The AI boom is starting to look a lot like the dot-com bubble of the 1990s — with thousands of startups rushing to market, only for most to collapse under weak business models. In this week’s episode of Lykken on Lending, David Lykken and Pavan Agarwal break down MIT’s recent report predicting that 95% of AI startups will fail, the disappointing reality of ChatGPT-5, and what this means for mortgage lenders and fintech innovators. They dive into the critical difference between probabilistic AI (guesswork) and deterministic AI (traceable, auditable, and regulator-ready) — and why the companies that master determinism will drive the next wave of disruption in capital markets and lending.
[David] Listeners. We’re in for another update on AI. There’s so much going on in the world of AI. We’re bringing you now weekly quick updates and we’ve got again our AI expert. Mr. Angel AI himself, Pavan Agarwal. Good to have you back, friend.
[Pavan] Good to be back and some people think I’m a ballerina company, but no, it’s not. It’s not a ballerina company.
[David] I love the logo because it talks about the elegance of this and of what this is, and it’s an art form. Technology, when we bring it forward, like Angel AI is, it’s an art form, but we’ve got some updates I want to get into today, so share with us the latest things that’s important that people be aware of, as things in the AI world is fast evolving what’s that?
[Pavan] Yes, yeah, so the big news that’s everyone the AI world is talking about. Last week, MIT released a study I think it was last Thursday, that said something like 95% of AI startups are failing or either they’re failing or they’re going to fail.
[David] They’re going to fail, the revenue model if they’re trying to figure out the revenue model.
[Pavan] Yes, yes
[David] What is behind that? Share a little bit more about that, Pavan.
[Pavan] It’s just, you know, very reminiscent of every big technological breakthrough, like in the 90s. we saw this. You and I are old enough to live in the 90s, I think a lot of people in this audience are too young for that but when the internet boom came out, you saw so many tech.com startups, right, and they’re selling. Everyone’s putting up a website and trying to sell something and it seemed like the money would never end and everything would like these things would go on forever. Like it’s like the money would never end and everything would like. These things would go on forever and it went on for years, whereas on the AI side, it’s only gone on for a short time period. I mean, I think it went on for like a decade in the dot-com period and what happened was you had all these websites trying to sell a product and anyone could create a website and it and it looked like, it looked cool, but it didn’t. But after you ordered something, it was nothing much behind there. So, the markets quickly figured out that having a website up for e-commerce meant nothing, right, and so today in the AI world is very similar, like there are people throwing up AI apps left and right, which all they’re doing is taking ChatGPT or Gemini, putting a new picture frame on it and saying I have this great new AI idea and give me lots of money and we’re going to change the world. And of course it’s not that simple.
[David] It doesn’t work like that.
[Pavan] Yeah, and then on top of that MIT report, we had a disappointing release of ChatGPT-5.
[David] Yes, we talked about that last week.
[Pavan] Right, and what happened was, you know, Sam Altman hyped up ChatGPT-5 like it’s going to be the end all, like it’s going to be, you know, artificial general intelligence or real close to it and all this kind of stuff. So, all this hype was built up and he had to do what he had to do, because that’s how you raise money, but then ultimately, you have to deliver too, right? and what he delivered was you know. So-so it’s actually, it’s a good release, right.
[David] I’m hearing a lot of everything. Allen Pollack, who’s on my regular podcast, says this was an amazing release, but it lost its personality, it lost its empathy, it lost some key ingredients that people had gotten used to in their first release. And so it’s how you add, without losing is what my takeaway was for that article.
[Pavan] That’s part of it. But the other thing the big disappointment in ChatGPT was, you know, Sam Altman over-promised and under-delivered. Right, because if you had…
[David] As if that’s a new problem in technology. We’ve seen that for decades.
[Pavan] Right. So, the thing is like he promised it to be the ChatGPT to be the end-all, and what it came out to be was an incrementally better release. Yes, they made some mistake, but that’s tech. They lost some features, but they’ll put that back in. That’s not a big deal, but it wasn’t the end-all like he had promised his investors that it would be. So you have these two things happening at the same time. One is MIT comes out with this report that says 95% of tech, or AI startups are going to fail.
[David] Yeah, are going to fail
[Pavan] And then ChatGPT-5 was far, or is far, far away from the promise of artificial general intelligence, so you had this compounding effect. You had ChatGPT -5, which is nowhere near artificial general intelligence, and you know, you and I have talked about this before it’s not going to happen. It’s not going to happen anytime soon and we’ve never really finished that last podcast we did two years ago about why AI will never be conscious. So, it’s probably a different conversation we can have on that, and then so you had that and then had MIT saying 95% of startups. It’s kind of the dot-com moment for AI just happened right, dot-com bust moment.
[David] Yeah, and I think also, but it was also underscored with what the president of Microsoft said we’re abandoning the AI, some of the AI stuff, because we haven’t figured out how to monetize, monetize it yet. And so where is AI going? If we’re seeing companies failure 95% failure rate, where are we going with AI?
[Pavan] So, what it’s going to be is the herd is going to thin out, just like we saw in dot-com, and the ones with real science, real technology, are the ones who are going to survive and dominate, and the others are going to either fall or get merged or consolidated in dominate, and the others are going to either fall or get merged or consolidated in right. So, and that’s that’s what happened in the in the dot-com cycle, internet cycle, it’s the beginning of that cleansing cycle has started right now, and Microsoft’s CEO Nadela said this like two months ago. So he was, you know, he was ahead of the MIT report and on that note, if any of you watched the All In podcast with Chamath and David Sachs they were actually on Friday. They had a podcast where they were talking about these things and it’s interesting, Chamath basically said in that podcast the exact nearly seems like he was reading from my press releases he said almost verbatim If you go back. If you look reading from my press releases he almost verbatim If you go back. If you look at all the press releases on Angel AI, at the bottom it always says that AI, this is an AI that’s deterministic, not merely probabilistic.
[David] Explain what you mean to that for our audience. Dive in that a little bit.
[Pavan] I’ll be back up a little bit, right, okay? So let me back up a little bit. So for those of you who don’t know who Chamath is, Chamath is one of the original founders of Facebook and David Sachs obviously he’s the AI…
[David] And if you can send us a link to that podcast so those that want to click through it are going for extra credits on our podcast. They can click on and go watch it. So send us a link to it so we have that.
[Pavan] Absolutely yes, so it’s a quick, like it’s a clip. I sent you already the link. It’s a quick 10-minute clip of Chamath and David Sacks talking about this thing and you know I obviously all know David Sachs is in the Trump administration. He’s the AI czar in the Trump administration. Yes, he is, and you know very capable, made billions in technology. So that’s setting the stage of who’s talking about this stuff. So Chamath is saying now that you need AI. That’s deterministic, not just probabilistic, because LLMs and they’re great, they’re amazing, okay, and they’re good for creative work no-transcript. You’re going to create something out of thin air. You’ve got to guess that stuff, but when you’re actually trying to do work that you can count on depend on.
[David] Yeah, like underwrite a loan.
[Pavan] Yes. Like wire money, Like if the AI says, go ahead and wire a million dollars to closing. You don’t want it to guess at that. It has to be deterministic and in banking especially in banking and also in medicine is you don’t want to be guessing at that number one and number two, you need to be able to see why it came to that decision, right? So, and somebody people watching this who have used ChatGPT and Gemini can say well, I could ask them. I could ask ChatGPT and Gemini say tell me how you got there. But what you don’t realize, all it’s doing is just taking your prompt and feeding it back into its algorithm, and it’s guessing those words. Also, it’s guessing the explanation. It isn’t actually a log of how it got yeah it doesn’t actually keep track of the steps it went through to get to that conclusion and then spit that back to you. No, it actually puts that prompt back in and then it makes a guess as to how it got there.
[David] You’re really raising an important part because when we go into underwriting or any aspect of what we have to do, we have to have a clear, clear logic trail of how we got there so that, from a regulatory standpoint, any other aspect of auditing quality control we got to have a clear trail of how we got there and that’s what you’re talking about deterministic.
[Pavan] Correct, correct. You have to, you have to, you have to. Ai has to say we’re going to do this and this is why and it’s got to be right, when you’re wiring, you know, hundreds of millions of dollars a day, or billions of dollars a day out for funding loans or any other bank transaction, you can’t even make a $1 mistake, right, you’re in big trouble. And then also, on top of that, as you know, the OCC put out guidance, actually three years now ago, that said it’s okay for a bank to use AI, and by bank it also means mortgage companies. And by bank it also means mortgage companies. It’s okay for a bank to use AI to make decisions and run their processes, so that, as long as you can also give us the reasoning as to why it made those decisions, if you can’t tell us why the AI is not acceptable.
[David] Okay, is it also asking how you made the decision, why and how?
[Pavan] What is the thought process to get to that? And the technical term in the AI world is called chain of thought, right? but the problem is that when you understand how neural nets work, is there’s like you can’t explain their chain of thought. I mean because no one knows how they get the weights and how the network gets trained. When it gets trained, it just does these calculations and it gets created, but no one can explain why it got there. So when you ask it for chain of thought, the best it can do is guess. This is what I probably was thinking, not, this is what I thought. This is very different. If you want to understand probabilistic and deterministic is you can say If you want to understand probabilistic and deterministic, as you can say that if I’m going to drive from here to Las Vegas, I’m going to guess you took the 10 and then you took the 15. And you exited Las Vegas Boulevard right. But I don’t actually so that’s probabilistic Okay, but I don’t actually know.
[David] You don’t actually know that you took that route.
[Pavan] You could have taken one of the side roads through 29 Palms or something like that, right, or you could take it Route 66. So deterministic means, as you drove from LA to Vegas, every turn you made and every road you drove on, it was logged and tracked, and so that when I ask you, David, how did you get there, you print out this log and say this is how I got there.
[David] That’s a really good analogy for what we’re talking about here, right.
[Pavan] Yeah, so that’s how you know. So, to do real value right, and if anybody’s going to, bet their money on it. You’ve got to know for sure. You’re not going to depend on a machine that really doesn’t know why it did what it did. It just says do it when your money’s at stake. You’ve got to know for sure. And that’s the problem. And that’s why you know now that the hype is sort of dying down and people are realizing hey, wait a minute, this isn’t the hype is not real. Like that podcast. David Sacks has the same thing like what’s happened here is that people have figured out that AI is like any other tech cycle. Right, this sort of hype for artificial intelligence is not, it’s just hype. It’s not happening anytime soon, although if Yan LeCunn has his way, he’s going to get there.
[David] Yan LeCunn. Can I explain who Yan LeCunn is?
[Pavan] He is the father of LLMs. He’s like the world’s leading researcher in Artificial Intelligence. Him and Dr Ben Gorzo, who I know well.
[David] Yeah, Ben Dr Gorzo,
[Pavan] He’s one of the top minds at AI, and they’re working on entirely new frameworks, and Dr. Ben Gorzo has been working on completely different frameworks than LLMs for a long time, and so is Yann. Yann is now working for Meta, and so they’re off to new principles of Artificial Intelligence. So I think, when they finish their breakthroughs, I think that we’re going to have another big boost and another big wave of AI.
[David] When do you predict that happening?
[Pavan] Oh shoot, I mean, it’s like anything. You know when is high-temperature superconductors going to be here, right? You just don’t know. I mean you know. Theoretically, you know Well. I mean, I don’t even think it’s proven. There was a Nobel Prize won by a guy who proved that you can’t have high-temperature superconductors. Anyways, I have to go research that I digress, Sorry. It’s like with any other science right.
[David] What I was trying to get to was the rapid rate at which things are developing. So we’re locked on to ChatGPT right now, but there’s a whole new. The guys that created all the LLMs that are there now are already saying that’s done, we’re over with that. We’re over that. We’re now working on the next, not generation, but the next whole chapter or era of what’s happening. Yes, pretty interesting.
[Pavan] Right, I think what has the advancements that have been made with AI the last five years? I think there’s a lot of work to be done to be commercializing them and implementing them correctly. Okay, and I also think there’s another wave, big wave about to come, as concept-based AI that understands entire concepts are rolled out and implemented. Okay, and tomorrow we’re going to be releasing a press release where the latest build of Angel AI has what was in our lab for years is actually out in production. So if you use Angel AI, it’s there, and the latest build of Angel AI understands concepts and you can tell when you look at the difference between how it answers a question versus how, say, Gemini answers a question.
[David] And the results are deterministic.
[Pavan] The result is deterministic right because it understands. It understands bigger pictures.
[David] Science is getting more and more exciting. Pavan, where this is all heading.
[Pavan] And look, we’ve got some really, really smart engineers working on this stuff. And we also read all of the other research that’s done and we learn from other people’s research and then we add our own ideas to it. So, the reason it seems like we’re always ahead because we’re always reading the most current research that doesn’t make it into mainstream pop culture for years. So, we’ve already read and analyzed it and have made conclusions as to well, yeah, we’re going to pursue that direction or not pursue that direction, right. So it boils down to reading and staying ahead of the research. And second is having good instinct, right? If you have good instinct and if you understand the tech, you’ll know which research direction is worth pursuing and which one is not. And those are the things that we’re doing.
[David] Yeah, we understand how deterministic I think you did a great job of explaining that versus the what was the other one problem? Probabilistic and so now we understand that, now we’re talking about your latest version of AI is going to have the ability to perceive and answer things that are on a perception basis. Where does that come into app? Where’s the practical application of that?
[Pavan] Well, I mean the example, simple example is you ask it a scenario question and it will cut through all, all of the noise and give you a very simple, precise answer and you go in there, you try right now, right, and, and as an example we had a broker who asked it a question about he has an FHA, he wants to have to cash out with the HELOC, the guy has a second lean HELOC and he had a late recently on HELOC, less than 12 months. Right, and if you on one read of the guideline, you can say you can do it, because what the FHA guideline says is you can do it, provided that you downgraded to manual and that it has extenuating circumstances. Okay, so when he asked Gemini, Gemini said you could do it. But you know, see, talk to your lender, because Gemini read the first, didn’t understand the whole concept. The whole concept is there’s all these layers and they all have to interconnect together and then there’s your answer. So when you ask the same question to Angel, she actually understands the whole concept and says well, no, you really can’t do it. And the edge case of when you can do it is so narrow. And in that particular situation, Angel had all the information about the file and it just flat out said no because it didn’t meet the external circumstances and all the other stuff. Because, if it, got downgraded to manual, it wouldn’t have qualified. As you know, because FHA manual underwrites you have to have. I’m not a guideline expert. But you have to meet certain criteria. I’m not going to try to quote the criteria and get it wrong, but that particular loan didn’t meet the manual underwrite guidelines, right? So the AI understood the loans, the entire loan, the concept of the entire loan, understood the concept of the guideline. So if you’re going to make a Venn diagram, right, it says here’s the concept of the loan, here’s the concept of the guideline and there’s no intersection here. This loan is not eligible for manual underwrite and he had a HELOC late and therefore it has to be manual. And guess what? The two don’t connect right.
[David] Very interesting.
[Pavan] That’s a level of thinking that human beings do right, yeah, right.
[David] We’re getting closer with the latest version with the perception ability.
[Pavan] Exactly, very exciting.
[David] Exactly, we got to wrap this one up because I want to say something for next week. It seems like there’s so much developing, but I want to just say kudos to you, Pavan and your team. I was just on a call. I’m not going to name the major lender. It was like one of the top ten retail lenders in the nation. I got invited into the call. Lord knows how I got invited, probably because I work with both companies you and as well as this other lender and they invited me into this call and they were talking, their eyes opened up. You said you mean, we could do this. We do this because it’s perception, the perception of what you’re talking about right now. These new tools folks are going to take over and run our business. If you’re not using them, you’re going to be left out and you’re going to. Well, how do they do that? Because they use technology that you’re not. Sign up for Angel AI. You can go to angelaicom, learn more about it, go to Ask Angel and actually start using it for free.
[Pavan] Yeah, and if I could. If I have a minute, I could add something about that call you just had so the reason that call happened. I mean, in a way we’re competitors right. but I don’t believe like anybody’s a competitor in the sense that I don’t take that attitude. We never did right. It’s like it’s a small industry man.
[David] It’s collaborative.
[Pavan] The way I was raised and the way you were raised, and the industry, if you remember from the 80s and 90s everyone, we all worked together. We all helped each other. And I really long for the days where the industry starts thinking that way again back to that. Right. So that’s number one. Number two is we have brought AI to the capital markets, and that means we have tapped into hundreds of billions of dollars of capital that’s been sitting on the sidelines, and we’re bringing it into America and being deployed into mortgages. So therefore, we’re creating products that didn’t exist before, right, and so we could get creative and create new products at very low cost.
[David] And it’s being underwritten. The nice part about that is you’re not taking a risk because it’s unproven. You’re using Angel AI intelligence to start creating probabilities and the success of this product intelligence to start creating around probabilities and the success of this product. That’s why this one investor overseas investor, foreign investor is coming in with billions of dollars and doing it through Angel AI. It’s really exciting. People got to stay and pay attention to what you’re doing, Pavan. That’s what I told. That’s why I’m glad I got invited to that call, because I said guys, you do not realize what is happening as a result of this technology. Kudos to you, Pavan, for what you’re creating. I get so excited about it. I’m so glad we met 44 years ago. I’m so glad to have the opportunity to work with you today, so excited about it.
[Pavan] Thank you, David.
[David] It’s a joy.
[Pavan] Cheers.
Important Links
- Chamat Palihapitiya discussing about AI
- Press release about Japanese investor bringing AI powered innovation
- SunWest Mortgage website
- Angel AI website
- 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.