Mike Gonzalez de Benchmark AI: inversión con IA a nivel de cartera, menor frecuencia de operaciones y comparación de modelos de frontera

[3:49:00] ?: Do you guys have a good lunch?

[3:49:03] Anthony Pompliano: So I don't know if you guys know this. I knew this but forgot it earlier when we were talking.

[3:49:08] ?: Remember when I was saying all this stuff? This is being streamed live.

[3:49:12] Anthony Pompliano: So all the people on Twitter when I was like, I'll just tell you guys. I went to lunch, I just saw all these texts from people. I'm like, Jesus Christ. All right, I'm very excited about this. My friend Mike Gonzalez is gonna come up here. Mike has, we worked together at Facebook. Mike was one of the finance leaders that helped turn Zenefits around once they got in a little bit of trouble. He then built and sold a company to Paylocity and now has been working on a very interesting AI-driven investment strategies, which I know I've gotten a ton of questions about, so I feel like we got. Somebody here who knows a lot more about it than I do. And I think this is the first time he's publicly talking about any of this stuff.

[3:49:51] ?: So it should be pretty fun. But Mike, why don't you come on up? But there's supposed to be a walkout song?

[3:50:04] Anthony Pompliano: There is no walkout song. All right. So the number one question that I get when people start talking about Sylvia stuff is they're like, OK, I'm in the mindset of AI finance. When will AI be available to actually trade for me? So I want to give the AI money. Sit back, chill, Netflix and chill, and it just makes me money. You've been working on this now for quite a while in terms of figuring out where are we technology-wise and that actually becoming a reality.

[3:50:35] Mike Gonzalez: Thank you for having me, first of all. Happy to be here. Thank you all. The use of AI for investing is at a very unique and interesting time. I think all of us are probably on X, and we go to X for a lot of information. I think Anthony is definitely in that camp and he's producing some of the information. But I think a lot of what you see on X is often it's a lot of marketing. And I think you see a lot of these trading bots and these things that look like AI and they're doing this stuff and the account goes up. But I think there's a lot of questions around, you know, what is actually behind some of those systems and some of those things. And I think the approach that people take. Is also very important. What are you looking to do with AI managing your money or investing? I think a lot of what we've seen so far, it looks like people trying to apply quant strategies, trading strategies, and using AI to make those decisions. That's super interesting. I think quant is kind of where machines and the application of technology really influence financial markets in terms of these, you know. Technical strategies that are engineered that deliver some type of performance, some type of alpha. And the way that AI works, and I think you probably heard this. A lot of times, but AI is probabilistic. It's not deterministic. And code is traditionally deterministic. So when you define a quant strategy, it just runs the rules. And it runs them ruthlessly. And it applies the same decision making every single time, even if the data changes, applies the same rules. When you build a system that uses what we call AI today, You're introducing something that's probabilistic. You give it a prompt and you get an answer and you don't know what that answer is going to do. And so quant trading, it's a unique approach. We look at it very differently. And so when we look at those approaches, we don't try to compare what we're doing to a millennium or a citadel. These guys who are quants, they're doing hardcore. We're thinking about it more from a portfolio perspective. And we're thinking about asking questions. If you try to replicate an investment team, a portfolio manager, people who go out, they look at the market, they come up with a strategy, they go come up with watch lists. They go research those equities and then they build a portfolio, oftentimes over a longer horizon. And I think we've been doing that for a very long time. And I think we have sufficient data to say that that works today. And we have a research record. On the quant stuff, I don't know. I see this stuff on X. I see this stuff on Twitter. There are going to be people that use AI. To deliver quant-based strategies. And I think you hear Jane Street's investing a ton in tokens and AI. But I think for most folks here, you're thinking about your own portfolio, right? And I think that's what a lot of people are doing. I think for managing a portfolio, I think we are there today. And there's just a ton of questions that need to be answered to figure out how do we apply it? How do we apply it for any given... Portfolio, any given strategy. We've asked ourselves a lot of those questions. We've answered a lot of those questions. And I think for Benchmark AI, we're now at a point where we're ready to start formalizing those into financial products. So I think we're, the answer is I think we're there.

[3:54:18] Anthony Pompliano: Talk a little bit about one insight that you gave me that I thought was very interesting is the more that the AI tried to trade, like it was doing it every day, it actually seemed to have degradated or worse performance than if it like rebalanced once a month or something, right? Like, do you have any insight into maybe the people who are chasing the like, hey, I want the algorithm to day trade for me? That actually may be just the wrong approach and you get outperformance by like making less decisions?

[3:54:43] Mike Gonzalez: That is true in our data. And I think we've run a series of experiments. And I think that was one of the kind of key questions that we asked, which is the frequency. And frequency introduces a lot of dynamics because you are increasing the cost of the system. And even though we have portfolios that rebalance monthly, the system is still running every single day to do research. It's really about when do you allow. The portfolio manager to make a decision. And weekly trading, daily trading, weekly is worse than monthly, daily is worse than weekly. And so it's almost like humans. It's like you start trading more frequently, typically you do less. And I think the more patient you are, then you typically see that the performance is higher. And it's not just the activity. It's also things like taxes and turnover. Once you start applying a lot more transactions, you're trading a lot. I think this is often sometimes what people miss, even in their individual brokerage accounts, is just how much some of that churn and the tax impact and the transactions actually adds up and eats away versus more of a kind of a buy and hold strategy. And I mean, Ken Griffin is, you know, probably one of the folks who's been, you know, you know, talking about this a little bit more. I mean, I think he was public saying, you know, many months ago that, you know, AI is at the point where it's doing the same type of work that his PhD researchers are doing. And he's also said that he thinks the alpha in the market is going to be more over long term horizons. You know, it's going to be more over the long term. This is from the person who. You know, is the market maker. You know, he's the one that's made a lot of the money on kind of those quant market making strategies. So the longer term nature of where alpha is, it's reflected in like the data that we see, but it's also you also kind of hearing it from some of the other kind of prominent figures on AI.

[3:56:45] Anthony Pompliano: What I see is a lot of the models, they're constantly competing over who's more performative, who is lower cost, and they've all these different ways to try to compare themselves. What are you seeing? Because for context, with Benchmark AI, they have run the same type of approach through Grok, Gemini, Claude, ChatGPT, et cetera. Do you see one model performs better than the others, or is it kind of different strategies for different models?

[3:57:10] ?: It's a great question. It's a very interesting question.

[3:57:13] Mike Gonzalez: I think... I'll take a step back and talk a little bit about how we got here. Why am I even building this system? I think Anthony mentioned on the intro, my background has been in finance and product building. I met Anthony at Facebook where I built our financial systems, our budgeting systems, our long-range planning systems, so the tools that produced Meta's earnings and revenue forecasts. As an entrepreneur, I built... Corporate finance, startups. After I sold my company, I had every intention to go build AI for CFOs. It's the space that I knew. I knew the players. I knew the problem. And when we went out to go start thinking about, okay, let's go build this. Let's build this new system for CFOs. One of the first questions you ask is, okay, what is our stack? What's our architecture? What are we going to build on? And this is about two years ago. And at the time, it wasn't really clear who the best frontier model was. And so our idea is like, well, what people are doing today, you benchmark it, you build a benchmark around everything. And so we built a benchmark. And that was the question that we wanted to ask ourselves, which is, if you apply any of these frontier models to a series of finance tasks, who's going to deliver the best results? And so when we started to package all this up, we were like, OK, we're going to build this one so we could figure out what frontier model we're going to build on. But we're going to share this publicly, so let's make this story a little bit more interesting. Let's have them build portfolios and have these finance tasks get put up against the ultimate benchmark, which is the S&P 500. So that's kind of where this started. It was more so a side project that we got working on. And so when we built the system, we said, okay, what is a frontier? ChatGPT, Claude, Grok, Gemini. At the time, those were the four. And so we built... You know, each, all of these different finance tasks, and we built the system for each one of these frontier models. And it's not an easy thing to do. I think it's very hard to get one model to do your work. We wanted to do it across all of them. And, you know, when you run these types of experiments, you need to see a lot of different data. And so we did, we went about a lot of experiments around different strategies. The first actual question that we asked ourselves was, if we give AI risk parameters, will they follow? And so we gave them each three strategies. One was a risk adjusted strategy, which is for every unit of risk, you know, measure the unit of return and optimize for risk adjusted performance. Another one was more around that long term nature of these things, which is like, all right, we know that a lot of the performance in markets is driven by short, the short term nature of it. So let's think about a little bit more long term. And then the other one was just max alpha. Like, let's just see, let's just let these things rip. And like, let's just see how much performance they can drive. And what you saw was the strategy correlated to performance. You know, the risk adjusted portfolios, three out of four of them are beating the S&P 500 in this research set. And all four of them are less volatile than the S&P 500. Max alpha, huge returns, but also a lot more volatility. And so they are listening. The question about which model is better, there's a lot of different ways to answer that because it's not just about return, right? A lot of times you're looking at risk-adjusted returns. And so you can define a model, you know, what is the best performance by a range of things. And I think we've also run a bunch of different strategies, you know, after, you know, we understood can we solve risk, we looked at what about, you know, applying custom weights versus equal weights, you know, like if you just ask them to just pick stocks. What about large cap versus mid and small cap? A lot of people do a ton of research on large cap. Mid and small cap, less research out there. So, you know, we kind of had a thesis that there might be more alpha in the mid and small cap. And so if you look at across all those portfolios, I think you'd see that the leaderboard kind of shifts, but there kind of is one that kind of sticks across the top, especially in absolute returns, and that's Grok. So that's been an interesting learning for us. I wouldn't say Grok is the best model because sometimes you look at Sharp, for example, and the Sharp ratio measures risk adjusted return. So it's not just looking at, you know, it's not just looking at absolute return. And you'll see Gemini or ChatGPT kind of like bounce up there on for the Sharp. And that matters, right? Because it really depends on what is your definition of best. And I think, you know, everybody in this room probably has a different definition of what best may be.

[4:01:54] Anthony Pompliano: You think about the implications of this on the market. Somebody asked me a question earlier of if everyone has an AI trading in the market, does that just arb away all of the alpha? Does that just completely, you know, if indexing became just like the market is the index, does AI take that to the extreme?

[4:02:14] ?: It's a great question.

[4:02:15] Mike Gonzalez: It's a very interesting question. I think my answer is I don't know, of course. But when I do really think about it. My actual answer is no and yes. The reason it's no is because when you look at the different models and you look at their stock selection, even if you give them the same exact inputs, so everything's the exact same, you'll see that the four different frontier models, they pick very different stocks. So even that same strategy set, two-thirds, if you look at all the tickers that were selected across the four models in these strategies, Two-thirds of the tickers will be unique to one model, you know, so the actual overlap is pretty low. And so when you think about is AI the market, that presumes that everybody might be using the same exact AI, or everybody's using all the AIs, and we know that that's not the case. I've been spending a lot of time with some of the big multi-strat funds, some of the people that work there, some of the people that work in their AI groups, right? So think big multi-strat fund, think somebody who's like on the AI team. And it's interesting because I'm talking about some of the things that we're learning and they're a claw chop. And that was super interesting to me to think about, you know, biggest of the biggest fund and they're not experimenting with some of these other models, you know, and Gemini is, you know, very, it's built on.

[4:03:38] ?: Google's database and search. You know, Grok is built on X. So they're very unique models.

[4:03:42] Mike Gonzalez: Even if they're all at the frontier and you see some benchmarks where they're all kind of like in the top right, they are very unique, you know? And so if you're an enterprise multi-strat fund or a big kind of portfolio manager, does your CTO or your chief security officer, do they let you just go, you know, build a bunch of stuff? So there's things like that. And there's also just my reflection on the system and the system that we've built. The number of decisions to get that system, there's just so many decisions. And then even thinking about, there's only a couple other public companies that are doing what we're doing, and it's more on the quant side. I think I kind of see that. It's kind of like what you see on X, where it's trading bots, it's quant, not so much on more of kind of portfolio construction and thinking a little bit more long term. So that's the reason why the answer is no. The reason why I think the answer is yes is because I think there will be companies like Benchmark AI. And what we really do is we just research the application of AI on investing. And we're going to package up that research and stuff like this. It's going to get out there. You know, I think all the research that we're doing, it's going to get out there. And the way that we think about the business. We have a system and that system produces, you know, financial products. And on one side, you know, we are looking at private funds, you know, stuff that's, you know, a little bit more, you know, private bespoke has, you know, specific agreements. On the other side, we want to create an index and we want to create indices. And when you think about the S&P 500, there's a committee and they pick the stocks. When you think about Nasdaq, it's more rules-based, but they do have some discretion, like they just put SpaceX in, there's things like that. So if you got the traditional S&P 500 doing research, Benchmark AI is doing the research, and we're going to use our frontier AI systems to produce indices. And I think that's where AI might... You know, kind of become the market if there's strategies like this that package up everything AI is doing and the learnings and then it becomes an end to see. And then it's like the Q's, it's like SPY, it's something like that where it's like, you know, you just have these systems that are just driven by AI.

[4:05:59] Anthony Pompliano: I know we're running out of time, but what do you think the biggest area of opportunity is in terms of like things that are missing that would help you be better? Are there certain products that you want, certain features that you're like, hey, if somebody could build this, we could plug it into our system?

[4:06:11] Mike Gonzalez: That's a great question. I think there's anybody from like the SEC listening. I think the beautiful thing, and you and I know this, you and I have been talking about this for years, but the beautiful thing about an ETF, if you have it in a basket like the SPY or the Qs, when they rebalance, there's no taxes on that, right? You only pay taxes when you leave the actual ETF. And so when you think about a product like this, or say somebody's in their individual portfolio, right? If you want to buy a basket in your portfolio, And you agree to sell that basket, you know, when do you pay taxes? It's not as clear if it's an individual brokerage account, but if there's something like, you know, us offering some of these products, ideally if it rebalanced, people don't have to pay taxes,

[4:06:54] Anthony Pompliano: things like that. Tax mitigation, never is a boring topic.

[4:06:59] Mike Gonzalez: That's the thing I'll leave with, is like I know there's all these interesting securities, all this interesting stuff, but really when it comes to investing, and I'm like building an AI system for investing. I think it really boils down to just wealth preservation and just compounding, right? It's like the SPY. It's the Qs. It's whatever this index is, and it's just like set it and forget it, let it go. If you love investing and trading like I do, then 3% to 10%, just a little bit on the side just to have some fun with, but for the rest of it, index and chill.

[4:07:30] Anthony Pompliano: What is the website people can go check out?

[4:07:33] Mike Gonzalez: You can go to benchmarkai.app and we're going to be introducing a lot more of our research and all that. So appreciate you guys checking it out. Any feedback, you can email me at mike at benchmarkai.app.

[4:07:45] ?: Amazing. Mike Gonzales, everybody. Thank you.

[4:08:00] Anthony Pompliano: Am I supposed to give you the microphone? That's okay. I don't want to get in trouble. All right. Next up, we got Ryan Cohen. Ryan, I don't know if you were there or if you could. Oh, there you are. Look at your beautiful face. How are you? How are you?