You know, running a small business, your internet isn't just for email. It's how you process payments, manage your team, and stay connected to customers when it matters most. Whether you've got one location or five, reliable connectivity helps keep your transactions moving, your team connected, and all the tools you rely on within reach. Because when your connection just works, you're not busy troubleshooting. You're focused on your customers and growing your business. AT&T Business can help you stay connected and ready for what's next. Visit att.com/smallbusiness to learn more.
Right, people are very willing to pay for things if it solves their problem. People are not willing to pay for things that they find annoying, that they don't find valuable, or they find ineffective. And so, it's kind of like advertising. People don't hate ads. People hate ads that don't matter to them. But if you've ever seen like an amazing ad, look at the Super Bowl ads. People literally tune in to watch the ad. People are willing to pay for products as long as it is helpful to them. They expect to pay for products that are helpful to them.
Running a business means solving problems. I tell you how the smartest entrepreneurs do it. Hi, I'm Jason Feifer, editor in chief of Entrepreneur magazine, and this is Problem Solvers.
I'm obsessed with this idea in entrepreneurship, which is this. You know, we're all trying to solve problems, and we work on them, and we work on them, and we work on them. And sometimes we step back and we're like, "Ugh, it's still a problem." And somebody once said to me, "No, no, no. The way to think about it is, do you create a better problem?" Like, there's always going to be problems. The whole journey is problems. So, the way to track progress is not, "Did I solve this problem?" but rather, "Did I create a better problem?" Today's episode is really about a better problem, which is this. In many businesses, in many ways, whether it's technology or breadsticks, we're going to come back to that, you can have the problem of a customer is so engaged, they like the thing that you have so much that they are costing you too much money. What are you supposed to do about that? Isn't that supposed to be a good thing? The customer is there. They really like your thing. But it's costing you money serving them. What are we supposed to do about that? Today we're going to talk about AI, but we're also going to talk about breadsticks because my guest today, I don't think he serves breadsticks, but he'll tell you about it. Maybe he does, he'll tell you. His name is Anthony Pompliano. He is a prolific entrepreneur, investor. He's had his hands in all sorts of things. Extremely impressive. And author of a book called How to Live an Extraordinary Life. And today we're talking about his company Silvia, which ran into this exact problem. We'll talk about it in a second. Anthony, welcome to Problem Solvers.
Anthony, okay, let's get to Silvia and that problem in just a second, but quick, let's just back up here. Tell me about you. Tell me about this great book.
Yeah, entrepreneur and investor for probably last 20 years or so. And as I've gone through my life, you know, I've always tried to seek out great ideas. And usually it's learning from somebody else or from making a mistake myself. And so in my phone I've got a whole Apple note. And basically anytime I come across an amazing idea, I read it, I hear it, I learn the lesson myself, I just write it down in those Apple notes. And about three or four years ago I decided that I was going to sit down and write some of these into a book. And so the first volume of How to Live an Extraordinary Life was 65 Letters. And I was trying to think of how do you, you know, kind of share these lessons? And I thought, well, the people that I care the most about that I want to learn this are my kids. And so why don't I just write it as letters to them? Each chapter is a single letter and it has Here's the idea. Here's how it impacted me. And here's how I think you can apply it to your life. And the beauty is that those letters they apply to a 75-year-old, you know, person who's retired and and just wants to read something on the beach all the way down to, you know, a teenager who says, "Hey, I I to learn as well." And the volume two that is coming out this week, that was 50 more ideas. And what I learned, I think, is you know, writing a book is a skill. And it's a skill that when I wrote the first one, I don't think I really had. I kind of just wrote, and then we just put it into a book. With the second one, I I'm almost ashamed of the first one, you know, in terms of of the structure, et cetera. But with the second one, you have had some people read it and you know, whether it's the Robert Kiyosakis, you know, kind of famous authors, all the way to some of, you know, my friends who are financiers, et cetera. They're all like, "Hey, man, this one's way better." So, so I'm excited for people to read it.
That's awesome. You know what? As a guy who's written a book, I can tell you, number one, you haven't grown as a person unless you're ashamed of your first book because you it just means that you grew so much more than it, but also, I did the same thing. I wrote the first four chapters of my book, and I'm a writer by trade. I wrote the first four chapters of my book, and I sent them to my editor, and his feedback was basically, "So, you wrote four magazine articles, and they don't feel like they have anything to do with each other, and they it's not a book. You're not writing a book. You're writing a bunch of magazine articles." So, I had to learn how to write a book, and and and you did, too. And that's and that's growth. But look, you just to be clear, when you're talking about great ideas that you come across that you that you identify, people should understand we're talking about the kind of person who was an early investor in companies you might have heard of, SpaceX, Airbnb, Ramp. You've raised more than $750 million for your businesses, haven't you? So, like, you know, you've got your eye on really sharp ideas and what progress really looks like.
We are we're trying every day to figure out how to solve problems. So, this is the perfect podcast for us to come talk about it.
Appreciate it. All right. So, you got this company called Silvia. It's an AI financial assistant, works as a personal CFO. Take me into this challenge that I had teed up at the top here, which is that you built the thing that people loved and then your power users were actually costing you too much money.
Yeah, well, if you're thinking AI every time that you ask a query, basically you're using compute, right? Or or consuming tokens. And so that costs money, whether it's your own model or you're using one of the frontier models. And so when we first launched the product, it was free and we felt like that was really important is to have people come in, try it, understand how it works. Now, with Silvia, what the you know, kind of the pitch, the value proposition is basically when you come in, you attach all of your financial accounts. So your bank account, your brokerage account, your crypto accounts, you can attach your credit card, all your alternative assets, so private investments, real estate, cars, collectibles, etc. And then you begin talking to Silvia, but she has the context of your personal finances. And the best way to understand the value here is if I go to Google and I just ask how do I get my tax rate down, Google will return me a bunch of very general links to websites about taxes, but it has no clue about my personal situation. Same thing with all of the, you know, most popular LLMs. With Silvia, when you ask that same question, she will go into your portfolio and look asset by asset "Hey, this stock, you're underwater, you could tax tax loss harvest this one. Hey, this piece of real estate is in an opportunity zone or you could bonus depreciated." Now, your point about the problem is your power users, they're asking a lot of questions. They ask the most amount of questions. Well, every time they ask a question, it means it's costing you more money. And so you want people asking as many questions as possible because that ultimately means that they are engaged. They are your power users, right? They're the most valuable ones. But, at the same time, every time they ask that query, your expenses are going up. And so how do you settle this? What I remember is a couple years ago, maybe a decade ago, there was this famous PowerPoint that went viral online. And it was a private equity analyst had gone to a bunch of Olive Gardens. And he basically put together this entire PowerPoint and he basically, you know, said, "Hey, they're serving too many sticks of bread and too much salad."
Because because because Olive Garden was known for the unlimited breadsticks and salad. It's
Dude, I I grew up in a family. I I grew up in a family with five boys, we're Italian, and it sounds insane, but it was like a treat. Let's go to Olive Garden, right? Like, you know, when you have five kids, you don't exactly go out to eat all the time. And so, one of the places that you can go that seems like it was really upscale, but probably was less expensive than if we went to a really nice place was an Olive Garden. And so, part of the beauty of going to Olive Garden is your parents knew that they're going to bring tons of breadsticks and tons of salad, which means that you're probably not going to order that much food, right? Or you're going to order less food. Well, the opposite side of that is Olive Garden is feeding you tons of breadsticks and salad that they're losing money on, and you're not ordering as much food. And so, ultimately, this PowerPoint basically showed, "Hey, if when you greet the person after they sit down at the table, do not bring the breadsticks and the salad yet. Instead, take their drink order. Go away. Bring the drinks back. Then, you can ask them for their food order, and then only after they've put in their food order, then bring out the breadsticks and salad. So, they're still getting it before the meal. They still feel like there's a little bit of this brand promise of free breadsticks, you know, and salad, but you have taken maybe the first 15 minutes of them sitting at that table, and they don't have breadsticks and salad in front of them, which means that they're less full, so they're going to order more food. And then, it also means that you're not going to run through two or three turns of that food. And so, naturally, this just changes the unit economics of those people who come into your restaurant."
Right. So, okay. So, this is So, So, I appreciate that you have done exactly as I promised. We have connected AI and breadsticks. But now, loop me back in here because you've got this problem where you've created this really useful tool in Silvia, and just the way you explained it, and the way in which it's going to understand at a nuanced level somebody's finances, of course, if somebody's very interested in their finances, they're going to be asking all sorts of questions. And so, how do you now do the breadstick thing, which is to not pull back or well, actually, let me say let me say let me not even ask that. Let me say, what was the first thing that you thought to do because some people might say we need to just limit maybe the number of questions that people can ask or somehow squelch down people's usage of this tool.
So, we basically said, what are all of the options available to us to address this, right? One, should we just let people keep using it as much as they possibly can, spend whatever it takes, and just raise more money or you know, just pay for it and it's worth it. The second thing was, are there things that we could do where they're still asking the same number of questions, but maybe we don't need to consume as many tokens. So, in essence, get the cost down on a per query basis. The third thing was, maybe rather than send all of the queries to the most powerful model, could some of the lower cost models actually answer the question just as well, and so we could kind of address the cost that's the main point. And then fourth was, should the user actually bear some of the burden or should the user actually pay for some of this as well. And long story short is, we basically got to the point where a little bit of everything ended up being the solution. So, today when you come in and you ask a query, one, it is completely free to use unless you're a power user. But two is that we actually will take your query and we will route it to the model that is best suited to answer the question. So, one of the aspects that people don't really realize is if you ask the most high-powered LLM, what's today's date? It usually takes a little bit longer than if you just ask your friend. But that's something that your friend knows almost immediately. What your friend can't do is, you know, a Monte Carlo simulation or some crazy calculation that AI's really good at doing. And so, the AI goes through a lot of the same kind of infrastructural sequence to answer what is the date as they would do to do some crazy calculation. And so, it's just slow. So, if you can answer it accurately and faster and cheaper, it's a win-win-win. The user gets a better experience, this latency is lower, and the cost is lower. And so, that model router is really impactful because it basically routes queries to where they need to go.
And and that didn't I'm going to ask like a dumb technical question here, but so I assume that that model router, as you're calling it, didn't exist in the earlier version of Silvia. And so, what you had to build here was a layer that essentially assessed the complication of the question and then made a decision about which model to use. Is that right?
Correct. And if you look today, there's, you know, companies like Open Router and many others, they basically provide the ability to do model routing, but the intelligence of which one do you route to is very specific to your infrastructure or your use case. And so, we built this from scratch. You know, our engineering team is very proud of the fact that there's kind of two core components. They use the first 500 characters of a query to predict whether it will be high effort or low effort. And they do that with incredible accuracy so that they're able to route it to the right model. But the second thing is that they are able to make this decision in sub 1 millisecond. Meaning that you don't increase the latency for the actual model router to be inserted into this. And so, that alone was able to divert quite a bit of cost because now all of a sudden we were able to take a double-digit percentage of queries and say, "Hey, we don't need the most powerful, most expensive model to answer this. We actually can give the user a better experience and it be lower cost for us by routing that."
Yeah. And And before we get deeper into your solutions, can we just spend one more moment talking about how you made the decision to pursue the solutions in the first place? Cuz you just laid out a really useful breakdown of all the different ways that you could attack this problem. But what were the internal conversations that you guys were having? Cuz I assume you you know, there has to be a cost-benefit here. It's like, "If we do this, then what is going to happen to the user experience? If we do that, like are we going to impact something that we've a long time now building habit or trust on. So, what were those conversations like?
There's a concept from the military called bracketing, which I think a lot of people should know about but but don't use quite often enough. And the idea of bracketing is basically, let's say you're shooting a mortar. For for most people, just think of it like you're shooting a rocket, you know, towards a target. It's very hard to aim these or kind of dial it in. It's not like the movies where, you know, you just kind of just lock onto the target and the and the missile just follows wherever it goes. And so, when you're doing this, let's say that, you know, you realize that there's a target you want to hit. So, you shoot the first mortar and you shoot it 100 yd. And it falls short of the target. You know in your head that target is at least 101 yd, if not further. And so, you could say, "Well, I think I was pretty close. Let me try to do 150 yd and see where you land." Well, the better thing to do is say, "Okay, I know I'm at least 100. My next shot, I'm actually going to try to shoot past the target. So, I'm going to shoot 1,000 yd. And when I shoot 1,000 yd, I go over the target. Great. Now I know that my target is somewhere between 100 yd and 1,000 yd away."
You're calibrating where the target is.
Correct. So, now what I'm going to do is I might shoot 600 yd. Am I short? Am I long? Did I hit it? Oh, I'm short. Okay, well, now I know that I'm somewhere between 600 yd and 1,000 yd. Let's shoot 800. And you just kind of walk yourself into where is that target. It's a much better way to do this than go from 100 yd to 110, 150, 200, 250, 300. And all of a sudden you realize, "Oh my gosh, I missed
