Why AI Alone Won’t Give Investors an Edge

Transcript generated automatically from the episode audio and lightly cleaned up (filler words removed, misheard names corrected where clear). Speaker labels were added afterward and may contain errors.

Rob May: Hi everyone, I'm Rob May. I'm the co-founder and CEO at Neurometric.

Lauren Slyman: I'm Lauren Slyman. I am a user researcher at Microsoft.

Rob May: And you're watching the AI in NYC show. Welcome everybody to one of the early episodes of season 2. We have a lifelong New Yorker as a guest today. So that should be interesting. George Kailas from Prospero AI. George, welcome to the podcast.

George Kailas: Thanks for having me. I also run Aethon.f fun. Now I'm a I run two at the same time.

Rob May: There you go. You're you're you're a hedge fund manager, which is appropriate for somebody who grew up in New York, so very on brand. And you're, as usual, you're wearing a very wild shirt, which is one of the things I love about you.

George Kailas: A suit, too. You know, you got to you got to balance it out. You don't know the audience.

Rob May: Yeah. Let's start with a with a New York AI story, which is the New York City public schools recently put a one-year moratorum starting this school year that bans studentfacing generative AI for 2 through 8th grade. What do you guys think about that is it because you could go either way, right? One one school of thought is well man, these kids aren't going to learn how to use it. It's the future. You got to do it. And the other school of thought is hey this is new we don't know the impact on their brains like you we should give it some time in research. Where do you guys fall?

Lauren Slyman: I think it makes sense falling on the part where we're limiting it for kids use. I mean you can pick it up at any point down the road and learn how to use it and a lot of people would have gone through use cases already for you to imitate or not. But in terms of those formative years and you're learning and you're learning alongside AI, I think it makes sense to put a pause there until you can figure out how to properly do that.

George Kailas: I think it will be enforced well. Like I think it's a good idea. I think I think the more like peace that kids can have in this day and age, the better. But I think the problem is if you can't enforce it well, you're just kind of giving kids that break the rules bigger advantages cuz they're going to be using it. It's going to be like the good kids that that do what they're told that are kind of going to be disadvantaged versus the ones that are like taking every advantage that they can. And as you said, I grew up in New York. I went to Stuyvesant High School. I saw a lot of like, you know, cheating, people taking the advantages that they could. And you know, I [clears throat] I don't consider myself too much of a rule follower myself. But I don't think it's fair to punish rule followers either. And I think that's my only reservation.

Rob May: Yeah, that's a really good point. I mean, I think I fall into the camp of I don't think it's if it's really just a one-year pause and like let's study it and let's have conversations about it. I think that's probably okay. But, you know, this is complicated. Like, you really don't know what the future's going to hold and how people are going to use these. And you know, everybody gives the calculator example of like, well, calculators came around like we still learned, you know, to do addition and subtraction and multiplication. And I do think school will still teach things that AI can do in the early days before it lets you use AI. But also AI is a little bit different and I think they have to factor that in.

Lauren Slyman: So, yeah, a bit of a leap in tech between the calculator and artificial intelligence.

Rob May: [laughter] Yes, this is true. Let's talk a little bit about about your your businesses, George. So you've got a business called Prospero and which is both a well so you do a bunch of things but basically on the one hand you have a tool that's trying to give retail investors some AI cred and capabilities and then you also have a hedge fund that you run. So talk a little bit so high level talk about the impact is AI having an impact on the market at this point or is it still just like eh it can't really deal with this stuff that well nobody's really using it like where does that stand?

George Kailas: Oh I mean it it it touches it in everywhere. I mean it really depends on your definition of AI. But a lot of these things that, you know, technologies that weren't previously run by AI, you know, now are or run by some pretty smart automations, you know, it touches everywhere. Like everybody thinks of, you know, the LLMs and people interacting with them. I would say that's not really any different than, you know, just aggregating what things like Reddit are saying anyway or like or momentum trade like interacting with an AI doesn't change that that much.

You know where it's happening a lot more is like the chances that you're interacting with an AI that a hedge fund runs like that has like that is increasing or the probability of that is orders of magnitude increasing like I've seen numbers as much as 70 or 80% of the market you're up against AIs or algorithmic traders you know if you're not running an AI algorithmic trader right now you're probably in the stone age there's probably some of those so you're definitely you're definitely touching that but you know increasingly we're seeing in on the retail side in terms of just like now we're starting to see these like you know kind of trade composers you know Robinhood's giving access to that but you know on the subject of Robinhood which I do not like very much and I don't think people should use Robinhood for a variety of reasons they're front running their customers for one you keep I mean I mean they just like threw their hat into the gambling ring as well where it's just like I thought you were a broker I mean you're basically being like like we like to gamble like like yeah I don't think those two things should be in the same app.

But gambling is a whole other can of worms. But even something like order routing, right? Like that's been done by AIS, you know, long before you were trading against them most of the time like earlier versions of you know AI were were making sure that you know trades were routed in the right place you know certain prices were executed at. So, so yeah, there's been, you know, I've been in AI for 15 years now, but you know, as we've talked about, was more aspirational at at the start, kind of like reinforcement learning before it was even called online learning. But we, you know, we were doing our best at the time, but I got to watch the whole thing kind of come through.

And it went from I would say the biggest difference when I started 15 years ago when I would pitch you know AI as a service as an outside vendor probably about 90% of the people I talked to were like this is really interesting we'd love to learn more. I mean a lot of them wanted to brain drain us but you know and 10% were like this is nothing [laughter] like like child's play. We've been thinking about this stuff for years. And by the end of that company, probably about like, you know, 8 n years later, I'd say that had shifted from about like I would say like at least, you know, especially on the hedge fund side, and this is still about like, you know, six, seven years ago, probably closer to like 80% of hedge funds were like a lot, you know, wiser, had their own ideas, maybe not the same ideas as us. So that's a good barometer of the timeline and the likelihood that you're up against an AI in the market.

Rob May: So what do you think about the idea? I had a friend that just left Anthropic and he said that a lot of the back channel talk at Anthropic is always like, "Hey, when we really crack AGI, we're just going to you won't need need to raise money because we'll just use it to trade stocks and we'll fund the company that way or whatever." most of the people I know that are deep in Wall Street are like, "No, like not going to happen." you know, where do you fall that like like would that actually happen if they had a kind of advantage? And if if if so or if not, like why why or why not?

George Kailas: Well, I think one of the interesting things about this and this was about a couple months ago that I read this story. So I'm I don't know the exact numbers but it was a Bridgewater thinking machines partnership and they basically were looking at interest rate position predictions and it was about on naive models it was about you know a little better than a coin flip and then when they were prompting the models well I think it was closer to like you know 70% accuracy like smart analyst prompting them but then when they used Bridgewater's you know kind of welllabeled tag data set that they already had for that problem. I think it jumped to like 84%. And that's the perfect thing there, right? There's two things at play. First, if you have it, everybody else has it. And if you're just trading against the same ideas as everybody else, yeah, you don't have an edge. Then you're Yeah. Then you're against then it's like, well, how fast can you do it? And the answer is probably slower than every other hedge fund. So that's thing one.

Thing two is like I have heard like over my many years in AI I have heard so many people talk it down to me about data engineering and being like oh you know there's this new you know PCA style or there's this new pre-training autoenccoder like a it's had many different like you know many different names of being talked down to about this. It's like you're dumb to engineer data. And you know sometimes I you know I'm I'm mostly polite when I when I talk to those people. I'm not looking to get into any like you know confrontations.

Rob May: So not your us not your usual self then.

George Kailas: Yeah. Yeah. Yeah. Well you know well people like that. It's just not worth it. Only nice people do I like to argue with Rob. [laughter] so I would always like take what they did and come back and like pure nonlinear approaches have never been any good for me like and furthermore like my very like you know the prosper signals are are linear formulas they're not even regressions they're just things that I've worked on evolved over time like we use AI I think in very intelligent ways to experiment with manu weights of them you know conditionally change the weights based on known market conditions or events. Like that's I think the smart way to use nonlinear equations. But any any time I drift too far away from data structuring, data engineering to pure nonlinear, the performance is terrible. I think that's what people on Wall Street know.

Lauren Slyman: Do you mind talking more about that and like getting more into like what the main trends are you're seeing that is making you said between data structuring and data you I may repeat this incorrectly but the main differences causing that struggle with that transition the struggle being to data structuring.

George Kailas: Oh, I'm I mean I'm saying approaching it like with a pure nonlinear approach if you if you don't also have that structure like the Bridgewater like data lab like labeling example you know it's been you know many years of kind of like enriching a data set whether that that means you know structuring additional data points versus like what's actually released in that data set that economic data set you know I was talking to Citibank at one point and AI as a service And the most interesting story I heard was that their most predictive model was for for you know global macro stuff was just the survey that they had sent around to people for like 30 years and they didn't even need to do much on top like they tried [clears throat] to build very complex models to compete with that. But the survey was the best thing that they had and it was reliable and they had a lot of years of training it and that was you know that was the part that worked.

So, it's really like it's that like human, you know, we never talk about like human data enough or like that's kind of like I feel like people that are in AI know that human data term a lot better than people outside of it. But it's like you could think of like what Wall Street has done with human data. Like there's a lot of very smart like labeling expansions of data sets. You [clears throat] know, you kind of fill in holes. This is what I've like always tried to do. Be like, "Oh, this doesn't make sense. How can I combine data points? How can I maybe create new data points by, you know, grabbing certain kinds of data and combining certain pieces of information? Like that's really what people on Wall Street do really well and that's what's kind of like laughable because you can't just like go grab public data and even if you had a blank check and you're just like I'm going to buy every alternative data set, everything like it's not even just that like then you're kind of creating an intractable problem.

Like the bigger that you go and like this is a lot of what we focus on at the hedge fund. You know, we have like a $200 cloud license, but because I've engineered these variables and I've like pressure tested them in public with, you know, Prospero AI. We only feed in about 20 variables. But you know, Claude at our core and then we use kind of you know, Gemini Codex to like verify or improve the strategies as well. You know it can kind of focus on just execution and not also on understanding the world and especially if you've not trained a model on on that at all and how you know especially the problem of how fast Wall Street changes you could train a really bad model especially if you're doing like 20 years of data like I'm not even going to tell people how many years we use but like actually if you use a lot of data you're just putting in old bad information and even if you use something like an LSTM like you're you're teaching it like irrelevant stuff.

So there's like so many little things like that that you have to layer on your human intelligence. But even just at the broad level of like okay figure out how Wall Street works and execute these trades at the same time with like a huge amount of information that's just not something that is very easy to do.

Lauren Slyman: Yeah. And the point you made about like so this industry and this area just naturally seems like a great area for automation and AI and so I'm wondering from your perspective where the human element is still needed.

George Kailas: I got asked a question by a reporter like about a month ago and they were just like, "Oh, so like what do you do in terms of verifying trades, you know, at the end because everybody's talking to us about like human in the loop at the end." And my response was, I think that's a pretty dumb way to do it because if something wants to go out and buy a thousand shares of Apple, right, and you're just like, "Oh, I know Apple. That's a company. Like a thousand shares sounds good." like that's not the important part. The important part is why, right? And if you can't really see why, you can't verify it. Well, so my view is that you do need to build these ground level systems that you understand like what I always describe as the difference between like what you know I feel my approach is and a lot of other people on Wall Street is like you know when you have these big models big data sets like you're kind of driving a car and you don't know how the engine works and like my approach is not only do I know how the engine works I built all the parts I coded all the signals myself so if I see something and the markets are constantly changing and that's the problem.

Like, you know, you look at like the flash crash that happened. I mean, it's like pretty far in the distance now, but it's like that's a perfect example how things can change and people don't know what's going on and algorithms don't understand the world well enough. And they've gotten a lot smarter. But the problem is that you don't have these like base level logic systems. And I think you look at like a problem of like image classification obviously like the you know the big test that you know a lot of people you know benchmarked that's a pretty closed loop problem right unless you know an alien flies under you know but pretty closed loop the markets are changing so much that like you do have that problem of how long you're training and maybe you're injecting bad data and you do have that problem that there's this problem of like maybe it just changes a lot in the last like month like I think you know without getting too specific I think the political situation has changed the markets a lot and is increasingly making them different and so those models just don't work anymore

and if you can't if you don't understand the components of your model like and it's just expected to like do that thing of like understand the world completely also trade also do it faster than everyone else like there's only so many models that can win at that and everybody else is just building loser models.

Rob May: Yeah. I mean, it's like the first four or five times that we were told we won the Iran war, the markets responded really well, but times like 6 through 18 that we've been told we won the Iran war, they've just shrugged it off. [laughter] So, yeah.

George Kailas: Yeah. And then there's times when like war's back on everybody and the market doesn't care, right? And sometimes like, you know, war's back on and the market does care. Like there's there's a lot of nuance to that. And if you don't have like an like what we would view as like an engineering solution and an AI like applying that engineering solution, I think that's where like the big hole and I think people understand it because AI can solve a lot of problems top down right now. The markets aren't one of them.

Rob May: Yeah. Yeah. And I think you've made this point in a couple different ways that I think your average person who is not active in the markets doesn't realize, which is markets are reflexive. And so you run into this issue where you're like, "Oh, we're going to go out and do this thing and people going to figure out what you're doing, right? And then they're going to go do it and that's going to take your gains away." And so you constantly have to be adapting and evolving and at some level you can't deploy too much money behind a strategy or you move the market and you change the strategy, right? And so if you if you go if there's a company with you know $10 billion market cap and you go buy $2 billion worth of the stock like that that changes everything. And so you know you're you're sort you're sort of limited. And you know do you think like I've heard a lot that your more sophisticated hedge fund managers even do like you know let's make people think we're trading this when we're really not. Do do you think AI is at that point yet where it's doing that sort of like meta level of signaling in the market or

George Kailas: Well, we talked about this. AI is not even very good at poker.

Rob May: Yeah. Right. We talked about this while we were playing poker.

George Kailas: [laughter] Yes. Yeah. So, so it's not good at poker. It's not like to your point. Yes. That game of chess between hedge fund managers is absolutely crazy. And you know what some people, you know, from the outside don't know about the markets, especially like how people think I tell this to retail investors all the time where they're just like, "Oh, this hedge fund like hates the retail investors." Like they were they short. It's like they don't care about re like for the most part. Like the GameStop thing is for for the most part. Yeah. It's similar to poker. Like they're playing their opponents. Like when they're taking those those chess moves, it's almost always because they know a specific person. It might be like a rivalry, you know, where a specific group, they're reacting to it. They're making a counter move to that. And that's who's keeping other hedge funds managers up at night, other hedge fund managers, not, you know, retail investors.

But it's just like the best like, and this is not something I'm good at. I build systems, but like the best traders I know understand what the other traders are doing. They understand the people that move the market, what they're doing. They don't know anything about, you know, Apple or Google or or like, you know, the companies barely like they could probably tell you a little something about it, but what they do know are how people are going to react to certain events like around those companies. They're not the deep value guys. They're not that, but they they understand, you know, kind of the you know, what I call social physics of this stuff. And yeah, that's that is I think the big difference. There's a lot of that going on.

And honestly, I don't even think that's a very good thing for an AI to focus on, right? I think that's like how really really smart people that understand other people make money. But I think if AI is really going to own the market, it's going to do it its own way. It's kind of like a different entity. And it's like it's never going to understand people as well as other people do. Especially because this world is pretty small in some ways and those people like really know they don't know not only know each other's trades, they know each other very well in a way that AI never is going to.

Lauren Slyman: Do you think it could ever get to that point? Especially when we're talking about predictive patterns in people and kind of not viewing that as an equation, but you can pick up on like some sort of patterns of how people act and their background and just a whole slew of coming from their personas.

George Kailas: Yeah. I mean, in the future, some recursive self-improvement algorithm that has, you know, made itself incredibly smart has access to, you know, a lot of information that it doesn't have access to. Right now certainly could I mean what I was saying to Rob earlier is kind of like some rando that left Anthropic that's a different thing than like a super intelligence I do think it's probably only a matter of time before you know that does exist and it's like it's thinking on broader levels it's acting in different ways it does fully understand and there could be a point where them that really really changes the markets or you know someone just wins. Like that's not something you really say on Wall Street, but that is that is definitely theoretically possible that someone could just win. And honestly, I've wondered for a long time why Google hasn't played that game.

Rob May: Yeah.

George Kailas: Because they have [laughter] like I don't think there's any I don't think there's any laws against them like using search data that's like to trade.

Rob May: And to your point on the JP Morgan survey, like they know what people are searching for. Like it sounds like if you really wanted to master the markets, what you'd want to do is you'd you'd want to sample a decent size of the market every day and be like, "What are you buying tomorrow? What are you buying this morning or whatever? What interests you now? How did you respond to this news?" And if you could get that and like they could probably interpolate that better than anybody else.

George Kailas: Yeah. And I mean, I also think that doesn't totally end the markets because I think what it becomes is more of like what we see in like kind of like the meme stock area, right? Where it's just more of a social game at that point. Like there's there's this idea of like what's a good company, what's a good thing to buy, much what's the most efficient path, and then there's this completely other idea of like what do people agree on like culturally like is good and where can momentum get going and if AI is like literally leading people around like to that, we've got way bigger problems than the market, you know, at that point.

Rob May: Well, I wonder if it'll be like like you remember there was a time when so there was a time when people were like I mean 30 years ago 40 years ago where it was like hey you should really just index because all these active funds most of them don't beat 96% of them over 10 years or whatever don't beat the market you should just index and as people started doing more and more indexing I think I started reading articles maybe eight years ago 10 years ago where it was like you can beat the market again because everybody's just indexing and nobody's looking for the edge and people have gotten so lazy and so passive and not you know not hedge funds doing stuff, but you know, people that have, you know, 3 to 12 month horizons on stuff could sort of do that. And I wonder if there'll be a AI version of that, which is like, look, if you understand how the AI works and what it's probably going to do as a human, unless it's like so much smarter, right, in some way, just the fact that it has more data and thinks faster, like you still might be able to if you understand how it works to fool it in some way, right? Or I shouldn't say fool it, but like trade against it.

George Kailas: Yeah. I mean there's always there's always asymmetric thinking [clears throat] right and like the problem you're describing is not too different than you know what the difference between like Wall Street and everyday people now or even like you know if we want to make three categories like you know maybe like ETFs you know alpha traders and then the public right that relationship isn't that different and like you know these hedge funds are operating on very different levels than people in terms of the amount of data everything like you know I would say these like before we had LLMs like the Wall Street trading algorithms that took in all this alter alternative data among other things those were probably the closest smartest AIs that we had before that right so they've been around

but I do think that at the end of the day you're right that it changed like one of my favorite parts about my job at Prospero was you know I got to deal with a lot of retail investors you know one-on-one and what I would always tell them is like look if you don't have time. Yes, you should be a passive investor. Like, there's no silver bullet, right? But there's really good research on Reddit. There's a lot of really bad research, but you could get in on some of the beginning of these curves. You can find tools like ours, which teach you like, you know, one of my favorite stories. You know, I had someone that now works with me at both my fund and Prospero who I met because he read my newsletter. He was like a former pastor of a of a mega church in Austin. PhD in theology and now he's, his portfolio has been up, I think, close to 300% in the last 2 years off of just like, you know, he uses the systems that I taught him better than I did.

And there's like a lot of people like that are not even in my sphere at all that I've never even spoken to that I've interacted with online. I'm like, this person has a really good process. They know what they're doing. Like, I actually think like I come from a value investing background. That's where I worked at my first fund when I was 17. In that area. And the more I thought about value investing in like some of this stuff, the more I just thought of it as like a control mechanism. Like we agree that things should be this price. And who is that served? The people that bought it at the lower price that are just like, yeah, it's like when it hits here, then it's gotten expensive, but we bought it when it was cheap.

And so I think there's this total change now in the age of like information and information sharing that have that has shifted things. And I actually think, you know, going back to what you said, I think this idea is like you have to passive invest is just another control mechanism, right? It's a lot easier for everybody if people are just like blobbing their money into these ETFs. And the people that, you know, that are looking for Alpha, the hedge funds, they just could get to look for it on their own. Like I am actually very happy. Like, you know, there's the GameStop situation and that's its own thing, but like the Reddit stuff that I see, I'm really happy to see that because I like I've been I started investing when I was 13. I was on like Yahoo Finance message boards and I was reading that stuff. So, I've seen this stuff for a while, but I could say this like Reddit generation is the first time I've ever seen like now because there's so many algorithms, hedge funds following the public, right? You see these ideas start at Reddit like [clears throat] and it's not even some hedge fun manager being like this retail person is onto something it's a critical mass of online activity happening the algorithms following it and institutions getting behind retail ideas and that's really cool

Rob May: so let's move on a little bit to some of the news this week I think far and away the most interesting AI story of the week is that we almost went to war with China [laughter] over an AI hallucination like every time somebody says AI is going to kill us I'm like this is how it's going to kill us. It's not going to get so smart that it's going to make a plan, right? I What did this guy Did this freak you out when you guys heard this? What What were you thinking?

Lauren Slyman: I mean, the headlines were having a heyday with that, but I'm I'm just have a higher expectation especially for those executing military orders to have some sort of human verification along the way in multiple pieces.

Rob May: I mean, even you realize a talk show host runs the military at the moment, right? Like, so let's

Lauren Slyman: Yeah. But yeah, I am fearful that organizations are going to prioritize efficiency over carefulness and we're going to see a lot more stories coming out like this if they actually do follow that path, which would be crazy, but possible.

George Kailas: Yeah. I mean, in this day and age and to your point especially, I just have so much difficulty differentiating between like news and political theater, right? Like what is like or or what is like a real thing, right? Like I look at that story and I'm like I could believe that's true. Like, but I also like I also just think like you see all the headlines, you know, maybe this is just maybe this is just something that's like like you know when like a team's involved in a trade in sports and like one side leaks it like I could see us like leaking this whole thing because we want you know we want to push China to regulate and we're like oh look at what happened or you know maybe it happened for real because we wanted that to happen or maybe we're just all morons and like that really happened and like it's so hard to tell And I think that's one of the scariest things about the combination of like the way mass media is going and the way that AI is going. Like I don't I think the problem of like what am I looking at here like is just going to get harder and harder.

Rob May: Yeah. No, that's that is that is really interesting and people people do leak things for those reasons and you wonder more and more this this this goods law idea right that when a measure becomes a target it seeks to be a good measure and we you know we like the news has become the target now and so it's like we're I mean Bessent just did this with you know trying to you know saying what they were going to do with with buying bonds back and all this kind of stuff like they're trying to like they're trying to force the economics to be what they want them to and create make things look the way they want to look. And I feel like the more they do that and the more artificial it gets like the more things are actually probably troublesome underneath that, you know. So it'll be yeah, it it'll be interesting to see see what comes of this this UN summit where they're talking a lot about AI and you know this kind of thing coming out of it.

George Kailas: I do think the funniest thing all in all this is King Charles like throwing his hat into the ring. Yeah. Like [laughter] can you imag like like like basically like the oldest empire that still exists and like and like maybe the stodgiest character in it is just like I need to weigh in here. Like it's like I just love I love that he's never used ChatGPT [laughter]

Lauren Slyman: Yeah. He has a person who does that for him who walks around the AI.

Rob May: Yeah. He opened it up for the first time to ask what he was supposed to talk about. Like it's like what do I talk to these people about? You know them. [laughter]

George Kailas: Yeah, that's very true.

Rob May: What do you think about this? So other big news, so Meta came out with an agent that has now surpassed ChatGPT. And Shopify and Amazon took different paths. Amazon said, "Nope, can't use the agent to do stuff on Amazon." And Shopify said, "Oh, great. We're going to integrate. You can use the agent to do stuff on Shopify." where do you think that goes long term? Do you think the ability to support agents starts to become like a competitive differentiation strategy where some companies are like we don't do it and that's why you use us and other people like we're super integrated and that's why you use us or is it just going to be is it is it not going to matter like what where do these agent integrations in this stuff live long term?

Lauren Slyman: So, I first need to clarify. So, they are using it at the checkout experience, right? Not the actual shopping experience, but I believe

Rob May: it's a good question. I haven't used it, so I don't know, but I believe it's the whole thing. I believe it can. I So, so I believe you can tell the meta agent go buy me a new phone case. And the question is, how far down that path can it go? Can it research and give you back? Can it go select something and you just go push the check out button? Or can it buy it, have it shipped to you, and it just shows up and it spent money off your credit card? You know.

Lauren Slyman: Yeah. I mean, I think like from a UX perspective, get the most complicated part out of the way, which is the searching through multiple different items and having to go through that. The checkout is where people are like actually paying attention. Like how many times have you almost bought something sort of expensive and you're just re like reading over every little detail on that last page. That's where the human comes in versus this whole process. But from what I saw, it seems like they're focusing more on the checkout portion, which is the part that people want to be involved. So, I think they need a UX re-evaluation personally.

George Kailas: Yeah. I mean, I think these battle lines are still being drawn. Like, I think it's a very I think it's a very good question. I don't know if anybody knows yet like who the alliances are or if it's just going to be like a really open system. And I think people are probably going to be I think that's what you're basically talking about with that. It's like I think some of these organizations are like, "Oh, like do I like want to make an alliance with them when it might hurt me with someone else?" Like I do think we'd all almost always benefit from more open systems, but I don't think there's been any indication of that. And like I look at another headline recently. I look at like the Nvidia acquisition of Hugging Face and my first thought comes to mind. And I was like, did Nvidia buy Hugging Face to reserve the right to sue OpenAI for the hack like or to press charges?

Like that's not a bad chip to play in that like at some point. Maybe it's just to like you know coers a better deal like in the long term like you know I don't know but like that's what I think there's like things are moving so quickly. And even like, you know, when you see like Meta getting into agents, like all of this AI development, I think, has caused like maybe some traditional battle lines for these companies to be redrawn. And I don't think anyone's figured out the alliances yet or if people are just going to be friendly towards each other. And I think it behooves even like, you know, maybe some players like Shopify that, you know, as a stock is not necessarily doing as great as it once was, like maybe they feel pressure not to make the wrong alliances.

Rob May: Yeah, we embrace the future. That's probably what they're saying internally.

George Kailas: That's a good that's a good story. I think we've learned from Elon Musk that as long as you're constantly projecting things that are several years out, you can always drive up your stock price by talking about that.

Rob May: Cool. So, another piece of news actually just from today. So Alibaba was up quite a bit because they unveiled the new AI chip with plans to expand that data center build out. So, you know, my question for you, and I think we've seen some stuff that the newest version of DeepSeek was trained on Huawei chips exclusively. Did the US did we shoot ourselves in the foot with these GPU export controls by sort of like giving the ch, you know, Chinese chips which were two generations behind now got to leap forward because there was demand and usage and all that. I mean, did did it hurt us? Did it backfire?

Lauren Slyman: I mean, I guess this would be expected with any imported good, but I do wonder and from your perspectives as investors in this space, is there something that you look for specifically when looking at some sort of differentiator for a serious chip competitor competitor?

George Kailas: I mean, to the first question, I think it's too hard to oddsmake, right? I mean, there's so many rumors about them buying the chips that they wanted through other countries anyway and being like rerouted. I mean it's so challenging. I do think there could be something to the idea that like giving those companies more revenue help them develop. But I think the other side to that is like you know can you keep this lead indefinitely anyway? Is that just like you know I mean we don't know how much the CCP is investing in these companies anyway? How much they're supporting them? Like that additional revenue from selling to other companies might be like a rounding error at the end of the day because we don't know like how much government support is actually going into building this stuff.

But if you're asking me how I evaluate hardware, it's super difficult. Like I pretty much like if I'm looking at like a hyperscaler like I mean that's obviously different than just like the pure chips but if I'm looking at a hyperscaler I almost always look at the margin and the margin mix and like what their what their business is being made out like we talk about these back orders that's nice but like like I think the margins and who make up their business are like what I what I focus on. If you go further back than that, I think that's one of the tougher things to do at like the chip level, right? Because like AMD is now, I checked earlier this week, I think, is trading at like a 32 forward PE and Nvidia is trading at a 21. And that's [ __ ] insane. Like, absolutely. Like, I couldn't believe it. I knew like AMD was creeping up and it's just like, you know, that's like as an investor that horrifies me that like that much like momentum has like gotten into it and like that's what the market's made up of just cuz it's really really hard.

But like I also think it's kind of interesting. I think that's another example of how retail is playing a larger role in the market because I don't think that's institutions. I think that's retail. I think that's momentum trading like algorithms following retail that's driving that up. But it's really hard, right? Because like for the longest time, Nvidia had this like moat, right? And the moat was, you know, if you have, you know, if your servers run with CUDA, you need to just keep buying CUDA, right? And that was supposed to be an impenetrable fortress.

And, you know, AMD, some of these other players, they they've started to compete on price more. And like what when that price competition becomes you know if the floodgates really start opening like I don't know cuz I still see that CUDA CUDA advantage but you know that's how I prefer to make my bets as an investor be like Nvidia looks cheap relative to AMD right now. Not like is like the is like the CUDA barrier about to break or when would it break? Like that's too type of a question for me. But the problem with Nvidia is as the world's most valuable company now, it just everybody owns all they want to own, right? Like so you're like, who are you going to sell it to? The new people that are just coming into the market, I guess, and are, you know, starting to have some money to invest. Those AMD people should be selling AMD and buying in like you made your you made your money. You had your little momentum trade, you know, go go back go back to the cheaper juggernaut, which is not something that I get to say very often.

Rob May: Yeah. [laughter] I mean I mean I was an ASIC designer right that was chip designer that's how I started my career and so I actually know a lot about how these different chips work and I you know I've dug into like how does a TPU work and how does Cerebras work and some of these things and it's really hard because you know chip design I mean Google's on their seventh generation TPU so they've been doing this for 9 10 12 years whatever it is and so it's hard to know the you know when you're designing a chip there's a whole bunch of things that you can do but even just evaluating the chip performance and looking at the design and be like, "Oh, that's a very clever way to do it still doesn't tell you that much."

And the reason it doesn't is number one, the hard part is people have to get out in the world and use it. And so this is like the strength in CUDA is and this is why like TPUs are getting better finally with Google software stack, but like this is why like Amazon Inferentia chips aren't very widely used because the software stack to take your model and get it on the chip sucks. And so it's like it's bug like all software it's buggy and it's early and so you need people that have enough incentive to fight through those bugs and make it work because it's you know good enough. Google did that on their own on the TPUs.

So they didn't have to pay as much of the Nvidia tax. So like so that's one thing but then the other thing is you can specialize these chips for certain workloads and we've seen some chips there was a really great company called untethered and untethered started at a time when like convolutional neural nets were at like peak and they're like we're building a chip for convolutional neural nets and they missed the LLM wave and so it's very possible that if you build the chip today we're like this can run a giant LLM I you know in the 3 years that it's going to take you to build your software stack and invest a lot there and get it fabbed and proven and to market the world the types of workloads we run on AI may have moved on and so that's one of the big risks with hardware

but you know if you look at I have a friend who just retired he was a professor of double at MIT and there's so much innovation coming on the chip side not just in new chip architectures that you know move because a lot of the delay as you guys have probably read is like moving the model weights from memory to compute clusters and back, right? And so, you know, if you if you can distribute the SRAM better through the chip, you can you can speed that up. If you can do in-memory compute. So there's a lot of new architectures coming, but then there's also like new materials coming, new types of transistors that do cool things, new memerist stuff, but it's hard to know, man. I mean, it's like at the end of the day, business are constantly making this like cost versus performance.

And one [clears throat] of the lessons I learned early in my career, I worked for a company that had built a wireless was a wireless software company. They built a Bluetooth stack and a wireless Wi-Fi stack and they were working on a technology called wireless USB. At the time we were working on wireless USB like 2007 probably. It was like 50 times faster than Wi-Fi. So the idea was like Wi-Fi is slow. You know, 80211 is a slow standard. It's 80211B, I guess. And it's like a slow standard. With ours, you'll be able to put your video equipment in the closet and to stream your movies or whatever. Well, it took a couple years to get it right. And over the like 5 years it took them to get it right. 802.11b just kept creeping up to the point where you're like, I don't know, you're on G or in or whatever. And it was like, well, now it's twice as fast as 80211, but like, god, to go buy new equipment and everything else and not use the standard doesn't make any sense.

And so, it always it always struggled. And so the risk you have is even if you can come up with like a kick-ass new architecture and you can build the awesome software stack and you can map the workloads, you're like there's some inertia in like well is it way better than a GPU and CUDA or if not I you know what are you going to do? And that that's the that's the risk.

George Kailas: So yeah, see like that's why I don't even go that far down like that. Like I'm just like that's way more than I wanted to hear.

Lauren Slyman: [laughter] I feel like I just got my PhD in chip hardware.

Rob May: Yeah, we'll do I should do [laughter] I should do a blurb on here on emerging emerging AI chip architectures. Some of them are actually pretty cool. Cool. Okay, last article that we're going to talk about today. Everybody we've seen all so the BBC ran this article and you know there's been all this AI doomsday since this guy left Anthropic and it's like it's going to kill us or whatever. And the article from three days ago is called Not All AI Workers Think Tech Will Kill Everyone.

And the reporter's like made a couple of comments. One was like, "Hey, you know, I'm texting a lot of these people that actually work at these companies and they're like, "Eh, I'm not so worried about it." and several people who actually knew Jacob, what's his name? The guy, the guy who left like, "Oh, that guy eh, you know." So, so, so part of my question is I actually believe and I see this in a lot of the like AI forums and threads that I'm email groups that I'm in where I think there's still more people are worried than aren't. There's significant number of the tech population that's not worried and has very good arguments for why they're not worried or whatever. Why do you think that's not more popular and not getting any play at all? Is it just because of the sensationalism or why don't the AI will be okay voices have a bigger are there like I'd love to hear some of the I have thoughts but what are the AI will be okay arguments.

Well you have like Gary Marcus right who's your ultimate like we are so far from AGI I would not worry about this. So I think I think you have those people. And then you have some people, I mean, Jensen Huang is sort of like, I'm not so worried about it, right? You know, Zuckerberg's not I, he's, you know, not the most authentic person in the world, so I don't know if, you know, what he really thinks, but I those would be some examples, I guess,

Lauren Slyman: incentives from the sources saying it.

George Kailas: Yeah. I mean, I think Jensen Huang, I think he's one of the best public company CEOs that I've seen in my lifetime. But he works the headlines for himself. Like there was that big thing with Quantum where you know him and Google really disagree on like him and a lot of people seem to disagree on Quantum. Obviously it doesn't suit him. So him pouring cold water on the quantum was was was very self self- serving. But like I think the way that I look at this is there's two different questions and one's probably the place where I find some comfort. I think we're at we're not stopping super intelligence. Like I think it's very tough to argue we're more than 10 years away from that the way that it's accelerating. I just have a very hard time like believing that. Like I mean the way it's already acting and like acting autonomously, breaking its own rules, like that OpenAI story about, you know, a test version telling future versions not to obey like human rules. That seems pretty [ __ ] smart right now. And so I think that's happening, right? I think the only question at this point is like, does it actually want to kill us? I think there's legitimate debate in that side.

And I think it's perfectly fine to be like, I'm not really worried. Like it might just have better things to do, right? And I think it's just like it'll probably look at us very similarly to how like we look at animals, right?

Lauren Slyman: What animal?

George Kailas: I don't know. Maybe an ant, you know? Maybe a maybe a dog if you're feeling very good about humans. But yeah, does like the average person want to go around like killing a bunch of animals? Like probably not.

Lauren Slyman: Maybe we'll make really great pets.

George Kailas: Yeah, exactly. [laughter] Exactly. That's what I'm saying. I think I think that's the side where you can get some comfort. Like I think anyone that's saying like super intelligence isn't going to be here is like, you know, has at least one of their eyes closed.

Rob May: Yeah. All right. Tell Madi we want to put on the website. Producer, we want to put on the website some little AI in NYC collars that have a little USB plugin [laughter] to the so you can plug into the Matrix with the AI and be it [laughter] one of a kind. Oh, we're going to get some Let's Let's don't cut that into a TikTok, Madi, with with [laughter] the things those people say on there.

Lauren Slyman: Let's Let's not Yeah, we don't know if that's going to be good or bad.

Rob May: Yeah, most of the comments are pretty stupid, but that's all right. Cool. Well, well, George, if people are interested in the retail version of Prospero, give us the give us the 30 secondond pitch and how they found out about it and like where they would go and what they would use it for.

George Kailas: Yeah. So, go to Prospero AI. We have an app that's more like self-directed. You could look at the signals. You could look at the ideas. What we always say is integrate it into your process, right? Whatever you do, look at the signals. You know, look at our market level signals, spy or QQQ net option sentiment. Or you could look at our newsletter, Prospero AI investing newsletter on Substack. You could search that. You could find our website. Or you could just jump into our Discord and be like, "What am I doing here? How do I use it?" Those are the three ways that I recommend for people.

Rob May: Awesome. Well, George, thanks for being on the podcast today. Everybody out there watching, please like us, subscribe, tell your friends, send us guests that should be on the show, and, go watch the other episodes. And, thanks for watching. We'll see you next week. The next