← Excess Returns 
Excess Returns
Is AI Still in 1995? Gene Munster and Doug Clinton on the Next Phase of the AI Boom
00:53:03
Loading summary
Transcript116 lines
- [00:00]
Windows 11 Student Offer Announcer
Study and play come together on a Windows 11 PC. And for a limited time, college students get the best of both worlds. Get the unreal college deal everything you need to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft 365 Premium and a year of Xbox Game. Pass ultimate with a custom color Xbox wireless controller. Learn more@windows.com studentoffer while supplies last ends June 30, terms at aka mscollegepc trading
- [00:31]
Schwab Trading Education Announcer
at Schwab is now powered by Ameritrade, bringing you an expanding library of education with even more ways to sharpen your trading skills. Access new online courses, insightful webcasts, articles, engaging videos, and more, all curated just for traders. Plus guided learning paths with content designed to fit your unique interests. No sifting to find exactly what you need so you can spend your time learning to trade brilliantly.
- [00:57]
Doug
Learn more@schwab.com trading what AI really means, just as a first principle is we're converting electricity into intelligence right now. Like that's exactly what's happening. And so if that's true and the demand for intelligence is seemingly infinite, I think the demand for power is seemingly infinite.
- [01:18]
Gene
That's a very basic question that investors should ask themselves. Do you believe the management in these companies is competent at seeing the future? And if the answer is yes, then they should be rewarded for making these investments and if the answer is no, then they should be penalized for that. I do believe we're going to see in the near term, we'll call it a five years more acute knowledge worker unemployment than we saw around mobile or the Internet. But ultimately I think it does fix itself because people realize you got to get on board.
- [01:55]
Doug
We test about 700 different stocks across all these models and I'll give you the drum roll to see any guesses on on who's the top model right now before I reveal it.
- [02:09]
Justin
Doug, Gene, thank you guys very much for coming back on Excess Returns. You are in high demand these days. So the fact that we can get you for, you know, 45 or 60 minutes is we really appreciate it and our audience does too, because you guys always have a lot of great things to say when it comes to technology. And I always appreciate your ability to, you know, explain these things in a way that our audience, I think, can get, you know, a lot from these. It's not you guys can go in depth when you need to, but at the same time you can kind of talk, talk high level. So I think a lot of this conversation today will be high level. But then we'll get into some, some of the details too. Our audience can learn more and follow Doug and Gene of Deepwater Asset Management and also learn how Doug is and his team are building and constructing investment strategies, benchmarks and actually we're going to, I think have an opportunity to look at a pretty cool tool that you guys built over@intelligent alpha.com so lot to get through today. Thank you very much for joining us. And we wanted to start, I want to start Doug with and you wrote this tweet which we talked about I think last time you were on and we'll put up on the screen here. But you know, you wrote and this was back at the end of 2023. My highest three to five year conviction idea is that AI will culminate in a bubble bigger than the dot com bubble. It's the nature of major tech innovations to create bubbles. AI isn't close to a peak. We're in 1995 and, and I, I would think you can correct me if I'm wrong. I mean that's, that's kind of going according to plan. What wouldn't you say? And then I guess, you know, what do you, has anything sort of changed your view on this or what's the current state from your perspective?
- [03:52]
Doug
I would say so far so good in terms of the prediction that, you know, AI will ultimately be a bubble. Maybe it's a weird thing to say when you're sort of predicting a bubble. But the thing I think that has changed for us is if kind of the end of 23 was 1995, that would imply we're in, you know, 1998 now. I don't think we're quite in 1998. I think it might actually still be closer to 1995, 1996. I think there's probably still more room to go on the AI trade. Setting aside, when do we get to a bubble? I think there's probably still a few years left in the trade. When we think about what are the bottlenecks in terms of building data centers, in terms of powering data centers, I think is probably the biggest bottleneck. But then also the demand that we're seeing from these services. I mean, Claude Code I think has totally unleashed the ability of AI to really be effective in enterprise and productivity. And we're just really starting to see the beginnings of that being adopted at enterprise.
- [04:56]
Justin
What are your thoughts on and maybe you can explain what the Claude Mythos sort of is and the technology behind it and sort of how big a jump Some of these new models are in terms of the development, you know, as, as users of AI, we're kind of stuck in the current models that we have. But I mean, I know you guys know and test and look at some of these frontier models and then is there anything to be said for that development tying back to some of the incredible performance we've seen out of the semiconductor stocks and sort of other related stocks in the market?
- [05:34]
Doug
I, I do think there is a tie there. And really, if you, if you try to find like what was the real catalyst to a lot of the run that we're seeing now, especially in the semi side, I think it was probably when anthropic released Opus 4.6, it was late last year. And there was something about that model where I think it made this idea of using a coding tool, a coding agent accessible to the mainstream. Like you didn't really have to know that much about programming to really build code at that point. And the reason that's important in my mind is if you think about every sort of knowledge work that someone might do, I think it's all reducible to a computer program. And so being accessible, making that concept accessible, describing what you need to do, having it reduced to code and then just having a machine do it. I think that was a totally new paradigm. That really happened in November. I think it started to then really spread to the masses. You know, the early people got it November, December, January. I think it really started to catch fire and spread to the masses in roughly March. And that has coincided with this huge rally we've seen in semis where I think the light has just turned on for a lot of people that AI is truly powerful. We've had this question dogging AI for two years now, since this really began, of when will we see the productivity gains, when can it actually do something useful? And we are absolutely in the days of utility now. And I would even argue, you know, you kind of asked about the progress of the models. I think a year ago, these models, you could compare them to like a high school graduate. I think now the models are probably equivalent to someone who has graduated college, maybe two years in the workforce. And by the end of the year we'll have models that are people who are well tenured, five 10 year employees, PhDs. That's how fast it's getting down.
- [07:30]
Gene
Does that mean, Doug, does that mean we're going to be at general intelligence?
- [07:35]
Doug
Well, genius, you know, my, my quirks around the idea of general intelligence. Like I, you could make an argument that we're in general intelligence now. Like so, so many of these debates about AGI superintelligence are very semantic because I don't think there's one uniform definition of like, what is AGI? What is super intelligence? What I would tell you is if you go and use any of these models today, they are capable of probably answering or figuring out, you know, 95 to 98% of whatever you would throw at it with pretty decent accuracy. And so, I mean, is that general intelligence? That seems like pretty intelligent to me.
- [08:20]
Gene
Yeah, yeah, I think it's a kind of silly conversation, but it's one that kind of orbits around the utility of these models is when we get to general intelligence or some. Yes, they hallucinate. Humans make mistakes too. I want to pick up on another point you made, Doug. You talked about that kind of explosive growth that happened with clog code and the new model back in November. And of course Anthropics revenue going from a 9 billion to 45 billion run rate over a four month period. That's like breathtaking. But you said the, I think you said mass adoption or widespread adoption. Like the reality is, is that we're still not when it comes to vibe coding. Like when you say mass adoption, you mean like within people who like experimenting with tech? It's not the average person has no clue how to even spin up and start cloud code.
- [09:15]
Doug
Yeah, I think when I say mass adoption, I mean more at the enterprise level. And you just referenced those anthropic numbers, you know, going from 9 to mid-40s in just a few months. I think that is the definition of sort of wider spread adoption at the enterprise because almost all that revenue is incrementally from enterprises that are deploying these models. And I mean a few kind of just anecdotal data points there that I think are really important. The CTOs of both Uber and ServiceNow have both said that they basically burned through their entire budget for inference this year in like the first four months of the year.
- [09:54]
Gene
Oh my goodness.
- [09:55]
Doug
And now they have to go back to the drawing board because their companies and their employees who they're giving these models to, they're finding so much utility now in using Claude code or Codex that the amount that they probably needed to budget was like 2, 3, 4, 5x what they did. And so think about what that means for forward numbers and demand.
- [10:15]
Gene
I, I talked to, I'm not going to name the company. I'm just going to give a range of a tech company that has a market cap somewhere between 5 and 25 billion. I want to give a nice comfortable range here, but it's a real company and they mentioned that they think that automation could have a massive impact on their white collar, their knowledge workers. And I guess the question as we think about these models getting smarter, does it matter that there does the whole unemployment thing or the impact of jobs? Because I think that's what I hear in this conversation is like what does it mean for for me a lot of knowledge workers listening to this. How do you think people should view what some of what we've seen, some of what we're picking up on looking at how smart the models are?
- [11:12]
Indeed Sponsored Jobs Announcer
When you need to build up your team to handle the growing chaos at work, use indeed Sponsored jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications and more. Spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this show will get a $75 sponsored job credit@ Indeed.com podcast. That's Indeed.com podcast. Terms and conditions apply. Need a hiring hero. This is a job for Indeed Sponsored jobs.
- [11:43]
Lowe's Memorial Day Announcer
Pool days call for cookouts and lots of laundry. This Memorial Day at Lowe's, save $80 on a Char Broil Performance Series 4 burner gas grill. Now just $199 plus get up to 45% off. Select major appliances to keep dishes, clothes and food for fresh. Having fun in the sun is easy with us in your corner. Our best lineup is here at Lowe's Battle 527 while supplies last selection varies by location. See associate or lowe's.com for details.
- [12:12]
Doug
I think AI for any individual, it can either supercharge you or it can make you irrelevant. It's about that binary in my opinion. And so anybody who is worried and they haven't yet really adopted and embraced these tools, I think you need to go as fast as you can in the direction of figuring out how to use them to do your job better. Because I mean we've always had this thesis. I mean Gina and I have talked a lot about this at Deepwater and intelligent alpha. It's 80 20. It's Pareto again. The 20% of employees who are super high performers who figure out how to use AI, they're still going to be very valuable to companies. But it's the 80%, right? It's the marginal person. It's the person who's maybe afraid of AI. It's someone who's just kind of skeptical. I think that those people are in Danger, especially in the knowledge work side. And so there will be disruption.
- [13:06]
Gene
And ultimately do they just get religion and then able to kind of keep their job or do their jobs go away and does it matter?
- [13:14]
Doug
Some of them have to go away? I think, yeah, I think some of them have to go away. Naturally, if AI is as good as we say it is, if AI is good as we all think it is, it will replace some jobs. But new jobs will come, as they usually do, for different tasks that the models can't do. I mean, we've talked about the idea of what data is useful. Just kind of like conceptually, the most useful data in the world is data that the models don't have access to, just by definition. And so I think there will be jobs. We call them detectives, but people that go out in the world, can they find this useful unknown data that the models don't have and bring it back into the enterprise, Give it.
- [13:56]
Gene
Maybe a huge segment of the workforce are detectives.
- [14:01]
Doug
Maybe that's.
- [14:03]
Jack
By the way, that's the kind of question that people are asking a lot these days, which is if this is the most disruptive technology we've ever seen in a positive way, like how much is it going to be disruptive in the short term to get there? And you know, with all of the revolutions, the new jobs have come, but the question is, is the pain getting there going to be a little bit more or maybe a lot more than it's been in the past? Do you have any thoughts on that?
- [14:22]
Gene
Doug and I have debated this and I don't know where you're, where you're standing currently. My sense is that the next five years there's going to be more disruption than what we saw in other cycles. Of course, over the last 40 years, 40 years, 60% of the jobs didn't exist 40 years ago. So like, this is how humanity works. You know, the detective MO starts to gain momentum. But my sense is there's going to be some kind of a gap that will fix itself when education kind of changes. But it might be like a five year gap. And if I was going to put some numbers around this, I think we see a step up in knowledge worker unemployment. I use that, I think that is important to look at because I think it's representative how transformative and how useful these tools are. It's hard to say that because these numbers, we get numb talking about them, but they're like people's lives that are being disrupted and turned upside down. I do believe we're going to see in the near term, we'll call it a five years more acute knowledge worker unemployment than we saw around mobile or the Internet. But ultimately I think it does fix itself because people realize you got to get on board. I got to become the detective, I got to become the salesperson, the tastemaker. And they will kind of the free hand of the market will push them to develop the skills that are necessary to survive.
- [15:52]
Justin
Doug, I want to ask you, when you and Gene entered the debate ring, does he enter with the mean Gene handle?
- [15:57]
Doug
I'm probably usually meaner than Gene. Mean Gene's ironic for Gene because he's like the nicest guy ever. I'm the mean guy.
- [16:04]
Gene
Doug was talking about the enterprise and what's happened with Anthropic. And a question was we've seen OpenAI really push Codex and you can talk about some of the things that you've observed in terms of how good that is relative to cloud code. What's Google doing on this FRONT? We got IOUDIO coming up next week. You know, feels like they're still more focused on the making search better and Google Cloud. And I just haven't heard. Maybe I'm missing it. Like, what's their response to what's happened with Codex and Clog code?
- [16:38]
Doug
No, I think it's been unfortunately slow. And I would give you this perspective and I think a lot of different enterprises use these tools in different ways. At Intelligent Alpha, we think Codex is the best tool for actually writing code. So when we're putting something into production, we're using Codex to build that product. When we're doing like product development, when we were doing kind of earlier on stuff, when we're ideating, we actually, at least I do often I use Claude because I actually think it's, it's a little bit better of a thought partner than Codex or GPT5 5 at the moment, although 55 is really good. So I think you can kind of use these models in tandem. I think, I think that's the best way to optimize them currently. But we've also tested and played around with Gemini and Gemini Cli, which is basically their competitor to Codex or Claude code, and it's just not there. And I think it's actually a good point, Gene, where, you know, I think Google has done a very good job of integrating Gemini into search because a lot of people still just, we default to search. I default to search still all the time. I'll ask like, literally I'll ask an LLM type question in my search bar and I'll Get a decent answer usually from Gemini, by the way. It works. So, so they have a really great advantage there. But I think that they Certainly, of the three that we're talking about, of anthropic, OpenAI and Google, they're certainly the slowest I think, to really embrace the sort of coding revolution and really the agentic revolution.
- [18:12]
Jack
It does seem like on Codex, like when you talk to the elite programmers, like they were all Claude code people and it does seem like you're seeing like movement Towards Cloud, towards OpenAI Codex from like those elite type programmer people.
- [18:25]
Doug
Yeah, there's, it's funny like, and we've always said this because we see it as, we use the models to do, you know, portfolio management tasks with the tools we built, that intelligent alpha. But the different models do have different personalities. Certain models are better at better things. That's why there's all these benchmarks out there and you see different performance. But I do think that that is, is becoming kind of an open secret really is that if you want to write code, if you really want to build a useful product that's going into production, that's going to serve users, I think a lot of programmers are defaulting to codecs if they have a choice and if you're really just trying to do more product dev, then I think people are defaulting to Claude. What's actually interesting in that paradigm is there's like a higher order question which is, well, what's the bigger market? Is the bigger market to kind of do the higher order thing and ideate on product and imagine things and maybe build simple products or is the bigger market actually building production apps? I don't know, like, I think you could make an argument for either one. Certainly right now seems like the bigger market is for Claude, but we'll see over time.
- [19:33]
Jack
So Gan, on the model war, what are your thoughts?
- [19:36]
Gene
I'm going to defer to Doug, he's like deep into this.
- [19:40]
Jack
What do you think, Doug?
- [19:40]
Doug
Yeah, I'll, I'll tell you, the current rankings in my mind are GPT 5.5, Opus 4.7, Gemini 31 and GRO 43 are in my mind basically tied. And then there's everybody else, you know, we, we test a lot of these models.
- [20:00]
Gene
You put GPT at the top for
- [20:03]
Doug
me, GPT is the best right now. Yes. And before 55 came out I would have told you that Opus 4.7 was the best, Claude. So, so it does change. I mean the leaderboard does change almost every time a new model comes out because each Incremental new model does seem to be a little better than the one before. And think about the game too. I mean, these model builders all know they're all testing each other's models. They're all paying attention to the same benchmarks. And so when OpenAI releases a model or Anthropic releases a model, they want to be as sure as they can that everybody's going to sort of feel the same way. Like, hey, this one is, this is the best. I have to, I have to navigate to this one again. And 5.5 is the most recent model and so they're king of the hill right now.
- [20:51]
Gene
So one thing caught my attention today was this lawsuit that OpenAI has against Apple, basically saying that they've breached their distribution agreement. Apple, of course, is using more Gemini with Google and they're going to be, we'll probably hear at the beginning of June about them being able, you being developers, being able to more easily plug into different models. And my question is, isn't this a negative read on OpenAI? If they're out trying to take legal action on Apple, like, if things were like really cruising for them, wouldn't they just be like, we don't even need this, like, our demand's through the roof. But you mentioned it kind of caught my attention when you talked about GPT being at the top of the board. Because I've got this. I agree there's like fits and starts. And by the way, Rising Tide, I'm a big believer that OpenAI is in a great position. I think this is a trillion dollar plus public company, but just kind of reading at least the current score. It just seems odd that they would try to pick a fight with Apple.
- [21:59]
Doug
Well, I mean, you look at, you look at the Elon musk suit in OpenAI, but there's, there's a lot of litigiousness, I would say, amongst all these companies and you never know what angle they're trying to play. But I would say this, I think that, I mean, Jack asked a question a minute ago about is it kind of winner take all, is it zero sum? And I actually think that is related to what you're talking about, Gene. There's this perspective in the market that the model where it's not really zero sum, it's actually that there's going to be so much demand that whoever has capacity will be able to sell their capacity and therefore be a winner. Right. So let's say you're, let's say you have the best model, undeniable. Like you've won the game and nobody will ever catch up to you. You'll sell all the capacity that you, you have. Right. But if the demand for intelligence is as big as it seems to be, you're probably not going to be able to fill all that demand, given whatever your capacity is, because other people have agreements to use data centers elsewhere. Right? They have capacity elsewhere. And so then the second best gets their capacity filled and the third best and so on and so forth. And so I, I've kind of, I think that that view, and I've heard a few people kind of talk about that, I actually think that view makes a lot of sense given what we know about the market right now, which is the demand for intelligence. It feels like it's basically infinite. You know, all these model builders are capacity constrained at this point. And so, you know, if you have a model that is, it's really hard to do it this way, but let's just say it's.05% worse than the top model. But you have capacity, you're probably going to fill as much capacity as you have. That's my guess.
- [23:50]
Jack
Just does this play into the whole XAI anthropic deal? Because XAI was one that did have capacity.
- [23:54]
Gene
Right.
- [23:54]
Jack
And they've sold a lot of that capacity to anthropic.
- [23:58]
Doug
I think that's exactly right. And, and you know, if, if, if they had so much demand on their side that they were using that capacity, I don't think they would have sold it. I think that they are rational economic actors though, you know, and, and they said, look, we have all this extra infrastructure we built, we need to do something with it. And I think they also got the additional chip of opening up CLAUDE models to be able to use to XAI now. I think it's SpaceX AI internally so that they could use CLAUDE code, which was previously shut off to them and shut off to some of the other model builders.
- [24:35]
Jack
Forgetting about the revenue part of it though, on the model, it's the model's leaving each other all the time. Do we expect eventually one of these companies will jump way ahead or do we think they're all going to just be racing each other and they're going to stay pretty similar over time?
- [24:48]
Doug
I think for the foreseeable future. I think they're going to be pretty close. I think they'll stay pretty close. They're all too large.
- [24:57]
Gene
We've got, yeah, there's four, five is Meta now in that camp.
- [25:04]
Doug
Yeah, their new model on benchmarks, I haven't really Been able to play with it yet. We're trying to get API access on benchmarks. Their new model looks really good. Looks pretty capable.
- [25:15]
Gene
So we'd have. Remind me the name of Meta's model. I. I should know this. Willow or something.
- [25:21]
Doug
No, it's. Yeah, no, I'm blanking too. Llama was the old one. Let's pull it up here.
- [25:30]
Gene
So we've got.
- [25:30]
Doug
Was Spark. Muse. Spark.
- [25:32]
Jack
Spark.
- [25:33]
Gene
So we got gbt Gemini.
- [25:36]
Doug
Yeah. Claude Spark, Grok, Muse. Yeah.
- [25:43]
Gene
And then you got. Then the. And then on the other side of the planet you've got Baidu.
- [25:51]
Doug
Yeah. Quinn. It's different. That's open. Open source stuff.
- [25:54]
Justin
Yeah.
- [25:55]
Gene
But kind of western world. We got basically five horses in the
- [25:59]
Doug
race in the language model space. That's correct, yeah. And then you've got. Call it five open source, big open source players, largely in China.
- [26:10]
Gene
And in five years are there's going to be five still orbiting around the hoop.
- [26:15]
Doug
I would say in two to three years there's still going to be the general same structure we have. Five is hard. It's so hard to predict because it's moving so fast.
- [26:23]
Gene
Yeah.
- [26:24]
Doug
And I think like to. To. Well, to Jack's question though, like, here's where I think things could separate because basically right now all these providers are approaching the problem and in roughly the same way. You know, they, they all use transformer architecture. So the models are. Are built essentially the same way. They're for the most part trying to acquire the same types of data. So they're being sort of trained the same way. The one thing that I think is different right now, where it feels like anthropic's moving faster, is that they're using the model to improve itself. So they've got this recursive thing going. I think OpenAI is probably pretty close to getting there too, if not already there. They haven't really talked it as much and I think Google, it feels like IS and Gemini are probably further behind on that front. And so if there was a reason for one of these companies to get really far ahead of the other, I think that is the most likely reason is that somebody figures out a really powerful, you know, recursive loop where the model is just training itself super efficiently and the other providers don't figure that out because they're really not doing a whole lot that's different on like the training or the data side, whatever your
- [27:36]
Schwab Trading Education Announcer
thing, it could be anything. Canva helps you make that thing a thing. Canva is a simple online tool thing. It's A way to design, with our magic AI tool, things you can social media your thing, generate images or videos of your thing, make decks or presentations to show your thing. Whatever needs to be done for your thing. Canva can make it an even better and bigger thing. Canva, the thing that makes anything a thing.
- [28:06]
Columbia Omnishade Announcer
You can't reason with the sun. Trust us, we've tried. This summer, it's time to put that angry ball of fire on mute. Columbia's Omnishade technology is engineered to protect you from the sun's harsh rays that can burn and damage your skin. The sun is relentless, but so is our gear. Level up your summer@columbia.com to spend more time outside and less time slathering on aloe lotion. You're welcome, Columbia. Engineered for whatever
- [28:36]
Jack
and talking about the horse race of the models, this is probably a good time to pivot to Intelligent Alpha because you've got your own little race you're doing here in terms of this. But before we get into that, can you just talk about what Intelligent Alpha is and what you're trying to do there?
- [28:48]
Doug
Yeah. We started the project of intelligent alpha about three years ago. So it was mid 2023, it was a little after ChatGPT came out and we had this thesis that we wanted to figure out if language models could be good investors. Could they just beat the S&P 500? So we ran a bunch of tests. The tests look very favorable. And now kind of fast forward three years, we have two investment funds that we run using our language models, using our AI process to analyze stocks, pick stocks, manage the portfolio end to end. And within that, at Intelligent Alpha, a ton of the work we do is actually in assessing these models. Right. We want to know which ones are good at picking stocks and which ones aren't as good and why are they good, why are they not good? And so we actually just launched a new product called the Intelligent Earnings Benchmark where we use 12 different models. So we were just talking about the 10. There's a couple more that we kind of fit in there, but it's all the big players we were just talking about, like OpenAI and Claude. We also use a lot of the Chinese open source models to see how well they stack up. And we test them on the ability to predict a company's forward earnings, kind of the direction that those earnings are moving with the insight, hopefully being that if you get the earnings direction right, you probably get the stock right.
- [30:11]
Jack
It's really, really cool what you've done here because you're basically looking at each model individually and you're looking at how good it is at predicting these forward estimates, right?
- [30:19]
Doug
That's exactly right. And so if I scroll down here. So we have kind of our leaderboard if you, if you visit our site, Intelligent Alpha Co, we have our leaderboard here where we've run this process for several different quarters. We've got it going back to Q3 of 2025. We'll publish some of that data very soon. But we test about 700 different stocks across all these models and I'll give you the drum roll to see any guesses on, on who's the top model right now before I reveal it.
- [30:50]
Gene
GPT.
- [30:54]
Doug
Yeah, Gene didn't even cheat. I know he didn't look at this before. GPT is the top. And so we test these models just directionally. Did they get if earnings are kind of moving up or down. And then we also test magnitude, small, medium, large. We have buckets that have, you know, bands of what percentage that might mean for the accuracy. But yeah, as you can see, and we've seen this, I'd say across most of our testing there is a pretty consistent run for GPT. They've consistently been kind of the best model at the top and often we're also seeing that the closed source models, so the American models from OpenAI, Anthropic, Google and Xai, they all seem to stand out above the closed source models, which I think is a good thing. It's probably what you would expect given how much money is going to training these models. You would hope they'd be better at a general task like this. And so far through our testing that has been true.
- [31:56]
Justin
So this is all like financial statement type of data that's being fed or doing like natural language processing on earnings calls and stuff like that too. Like what is it, what are the inputs?
- [32:07]
Doug
I guess that's right. So we have basically built what's called a harness. And the harness is essentially a system where the, the LLMs can access a packet of data that we've prepared. So the data includes some of the things that you just talked about, Justin, the last transcript of earnings. What are some of the current estimates? Basically what is the street expecting for revenue and eps, historical financial statements, things like that. We package that all up into a consistent query that each of the models, they all get the same exact thing. So it's a fair test. And then we have them for each of the 700 stocks make their guess of, you know, where will revenue and earnings both go over the next quarter?
- [32:55]
Gene
What do you attribute the outperformance? I mean it's been a consistent outperformance, actually getting wider more recently. What do you attribute that to?
- [33:04]
Doug
It's I think a few things and I'll give you a few also just observations as we've done this benchmark and use this internally as we get these new model paradigms like kind of like we talked about earlier, you know, 55 seems to be better than 5.4 if we compare them head to head. And 5.4 was better than 5.1, which is model before it. Same thing has been true for anthropic with opus 474-645. And so I think part of the reason is the models are they're just literally getting better. They're just getting smarter, which is the most general term I could use. And that smarts, that general intelligence I think is reflecting in accepting this data and saying okay, here's the data that's been given to me. We were talking about base rates before we started recording. Here are the base rates. What are the expectations both for this company and also for the universe of large cap stocks. And here's what seems to be most likely to happen. So they're getting better I think just at that as, as kind of a general task.
- [34:05]
Gene
Do you think that fact that machines aren't emotional, you know, in the asset management business, we're in that business you stride to be objective and unemotional. But when you do introduce an idea to a portfolio, there's a natural feeling of wanting it to succeed. And I'm curious, do. Are the models quicker to cut off of a company, cut bait sooner than you think a human would?
- [34:36]
Doug
Yes, is is the short answer. Yeah. There's no sort of endowment effect that these models suffer from. They don't have any sort of, you know, bias because they did a bunch of work on something. Yeah. Thinking something's more valuable just because you own it already. The as far as earnings though, you know, like there's an adjacent thought to that which is these models aren't emotional but there's a funny byproduct to that which can be a negative. Right.
- [35:05]
Gene
They're not what promotional?
- [35:08]
Doug
Emotional. They're not emotional. Yeah. And there, there can be a negative byproduct of that which is when you need to make a really high conviction call. Like if you think a company is going to crush earnings. Some of these recent semi stocks that we were talking about earlier, there is a little bit of like a faith and an emotion in there because again, I'll go back to our conversation about base rates earlier. That's not going to be in the data. The model is going to feel like that's a risky call to say, you know, whatever. Lumentum is going to have an incredible quarter because the demand for optics is just off the charts right now. And so they might beat earnings by, you know, 30%. The models are going to be really, really hesitant to make a call like that because it just happens so infrequently in the data. And so that's kind of the other. The other side of the sword is, on average, these models are right very often. I think they're probably right more than the average human, but the average human might still have a really good, like, slugging ability. Like, if they get one call really right, they can still makes sense.
- [36:12]
Gene
So think of. Think of like, GPT is more. It's not gonna be up 40% in a year when the market's up 5, but it's gonna hopefully outperform kind of on a steady basis.
- [36:28]
Doug
Yep, that's right. Somewhere between that big swing. Somewhere between traditional quant and human. Yeah. Is what I kind of how I think about the models.
- [36:36]
Gene
Quant and human are the dog or
- [36:38]
Jack
the models that are best at, like, predicting the earnings revisions. Are those the same models that are the best at picking stocks, or do you see, like, different leaders in different areas?
- [36:45]
Doug
It's fun. It's actually, it's. It's really a great question because it is a little different. And so we look, we can kind of categorize that in two ways. Number one, the. The best two stock pickers. And this is something we haven't published yet, but I'll give a little preview. The best two stock pickers since we started doing this are Claude and GPT, in that order. And that goes back to 20, 23. A lot of different iterations of the models. And I would say CLAUDE actually gained some more ground more recently when their models were more powerful, in my opinion, than GPT. So, yes, there is a little bit of a difference. And then there's some things we do at intelligent alpha. We. We take the earnings prediction as like, one signal, and we put that into our process with a bunch of other signals and kind of marry it with other data. And so the, the way we kind of use the models to use this particular prediction is a little bit more like a human. You know, this is kind of one angle, right. Are earnings going to be good or bad? And then what is the relative valuation? I might look at momentum of the Stock, you know, do I think some of it's already priced in? You know, maybe earnings are going to be great, but maybe everybody already knows it. We kind of try to create a framework for the models to be able to think about things like that. But this is the fun part. To answer your question, if you actually take all that stuff away and just say, let's make a portfolio of the predictions for earnings, assuming that that is where stocks generally go, the best performing model is actually deep seek so far. Interesting in our tests. Yeah. And they were actually, if you go back to our screen, if you visit our website, they were actually in kind of the bottom half of accuracy. So they had good slugging as we kind of think of it. They had some of the big calls, really. Right.
- [38:34]
Jack
If we take a step back to investing in AI overall right now, how are you guys thinking about. I guess you're looking at stuff across everything, but how are you thinking about where in the stack to invest? Many people have said we'll move down from the infrastructure layer, we'll move to applications and other stuff, but it seems like the infrastructure layer is still on a massive tear. So, like, how do you think about that?
- [38:53]
Doug
Yeah, I'll give my quick take and then gene you, you.
- [38:56]
Gene
Sounds good. Yeah.
- [38:57]
Doug
Build them in. Because we have, we have, I think, lateral thoughts on it. I think about AI like the moment right now. What AI really means, just as a first principle is we're converting electricity into intelligence right now. Like that's exactly what's happening. And so if that's true and the demand for intelligence is seemingly infinite, I think the demand for power is seemingly infinite. And the thing that I feel most confident in still when we talk about this AI trade cycle is that we are woefully underbuilt for energy of almost all kinds. Whether we talk about NAT gas, I think nuclear almost has to be a big part of the solution to power all these data centers that we're building. That might mean small modular reactors, it might mean other things. I think alternative energy as well. Storing that is, is a huge challenge. There's a company that we've invested in in our private funds at Deepwater and our venture side called Antora, that does solid estate storage. I think that's going to be a huge theme. And so power to me is the thing that just, it makes the most sense that the demand is insatiable. It won't go away. Even if, you know, we start building data centers in different ways. If the model architectures change, if all these other things might evolve the demand for energy probably doesn't.
- [40:18]
Gene
My you know a lot of different data points you can pull out on this topic of like how much further do we have to go? A couple guide or maybe guideposts along the way here. One is that the currently we're getting stopped out and using these models more frequently today than we did a year ago. So within intelligent alpha. So what that means is demand as Doug hinted to talked about before, demand for the models is outpacing infrastructure. So we know we need more infrastructure. The second is that if you look at the kind of the key marker for this it's CAPEX by the hyperscalers. CAPEX growth A year ago at this time the expectations were that they would grow capex in calendar 26 by 10% over 25. It's probably going to be up 70% as it looks today. Next year the street's looking for about 10% growth in capex next year and our sense is it's probably going to be closer to 20 to 30. It's not going to be 70, but it's still going to be much higher than what people expect in part because there still is cash flow from these hyperscalers that continue to make these investments. On top of that, outside of the hyperscalers we're seeing industrial AI being built and sovereign AI. And so we kind of put all this together. The brain think of the data centers as the brain of AI and like the apps are an inference is the thinking around it. But the brain still is going to expand more than what people I expect. Quick finer point on the energy conversation Crash course on energy in the US 1958 was the first nuclear power plant and they basically ran a bunch of them. I think there was something like 50 of them or so were built till the mid-70s. And during that period the average increase in output of energy in the US grew on average 7% a year. I mean it is that's like wicked increase in in growth from essentially 1985 till 2022 it was essentially flat more people but more efficient H VAC systems. And so we basically saw that flatlining over the next seven to 10 years. This is from a White House paper also a Goldman report talks about that averaging increasing by about 3% a year, a little bit over 3% a year. And that might not sound like much, but 3% is a massive investment cycle. And so set a different way is that a lot of times the AI Infrastructure Conversation centers around GPUs and optical components cooling, things like that. But this Energy play is, even though it has had a move higher, is still underappreciated by Wall Street.
- [43:18]
Jack
Yeah, I was going to revisit some of the Dean's predictions from the beginning of the year. You've already, these are, you've mentioned something you got right here, which is one, that CapEx growth was going to be very strong, which I think we are definitely going to be right on that one. And the Nasdaq being up 10% or more at least so far. You're in good shape on that one. A third one we talked about though was IPOs and you had kind of at that point decided, I think that that one was not going to be right because you figured these companies might come out. But that's in the news all over the place right now. SpaceX and then maybe the anthropic and OpenAI. What are you guys thinking about that? Do you think those are going to IPO this year?
- [43:52]
Doug
I think they will. And we're recording this today on the day of Cerebras ipo, which last time I looked, which was probably an hour ago, I think the Stock was up 108%. So they've had a good day. Anybody who got in the IPO, good
- [44:09]
Gene
sign for future IPOs?
- [44:11]
Doug
Yeah, I think that's the bottom line is to me, I think that's a signal that. Not that like a SpaceX or an Anthropic or an OpenAI needed an all clear but, but I do think maybe the second tier of companies that might think about going public, they have to feel pretty comfortable with their prospects at this point after seeing the demand for Cerebra. So you think about a company like Databricks or maybe some of these other coding tool companies who have meaningful revenue in any sphere, which is Cursor or Cognition, which has a product called Devon. You know, these are companies that are valued in the tens of billions already and I think if they went public, I would have to imagine there'd be a lot of excitement around them, just like there is around Cerebras.
- [44:54]
Starbucks Frappuccino Announcer
Tomorrow morning is knocking. Stock your fridge now. How about a creamy mocha Frappuccino drink? Or a sweet vanilla smooth caramel maybe? Or a white chocolate mocha. Whichever you choose, delicious coffee awaits. Find Starbucks Frappuccino drinks wherever you buy your groceries.
- [45:10]
Jack
Do you think these, like, do these IPOs have an impact on the overall market like we've never seen? I assume these will be the three biggest IPOs of all time. Right when they come out, like, how does that impact I'm just trying to think about like the supply and like do you think that has any impact on the market when you IPO companies of three companies of that size especially
- [45:27]
Gene
Doug, like what does it mean for the mega caps? Yeah, we looked at assuming that source of funds.
- [45:33]
Doug
Right. Could be. And I think the indexes, I mean it's a huge question for them. And just as one reference data point, Aramco, Saudi Aramco, I think in terms of size and market cap was the biggest IPO ever. I think SpaceX will probably, will probably eclipse that pun intended. But Aramco I think was like a trillion dollar plus IPO for reference. The stock actually was up about 30% from the day it issued to about two weeks in and then kind of the market fell apart a little bit. So even for these massive companies it's not out of the question that you could have a pretty healthy move very early on at a tracex.
- [46:17]
Gene
So much more exciting too.
- [46:19]
Doug
Agreed. Yeah, biased but agreed. But you know, I think that what it means for the markets, what it means for potentially the other mega caps is as they get included in the indexes. And there's a lot of talk about how particularly for like the QQQ, the NASDAQ 100 index, there will be an early inclusion 15 days in for SpaceX. I would imagine that OpenAI and Anthropic probably get a similar deal. And I think you then do have probably a little bit of a source of funds coming from some of the other mega caps because those indexes are going to have to sell down and adjust their weightings across the various companies to get these new big guys in there.
- [46:59]
Justin
What do you, I wanted to ask kind of probably a boring question, but it's one that I've been thinking about sort of up until like maybe a month or so ago. I thought that the market was kind of maybe penalizing some of the Mag 7 and the hyperscalers for their investment into this and kind of questioning, you know, what is the payback going to be. But then like, I don't know if it was the earnings, their, their quarterly earnings that came out. Gene, you might have even been on CNBC that night on Fast Money or something like that because it was such a big earnings day and I feel like now it's, it's the, you know, the, at least the price performance has seemed to, you know, rotate back to the, the mag 7 is, is that kind of like. Right? And I guess what are your thoughts? Because in, in thinking through that I was like, well, Maybe Apple is the, is the play here because they're really not going aggressive into their, you know, capex spending and I thought the market might actually like reward that but it seems to have flipped the other way. So I don't know if you're, you have.
- [48:01]
Gene
Well Google had a step up too and just to kind of set the stage is that if we look at Tesla, Microsoft, Amazon, Google and Meta those five and Tesla is usually not included in the broader hyperscaler conversation but is relevant to this topic is of those five, Tesla talked about their capex this year being more than 25 billion. Three months ago they said it was more than 20 billion and stock traded down on that comment like meaningfully 3 or 4% on that comment. Meta bumped up I think it was like from the high end of the range from 175 to 185 billion this year. Stock traded down on it. Those two companies don't have cloud businesses. Google I believe they raised their bumped up what their expectations were like materially increased what they expect for CapEx this year and Microsoft did too. Amazon more or less was a wash. But both those companies, you know this the stock, if you look in after hours trading when those comments were made they take it took a few minute dip and then just came right back. So there may be something around investors feeling there's like a faster return on capex if you have a hyper, if you have a cloud business but that's about the I think the through line just in terms of how it trades around the quarter. You know the bigger picture is like the real takeaway here is that competent people believe that this is going to be more disruptive than what the market believes, what the analysts believe all street expectations are because they're putting their money where their mouth is. And so I see that as you know it's a very basic question that investors should ask themselves. Do you believe the management in these companies is competent at seeing the future? And if the answer is yes then they should be rewarded for making these investments and if the answer is no then they should be penalized for that.
- [50:03]
Doug
I think you could actually even make an argument if you look at I'll even narrow our set to just the hyperscalers, the cloud providers, Google, Amazon, Microsoft, one month on their stocks basically back to when they reported earnings to today. Microsoft is the worst performing than Amazon, than Google, Google's the best performing. And I think the part of the reason for that is going back to this excitement around anthropic anthropic obviously Premier Partner early, partner with Amazon. So if you're using aws, arguably most likely, and you have an AI product, you might be on Anthropic tooling. Google, they've signed a deal, they've made investments in Anthropic. And so my gut is part of the answer to that question is sort of what I think Gene's alluding to is all of them are building out this infrastructure, all of them are seeing this massive demand and two of them are seeing massive demand correlated directly to the hottest company in the space, which is anthropic. And I don't think it's an accident that their stocks are probably the two that have performed better than Microsoft, which is really tied to OpenAI still.
- [51:13]
Justin
What do you guys think about the opportunity in some of these second order space style, like the stocks that are currently on the market today that are, you know, have business lines or will have business lines in space? Is there anything there that kind of gets you excited? I, I personally am excited. I'm very interested in like seeing what comes down the road with that. And so I think what's the investment opportunity, I guess now and in the future there? And I think if you have a minute, if there's a company's business model that you know of that is like really unique, I think it'd be a good discussion because I think there are things happening, there's things coming down the pipeline that most of investors know nothing about. So I'd be very interested in your thoughts on that.
- [52:08]
Gene
You want to talk about that early stage investment we made on kind of like imagery, the satellite imagery company.
- [52:17]
Doug
Yeah, yeah. Well, I'll talk about a few different things. I think like space to, to us is exciting because it opens up this potential for new avenues to create energy. Going back to kind of what is like one of the fundamental sources of things we need. Energy provides the sort of ballast to create so many things in our lives. And also productivity. Right. So it's like those are maybe two things people don't immediately think of when you talk about space because it's like, well, okay, we're just, we're going to space. This is awesome. I was like, no, what is the purpose of going to space? I think those are two of the big things.
- [52:55]
Gene
You said energy is the first one.
- [52:57]
Doug
Yeah. Figuring out novel ways to extract energy from the universe and figuring out new ways to be productive. Those are the big things. And so from an energy standpoint, I'll go to the everybody's favorite topic to either really love or really hate. But orbital data centers. Google is rumored to be in talks with SpaceX to potentially create some orbital data centers. I think whether you believe in the physics of it or not is almost irrelevant at this point. I think the question is, is somebody going to try it? And if somebody tries it, it's going to be SpaceX, almost undoubtedly. And I think they will try it. And we should hope that they're successful because if they find that space is a place where we can put a lot of these data centers. One of the biggest issues with building a data center right now is getting local permitting done. It's brutal. No towns want to allow a data center in their backyard. And so if we can put them in space and if we can maybe even power them more efficiently in space, in the atmosphere, that'd be a win for everybody. It'd be amazing. And so that, I think, as an overarching concept, is probably the most exciting reason to go to space right now. The other one that I would give you is, and this has long been kind of this discussion about if we go to space and we're working in these zero gravity environments, does that open the door to maybe creating products that we couldn't create on Earth, where gravity is an issue? And I just saw this the other day. There's a company called Varda, which is a space company. They partnered with United Therapeutics to potentially develop and create drugs in space. I mean, I think that's really cool. Again, like, who knows if it works? But I think we should hope it works and that it creates something novel and new that we could never have done if we just maintained production on Earth. So it's those kinds of really, they're almost hard to conceive of things, but I think those are the things that we should be really excited about when we think about the space opportunity.
- [55:00]
Jack
Would you guys bet, yes or no, if we had to bet on, like, data centers in space, like, five years from now?
- [55:05]
Gene
I would say yeah, like, there'll be prototypes that will be. The performance will be pathetic, but they will be operational, which means that eventually we get there. It'll be effectively like 32 megabytes of Internet.
- [55:26]
Doug
I bet there's more than five operational, but less than 25. I'll give you a range. I think that's about right.
- [55:36]
Justin
Jack, you and I should have an internal goal about doing the first podcast from space. What do you think?
- [55:40]
Jack
Yeah, exactly.
- [55:45]
Justin
All right, guys, thank you very much. We always appreciate you coming on, sharing your thoughts with our audience, and we hope to see you soon.
- [55:51]
Doug
Can't wait.
- [55:52]
Justin
Thank you for tuning in to this episode. If you found this discussion interesting and valuable, please subscribe on your favorite audio platform or on YouTube. You can also follow all the podcasts in the Excess Returns network@excessreturnspod.com if you have any feedback or questions, you can contact us us@excessreturnspodmail.com no information on this
- [56:12]
Doug
podcast should be construed as investment advice. Securities discussed in the podcast may be holdings of the firms of the hosts or their clients.