You gotta use the right tool for the right job. You don’t take a hammer to do a solid job, right? If you’re a carpenter, you don’t use a hammer to cut a piece of wood. You use a solid to cut a piece of wood. Yeah, it is it’s the same thing. It doesn’t make the hammer worthless. I know. A hammer’s incredibly valuable, it’s the same thing here. Now we have to go back and peel back the layers and say, okay, what we’re solving a different problem. We’re working with loans that are really complex. And if you just take a look at one simple example, let’s say I’m in a loan file and I’m going to change the loan amount and I’m going to change the interest rate at the same time, right? and you know how complex that is, if you change the loan amount, you’re going to affect the qualifying ratios. You might affect the high cost test. You might affect your lock, You could end up kicking it out of the conforming product and it could become a jumbo now. Million things, right? and the same set of dominoes fall when you change the interest rate. So that’s why in all of you watching and whoever originated a loan, you know that when you go into your loan system, every single loan system out there. If you go into it, they lock everyone else out. Only one person can be in there doing their things and then when one person is doing their things, they’re doing their work sequentially. That’s how human beings can only work sequentially. We don’t, human beings don’t do five things on the loan file at the same time, right? If you just think about the fact that if you tried to change loan amount and interest rate at the same time think about how many rules get fired off at the same time. Because every loan system, every modern loan system has a rules engine. Behind it and those rules engine run in sequence and they’re very delicate and we all know what a mess the industry went through when trade were rolled out. People were months behind getting trade rolled out. In two, three years ago, when Fannie Mae rolled out the new 1003, nobody, but us got it done in time. We were like four weeks ahead and our test cases passed the first time and it’s because everybody’s rules engine is spaghetti and they’re delicate, one rule affects another rule, and so you make one little change, it takes an immense amount of expertise and knowledge to make sure that you don’t break the other rules and that’s why loan systems are so expensive they’re expensive to purchase and they’re expensive to maintain. Because you have a whole army of people maintaining those rules and if you were to take just two variables, loan amount interest rate you can’t change those at the same time without, going insane when I think about trying to catch all the balls in the air that thing pulls off. Now imagine changing a hundred things at the same time or a thousand things at the same time, it’s hard to imagine just two things, two people being in a loan system changing two things and without those rules trampling on each other. Now imagine a hundred, a thousand, ten thousand. Can I present a slide? [David] Oh yes, absolutely. [Pavan] You see the screen. So traditional loan processing on the left is you got one person in at the same time and everyone else is locked out. And AI driven and Angel AI, which is all AI doing this. We have thousands of bots in parallel in the system at the same time. [David] And you avoid the problems. That’s staggering that you can do that. I’m locked in the old, I’m locked in the picture, the image on the left, where it’s everyone’s locked out, you only work on this part, because of the sequential chaos that you create, by trying to visualize the picture on the right. [Pavan] In order for you to have a really working AI system, an AI loan system that’s fair, that’s going to deliver true fair lending, right? You need the architecture on the right. [David] I get what you’re saying, but explain a little bit more for our listeners that may not understand why that is. [Pavan] Here’s what happens in real life, right? you get a pay stub, let’s go on the left, You get a pay stub and you go sequentially, it reviews, it analyze, then a whole bunch of rules fire off, right? and these rules are usually written in some kind of a rule language, right? and then the database is updated and other things happen and then you’re done. Then you start all over again and it’s very sequential. And now these rules that you need a team of people constantly maintaining the rules and if something changes and those rules can break and you got to start all over again and it’s why every release of one of these systems is so painful. Because they’re so fragile. On the left is an AI driven system, where an AI system is saying, okay, it’s not thinking rule centric, right? It’s thinking data centric. It’s thinking, okay, it’s going to keep track of all of these interconnections, right? One updated pay stub, I’m recalculating income, I’m updating the database, I’m updating financial, I’m updating the clear to close and I’m doing the regulatory checks and all of these things, I’m doing these bots all simultaneously, right? one pay stub update, I’m going to spin off hundreds of bots in parallel, okay? and they cannot collide. If you think about all of the way these hundreds of bots can connect to each other and talk to each other because they have to constantly if you don’t know how to do it, you would write a bot that would have to talk to each other about everything that it’s doing, right? and the amount of permutation and combinations is something like the 10 to the order of a thousand to the power of 1,000, which is more than the number of atoms in the galaxy. That’s why AI needs so many microprocessors, what they call GPUs, so much computing power, because AI systems have to calculate through all of those variations, right? Find the patterns and find what they call, the technical term is gradient descent. Find the lowest point in the data and say, okay, the lowest point is probably the right point. If you’re going to walk away with some technical terms. GPU, gradient descent, right? [David] Gradient descent is the one I’m walking away [Pavan] and dimensional space, because all that data is represented in a high order dimensional space, every piece of data on your loan file another technical term you can walk away with is represented as a vector. It’s a vector in a hundred plus dimensional space and then you multiply all those combinations out. It takes a massive amount of computing power, if you’re not going to be rules based, you need all this computing power to test all the combination and find the one point where all the vectors intersect and say, yes, we all agree on this one point. If you think about it, loan amount’s pointing this way, and interest rate’s pointing this way and fair lending is pointing this way and high cost is pointing this way, right? But there’s a point where they all intersect, right? [David] I’m trying to get my head wrapped around. When I’m looking at this it’s just the one on the right looks like chaos, but it really isn’t. It’s the way of the future. It’s a way things are handled and that’s those that are operating in the old linear way on the left hand side are going to be just left in the dust with what’s happening Now, what about speed? How is the speed of which a decision is rendered. Angel AI underwrites a loan, if you give it an approved except if it’s gonna fund you’re gonna fund that and you haven’t made any mistakes But if it were to, You’re indemnifying that, if they fund the loan through your system. [Pavan] Exactly. I am that confident about, see that spaghetti that you see on the right? It’s so well tested and has been doing it for so many years and I’m so confident about the algorithm that it always finds the intersection. Ultimately if you draw a bunch of straight lines in a hundred dimensional space. The right answer is one where they all intersect and at a single point. And AI is about finding that single point which is defined as the lowest point. That’s a gradient descent problem. So, if you could find that lowest point, then you can say, okay, if I change loan amount, if I change this, upload this pay stub. This is where everything else has to be to have the right answer. [David] Interesting. So now let’s relate this back to the article. That was published today that we’re really addressing how does Angel AI deal with this problem or handle this situation on the right hand side? [Pavan] Yeah. It does it because it’s not using large language models, it’s our own proprietary technology. Again, we rolled this out when large language models were babies. When no one knew about them, there was ChatGPT 1. 0 and it was just an experiment, and it was already working. It was actually working long before that even. This engine that you see here was working years ago. We rolled this out a long time ago. We figured out how to solve this problem of finding the intersection dynamically and so because we can find that intersection, we could find it dynamically. It doesn’t matter what type of loan you put in there, we have all the data about regulations. We have all the data about the loan products and everything in all these vectors we have in there and then your loan file is another vector and we find the intersection and we do it like in a microsecond. [David] Yeah, it’s the speed at which it is determined that is just such a game changer, you’ve been doing a lot of speaking here lately. What is some of the things that you’re picking up from the audience? What’s interesting, the number of realtors that are showing up at your events is really impressive. They’re catching on to this. I’m wondering if the mortgage lending community is catching on to it at the rate the realtors are. [Pavan] That’s a great question, unfortunately I’m seeing the realtors way ahead of mortgage bankers on this. Way ahead of it. And I think it’s because mortgage bankers are scared. [David] We’re going back and talking about bias. I’m going to go outside and use an example, one time I was given a gift, a friend of mine, one of my family members won at a benefit or bid on an air to air combat. What we did is we went up and we trained as pilots, combat pilots. If you imagine for a short period of time, we went up and two Italian Marchetti high performance airplanes and we did a dogfight for an hour between Long Beach and Catalina. They gave us a cube of airspace out there. What’s so interesting about that and what I’m trying to go with this story is the wife, and oftentimes what goes out is there’s an experienced husband and his wife is flying against him and what’s interesting is the experienced pilot, a mortgage banker has all this preconceived training. You can’t go this fast. You’ve got to approach a bank and then turn like this. A wife who has no airline train or no flight training kicks the butt of an experienced pilot out there because all they want to do is just go kill their husband out of their high and have bragging rights out of it But I think we get locked in and it what kills the Pilots is preconceived ways we’ve been doing this. We can’t turn this fast. We can’t bank a plane this fast. I think that’s what I’m finding as I talked to more mortgage bankers, and that’s why I’m excited to have you on. A lot of people are commenting, Dave, you have Pavan on a lot right now. I go, it’s because of how rapid things are changing, how important it is people get this message, Pavan of what you’ve done and others are trying to catch up and I applaud anyone who’s getting into the space. Everyone get at it. We need to have everyone on it and the best thing I love about you is you just want everyone to be aware of what is available through artificial intelligence. Now you do have a decided advantage. You’ve been working on this for a better part of 40 years, but dedicated for 24 years, I believe you’ve been working on this uniquely positions, Angel AI. In a place where a lot of people are just not going to be able to catch up. I think you need to pay attention to listeners to what is going on. We’ve got too many testimonials. We’re going to be sharing more testimonials on this podcast of people that are having extraordinary success. We now have two mortgage bankers I’m talking to that have shut down their existing tech stack and have moved over to angel AI because it is so fast. It’s so affordable and it’s just playing in so well to where the market is going. [Pavan] I’m going to wrap it up with my favorite Tom Cruise quote, which is mortgage bankers and realtors in particular. Have the need for speed. Bring it full circle to your pilot analogy there, metaphor. The need for speed really is what realtors are gravitating towards and the speed of them getting answers and more importantly, is they love the mission. Okay they’re working with the free credit repair. They’re working with the free loan approval, they’re working with the free lead prospecting, the free websites. See, Angel AI, I purposely designed to be an open platform. It’s not closed, right? There’s no paywall. It’s open, right? that means everyone can use it. Everyone get in there and it’s like with any other open source or open source project like with Linux. Linux became the uniform global standard now. Everything runs on Linux. People laughed at it because it was open source, 30 years ago, 40 years ago. Now, everything runs on it. Even your laptop is running on it. You don’t even realize it, that’s the same idea. That’s why I think ultimately in any kind of product you build, the more open you make it, the better. I want partners. I want people to connect. I want people to use it. I want people to contribute. Realtors love this because it is open. It’s accessible to everyone. It gives and it doesn’t take. [David] What do you mean by that? That sounds like an altruistic statement there. Explain it. [Pavan] It’s like with anything and like in your marriage, David, it’s in all of our marriages is if you give, and without expectation, you end up getting back more than what you end up getting and that’s not meant to like, like arbitrage. That philosophy but the idea is that look, AI is new. People are uncomfortable with it. People are scared of it. And if I give it away then I’m creating a new market. This is a completely blue ocean strategy. I’m in a place where no one’s ever been before. This is brand new for realtors, for customers, for loan originators. I have to give it away so that they could come into this new market and into this new future and every time someone comes into it, you got to try it, and you’re like, oh my god, you’re not going to want to be anywhere else. [David] I think you’re the one that uses this analogy on another one of my podcasts. It was like Elon Musk was interviewed. There was a guy interviewing him about the rockets and this guy says Mr. Musk, I have a question. Why don’t you do it? Whatever, configure it this way. And he said it’s because, and then he thought and he paused. You go, you know what? That’s a really good idea and it wasn’t too long after that. So it’s when you invite more people into using it. This thing is going to develop and also reminds me of how a lot of universities do this. They build a new building. They don’t put down the sidewalks yet. They just let’s see where the paths wherever one walks once the paths are down, then they’ll lay the concrete down because that’s why put a concrete were in a direction where no one’s going to walk. I think that what you’re doing is laying down some new paths for our industry put on the this slide that’s still up here. And folks, if you’re listening to this, you can get the slides. We’re going to put them, connect them with this podcast. Pavan will provide those, but it is really exciting. If you have one more slide that just popped up there, that was pretty good. [Pavan] Yeah. This is the web of compliance. [David] That is a hilarious slide. It does feel like a web [Pavan] And not even SpiderMan can move through this web but Angel AI has no problem. It does these millions of calculations instantaneously, finds the intersection of all these vectors, and says, here’s the answer. [Pavan] This is so exciting what you’re developing, what we have the privilege of hearing, and I’m so thrilled to be in a unique place working with you on this. We’re going to be doing a lot more podcasts. Thanks for bringing this article to my attention today. And thank you for taking time at the last second. We just got a phone and said, let’s record interview now. [Pavan] Can I add one more thing? one of the things that this article that this researcher puts forward is that, the bottom line is you cannot use LLMs to make loan mortgage decisions. That’s the bottom line and then the bottom line, Angel AI does not use LLMs to make mortgage decisions. It does use LLMs, but for very irrelevant, periphery subsystems, if the LLMs went away, 99% of Angel AI will still work the way it is. Okay, 99.9%. You cannot use LLMs to make mortgage decisions. It then begs the question for all, there’s so many companies that are rushing right now to take an LLM or take a set of LLMs, integrate them together and use it as their mortgage backbone. I’m going to give them a hint, it won’t work and our researchers have, while we develop this and as LLMs mature, our researchers have kicked the, you know what out of the LLMs to try to make it work, and they cannot make it work. And we have some really good, state of the art researchers. We’ve had, at least, two major academic publications done on our research, and we got another one coming out later this summer. We’ve got some of the best researchers in the country working on this team. You can use LLMs to some degree, but You cannot use LLMs and expect 100% trusted answers. That’s the big difference. Angel AI delivers 100% trusted answers. That means every single time it answers you something, it tells you how to do a loan, I will put my money where where Angel AI’s mouth is, I put my money. So that if you do what she says, I will fund that loan. That’s the warranty and I challenge anybody, and I’m going to save somebody who’s trying to build a mortgage AI, I’m going to save you a lot of time and money and headaches and because the first thing you got to be able to say, ask yourself this question. Will you act 100% on whatever the AI says? Will you stand behind it? Will you fund that loan? Would you put your money behind whatever the AI says? If you say yes, then proceed. You’re on the right track. If you say no, you’ve making a big mistake. [David] Pavan, this is just real exciting stuff developments. Thanks so much for bringing this article to my attention. So glad you’re willing to come on tonight and for us to record this interview. It’s got to get out there. LLMs have a place, but it’s not the place and they cannot be front and center on this and a lot of people are going down the, down that path. And I appreciate you be willing to step up and say, guys, that won’t work. Listen to me on this. It’s not because you’re trying to just make angel AI, the end all be all it is, but you’re trying to help guide the technology of the future. And I really applaud you for it. Thanks so much for being here, friend. Appreciate it. [Pavan] Thank you, David. Cheers. [David] Yeah. Tell your daughter, thank you for finding this article and bringing it to our attention. All right, friend. Have a great rest of your evening. It’s great to be here with you. Always appreciate it. Listeners check out this article and share it and it’s so important that this message get out and pay attention to what’s going on. Yes, we are featuring a lot of podcasts on AI. It’s because that’s where all the activity, that’s where the changes and it’s changing rapidly. If you haven’t read the book, The Coming Wave by Mustafa Suleyman, get it and read it. It’s exciting stuff. These are exciting days. Pavan, thanks for being here. [Pavan] Thank you, David.Important Links
- SunWest Mortgage website
- Angel AI website
- Article about AI exhibits racial bias in mortgage underwriting
- LinkedIn – Pavan Agarwal
Pavan Agarwal is a renowned leader in the mortgage lending industry and a pioneer in bringing artificial intelligence to the financial markets. Agarwal serves as the President and CEO of Sun West Mortgage Company and Celligence International, LLC.