Andrew Freedman - AI Insights, Integration, and Best Practices

Ladies and gentlemen, welcome to the Business Brew. I am your host, Bill Brewster. This episode features Andrew Friedman of Hedgeye Risk Management. Andrew is an equity research analyst at Hedgeye. He covers TMT and healthcare and on the LinkedIn machine. He describes himself as intellectually curious and honest, with a passion for financial markets and a competitive drive to outperform. If you've listened to the show since inception, you will know that Andrew's been on the show

before. We have in the past had nice discussions about cable and I think we talked Matt at at a time. Either way, Andrew was tweeting a little bit about his perception of people not really appreciating what AI was going to do to either the world or markets or whatever. And I saw a tweet that he sent and I said, man, you want to come on and have a chat. So I quite like this conversation. I would say that this is more in the spirit of kind of how the

show was born. And it's a it's a more laid back discussion. And by discussion, I mean Andrew teaches me a lot in this episode and I do a lot of listening, but he talks about how he's integrated AI in his life. And I have a lot, a lot to improve on in the way that I use it. But it's certainly become a real part of my life. And Mary Meeker has a deck that I will link to the in the show notes. I encourage you all to flip through it.

And I suspect if you're a listener of this show, it's you probably are playing around with some of these LLMS at a minimum. So I hope you all enjoy the show. I enjoyed talking to Andrew again and he may come back and do more of a single name discussion in the not too distant future, but this one is more one for you all to chew on and think about. I, I kind of think it's Andrew Sharon best practices. So I hope you all enjoy.

I enjoyed recording it. Ladies and gentlemen, thrilled to be joined by Andrew Friedman for I think Part 3 here, also known as Free Bird. How's everything going, man? It's going well. Yeah. Yeah. This is great. It's been a, it's been a little while, but you know, it's good to be back. It has been a while I I prepped for this interview, or discussion, should I say, by using Grok. There you go. So So Grok has summarized some of your AI thoughts.

Boy, let's hear it. Well, hang on, I got to bring it up. But the first, so first of all, to give people some context, Andrew has been tweeting a little bit about people underestimating what AI is going to do. Is that a fair characterization, Andrew? I would say so. All right. And one of the things that Grok has highlighted is your prediction of unemployment. So I guess, I guess before we really get into it, you know, just kind of where's your head at and and why is this so much

on your on your mind right now? Yeah. So I guess I actually that's really interesting and we can get into a little bit later. But the fact that Grok picked that up is also kind of also addresses a little bit of the risks associated with using certain AI models, right? Because you know, it will pick up on the content that it's trained on, whether it has access to, right?

Which in this case is Twitter or or X, which is a great source of information, of course, real time information, but it's also social media. And there's a risk because some of those things can be taken out of context. Oh, well, unemployment is a risk. And I put out this kind of dystopian tweet the other day around this future state of the world and what it could look

like by 20-30. And of course, anything could happen, you know, if it's sarcasm or any type of I guess tivity can be misconstrued as fact, but so just kind of say that. But look, I, I would say I have spent. Let me, let me start over. I told myself, right, That I made a promise to myself that when there is a obvious technology, technological disruption, new technology that comes out, that I wasn't going to just sit idly, buy and be a passive observer, right?

Yes. And obviously everyone likes to financial podcast, right? We all want to try to make money with what we do in our time. But you know, I, I was young when the Internet first came out, right? I was probably 10 or 12 when things kind of really started to took off and I was like building websites then and doing all these different things, but obviously not in kind of like an age or place where you can actually really, you know, do anything with that and capitalize on the way.

You know, we're kind of at a similar crossroads today. And you know me, Bill, like I'm kind of cynical, right? I'm not like if you put me in the cynical camp versus the Galaxy brain camp, like I'm definitely not in the Galaxy brain camp. I think of you as data-driven. I don't think of you as cynical. I appreciate that.

I appreciate that. But anyway, the point is that this is really an important part time in our history and I've spent the last year and a half really just going down the rabbit hole in a really big way on this and trying to understand how it works. Partly because, you know, I cover Google, Meta software, all these, you know, very large mega cap companies that everyone cares a lot about and a is and the main topic. But what is AI, right?

I mean, from a analyst perspective, like I felt kind of like a fraud, like sitting there talking about llama this and grok that it without actually knowing fundamentally what these things are. So I went out and I bought a computer and I built a computer with two NVIDIA GPUs ultimately like spent the equivalent of like a Toyota Camry really on hardware. Yeah, yeah, yeah.

Wow And you know, kind of just trying to figure this out behind the scenes and not that you have to go out and spend like 30 grand on a rig, right to to figure to do this stuff like, I mean, there's cloud solutions, all these things, but you know, I, I wanted to understand how the hardware worked, right? I wanted to go through all the pain points and the problems of trying to host llama locally on your computer. What limitations do you hit trying to train your own models,

fine tune your own models? What type of limitations do you experience and then hit those problems and then solve for those problems? Because that's for at least me, that's been the best way to learn is actually by doing. And that way I felt that I could actually understand what an LLM is, right? How does it work functionally? And that's been a really fascinating learning experience.

And, you know, bringing it to kind of today and all those learnings, I think that it just kind of comes down to a lot of a lot of it comes down to like psychology, right?

And I know you're always like AI always enjoy listening to your podcast because you often get into like philosophical conversations and very like introspective in terms of human behavior and how the mind works and all the biases embedded there in. But, you know, I think it come kind of like this whole concept of like people not expecting or knowing a very smart people rather. Was the tweet that you reached out on very smart people don't really understand the implications of AI?

And I think it comes down to just the two types of people, right? You have people that come at things from a scarcity mindset versus an abundance mindset. Yeah. You know, initially with any technology transformation, people look at new tools as kind of replacements for existing tasks, right? But actually what AI can accomplish and what I've been trying to do and, and mind you, I actually came at this my scarcity mindset and then it turned into an abundance mindset. Interesting.

You know, you can, these tools are enable, like can enable things that you previously thought you couldn't do and really creating value above and beyond just core, you know, basically replacing your, your daily tasks. So there's a lot to talk about. It's, it's evolving like very quickly. It's amazing how fast the technology is moving. But I'll just, I'll just pause there, I guess. So what since you were using Llama, I'm curious for your perspective on this.

What happened to Llama? Because it was my default engine in perplexity. And now it has just been removed from the platform And like, I can't even use it if I mean, I could, right? If I if I did it but I I go through my primary I, I use Gemini and perplexity for the most part. Yeah, so I I can't speak to specifically what happened with Llamas availability and Perplexity itself. So Perplexity is a rapper, right? And. Yeah, but it kind of disappeared a little bit, right?

I mean, it's an open source code and it was getting adoption and then I've heard less about it recently. Yeah. I mean, look, I think as I mentioned before, like the rate of innovation is happening pretty quickly, right? And you know, really things only took off a year ago and the open source models had, you know, the, the open source models will always have a place.

But I think if you look at just Llama recently and the competition and the rise of the closed models with, you know, Claude and Tropic open AI Gemini 2.5 is a great example, right? They're just so much better, right? And, and they're scoring on benchmarks, they're leading benchmarks, not that you know, you can't game the benchmarks. You can of course game the benchmarks a little bit. And I'm sure there's some some of that going on.

But I can tell you this, much like I've hosted Llama, I've hosted DeepSeek, I've fine-tuned both models. I currently have like the premium subscriptions to all the different chatbots, Grok Gemini Ultra clawed, all these things. And there is like a meaningful difference between the models in terms of what they can do. Each model can. I would say like some models are better at coding, right? So like Claude is amazing at cloak at coding.

Gemini 2.5 has a huge context window and it's also really good at coding. So there's a lot of things that you can do with that that you can't do with with clod. You know, ChatGPT with their O3 model, their thinking model. It's very good at creativity and, you know, brainstorming activities. And so you can kind of use each one of these models as different tasks to accomplish certain

outcomes. You could also kind of hit them against each other, which I've done numerous times to kind of enhance the outcomes and, and also creating different agents. But for Lama specifically, look, I think it comes down to execution, you know, and we've seen what's happened with Meta and the turnover that they've experienced on their AI research team. That's a it's a huge deal, in my opinion, also like that culture of no bad news that's kind of

emerged recently. But I think ultimately the question becomes, you know, what is llama? Is llama better than any other models? And why should you use llama? And the answer is today it's it's not. It's actually falling behind

quite substantially. And when you're thinking about working with these models at scale, trying to do certain tasks and not just, you know, you know, basic summary, right or translation, like, I'm, I'm talking like not, not things that your approach from a scarcity mindset, but from the abundance mindset, like coding and creating, you know, like a 30% or even a 10% efficiency gap is a really big deal.

Think about it like, if you're not, and This is why Google search became, you know, that's a 90% share because it was the best search index. They had the best algorithm, they had the best data over time. That matters a lot because when you're dealing with information, you want access to the best information, the most accurate information, the fastest. And the risk of operating and incorporating bad information or wrong answers into your workflow

is extremely high. And especially with AI, because it's an iterative process of compounds and those risks can become exponential and those costs can grow and be quite significant when scaled out across multiple queries. Million, you know, millions and millions of users on the enterprise, thousands of employees, right? And so I think that's where Llama has kind of started to struggle. And you can see that like they

were initially took over. Like if you look at the developer adoption, which is always like a really, really there's like one I would say like key metric that you can look at to track the success of models is what are the developers using? Yeah. And you know, last year Llama was including myself. I'm not, I wouldn't classify myself as a developer. I'm more of a vibe coder than anything else.

But I did spend a lot of time with it and it was because it was the best at the tool, best model at the time. And now it's kind of an afterthought. And I would rather spend, you know, thousands of dollars on a closed proprietary model like Claude or Jeff, you know, or even 2.5, then spend the same amount of time but $0.00 on Llama because it's just not that good. Yeah, I to your point on perplexity as a rapper, I was looking up as you were describing it.

My man Evan Tyndale set me up on 4 mini in perplexity and now I'm on 4/4 point O Claude, 4 point O Sonnet, right. And I dude, it is crazy what these things can do. Like it is absolutely phenomenal. I don't I'm not a vibe coder. I wish I knew. I need to learn. I mean. You can though, you can't feel. Don't sell yourself short, man.

Like it's like you can do it. You just have to have the you know, the time and the curiosity and it's and honestly like coming at it from the learning mentality kind of. You always usually approach things like I know nothing. Yeah. And sometimes people misconstrue that as me not having conviction, whatever, you know, the reality is that it's important to stay humble, especially with new technologies.

And there's a level of intellectual curiosity, humility that you have to have if you're going to actually take advantage of AI and what it can actually do. Because you kind of, it's kind of like the death of the ego, right? Like and, and people instinctually find it extremely threatening. And how can AI replace me? Just like every single pass

technology before, right? We've seen it with, you know, Blockbuster, we've seen it with, you know, going all the way back to the printing press, you know, and I think that you just have to have an open mind. And, you know, I'm pretty confident, Billy, that you can do it, you know, and if you don't know how to do it, you just ask AI to give you a specific steps on how to accomplish your goal and then just follow those steps and and learn along the way.

Yeah, Well, interestingly, I have been, I've been in a rut lately. Not like a huge rut, but a rut. And I, I've sort of narrowed it down through working through why I'm in a rut is I think that there is a divergent between sort of like my alignment of principles and how I want to live and how I'm currently living.

And one of the, one of the issues that I'm, I, I think has spurred this thought is I read some of these research reports that AI is putting it like, I mean, if I use notebook and I use Gemini deep research, I look at the output and I'm like, I'm dead right? Like like I can't, I'm like, I

can't do much better than this. But but I think that that a more empowering question is OK. I just basically, I mean, I understand that they make mistakes, but interns make mistakes, employees make mistakes I make. That's exactly right. That's exactly how I've been. Yes, 100%. So it's like, OK, well, how now that I have these tools at my disposal for 200 bucks a year, how am I going to make myself more needed or how am I going to use them rather than being fearful of them?

And for a little while I've been kind of fearful. I've been like, what is my role in the world going to be? Because I mean, I don't know my, my hit rate. I mean, I was, I was following the news fairly closely. If I'm, I'm so glad I didn't put out podcasts over the past month because they would have turned out to be so wrong. The interviews that I would have done right. Like, I mean, I, I I would have classically chased some macro stuff. It probably all would have

turned out wrong. And it's just got me like thinking like man, I'm I really got to make sure that I don't just disappear in this world. Thank because the computers are probably better. There's a lot of thoughts in there, and I realize some of it's disjointed, but no. Hey, that's no, I think that's amazing. I think what you just said is spot on. And it's I, I was, you know, preparing for this.

I was taking notes and you kind of just hit on a key part of it, which is, you know, like the and the the psychological stages of this. All right, like, I mean, all tech innovation is you kind of have like you go through these different phases, right? Like, and it's kind of the grief. It's like the mornings. It is the grief stage. Smile it's like AI can't possibly do the type of nuance work that I can do. And then it's the anger phase, right?

When you see other people come on and creating all this content and you're like angry. It's like, this is just hype. You know, real analysis requires human judgement and then, you know, bargaining. Yeah, Yeah, I can help with some basic stuff, the simple stuff. But the complex work, like all that type of analysis, like how to call a quarter or like all this like thoughtful strategic analysis.

I, it still needs me. And then kind of like what you were just describing now, which is like the depression stage, which is like everything I'm good at is becoming worthless. Like, what is my value in this world? And it's this existential crisis that I think everybody is going to go through, Right. And you know, like I said before, I've actually, because I, I've been going through this,

right? I mean, and part of the reason and why I, I, I kind of went on this journey a year and a half ago was because like, I've, I've learned enough and I've studied history enough to know how this all ends, right? And I know how human behavior is. And I know that a lots of people are also not going to take the time to necessarily study history and how history has

worked. I, you know, kind of just basically like immediately jumped to all those and try to just disregard all those feelings that I still have and just move right to the, the final stage, which is acceptance. And that's basically where you just need to find new ways to create value in this world. And I think that with that, there's never be going to be, there's always like there are, there are grifters out there,

right, Lots of them there. You know, hopefully and hopefully with like new technology in this, like that long tail will get cut off. You know, I operate, we're hedge. I work for an independent research firm. I like to think I'm really good at my job. I get my fares, I get stuff wrong just like everybody else, But I think I'm, I'm pretty good at what I do. But I also know that the sell side has has just kind of become increasingly less valuable over

time. And there's a lot of shops out there, but there's also a lot of long tail analysts that just wake up everyday. And they're kind of glorified administrators just working on connecting investors with management, you know, doing NDRS, those types of things. You know, they're writing earnings recaps and preview notes. They're just regurgitating management very, very little, like low value add, not a ton of intellectual rigor and, and integrity in their work.

And like, that's an example where AI, as far as the written research goes, like, you know, for those types of people, they're in a lot of trouble, right? Because AI can come in and through that kind of scarcity mindset, just come in and replace what they are because they're already operating from a scarcity mindset, right? They're just trying to keep what's around them, you know, maintain the status quo. And I think it's those like, you know, those types of people. And I know that sounds

pejorative, but whatever. Well, look for like an earnings recap or or an earnings preview. To your point, I don't see why. I don't see why AI can't write that every bit as well as he can and can. For a lot of it, yeah. I mean, especially if it's, you know, like especially if it's just kind of reading the transcript and going off the

highlights, right. I mean, there's certain nuances related to being able to update the model, build your forward projections, you know, dissect the quarter, look at things critically in the sense of, well, this is what management saying. What's my independent view of it relative to my thesis? Has something changed? You know, management saying XYZ. But I actually think that's, you know, their bias. I actually think it's, you know, ABC, you know, that type of thing.

But again, that kind of comes down to the more like critical thinking components that AI today, you know, still struggles with because the models are still looking for context, right? We don't have, you know, AGI, right? We haven't kind of hit the singularity in that sense where they can truly mimic human thought and critical thinking, because the way that the LLMS work is they're stringing, you know, text together, right? They're identifying patterns and sequential.

They're very good at what they do, right? But at the end, but you know, like, I'll give you like, here's another example, right? So remember, what was what was it? It was it was with Doge, right?

And earlier in the year and there was a bunch of articles that were coming out about how the government was basically paying for a bunch of like New York Times subscriptions and you know, X and everyone on Twitter was like just freaking out about it. And you know, all this big like conspiracy theory because everything's polarized in this world and I'm not right or left, by the way, I'm kind of like just data-driven, like you said, like in the middle.

I I hate all polarization pretty much, but the but you know, I was like, I was looking into it because I was like, well, you know, you can just everyone's looking at USA spending that Gov right. And just like I'm looking at the big increase in spending that was associated with these subscriptions. And I spent a ton of time on USA spending that Gov like looking at government contracts with the

software companies in the past. So, and I know how complex it is, you can't like if you, if you're just looking at it for the first time, there's a lot of traps you can fall into. And so I, I looked at it and I strung together a lot of the different award IDs and contract IDs and it turned out that it was all related to some, you know, library. A lot of it was really to the Library of Congress grants,

right to digitize information. So really nothing nefarious or crazy, but of course, like he can spin it however you want. And then I asked, what did I do? I asked ChatGPT just after web search became available, because the reality is that a year ago, a lot of these models, the amount of access to web information, real information, and I asked it to kind of run the analysis and what it thought. And I gave it specific sources. I was like, Hey, look at

usaspending.gov. Look at all this stuff and you can go through because it's basically an agent. You can see it's thinking in real time, right? The different steps that it's taking. And even though we actually said that it was going to usaspending.gov and it was doing all these things, it actually its conclusion was still biased by an article that it got from auk public publication that was just a very basic editorial that reference the tweet of the person who caused all the.

Perfuffle related to that? Interesting. And they used that as context to anchor the entire analysis, right? And so in that sense, AI kind of failed, right? Because it wasn't objectively analyzing the information. It was kind of this circular reference that, you know, a lot of people like a lot of these models, a lot of these search tools can get caught up in.

And I think the danger, you know, the danger with that right, is in the long term, while there's a lot of healthy amount of skepticism of AII think we are going to be in a place where especially just given the this natural distrust like that people have now in society with like government figureheads, authority, all these things where they're probably going to actually be more willing to trust an AI to give them an answer. Well, that's a problem.

And so it's just kind of goes back to like the same problems that we've always experienced, whether it was social media, but just in a, you know, in a in a different. Yeah. And it's just in a in a completely different lens. And. I've noticed that I'll ask it for like company overviews and then I'll click through the citations and it's like some random guy's blog, which, I mean, there's a lot of good work out there, right? But it's like I don't actually.

Want that random guy's podcast? This guy building? That's where you know this transcript. Like who's this guy? Yeah, that's right. Yeah. I, I need to know a little bit. Let's limit this to like actual sources, right? Yeah. But I mean, look to our point earlier, that's no different than the Internet, and it's no different than a human that comes at an analysis with a bias

coming up with a flawed premise. It's incumbent upon us as users to make sure that we try to mitigate those responses. Yeah, no, absolutely. And, you know, I think this natural distrust too, I mean, and, and this all kind of comes down to like the initial part of like the whole premise of this, this podcast and conversation, right, is around, you know, limiting adoption, right? And just people not necessarily

appreciating what's coming. And I, and I think a lot of it has to do with like the fear of uncertainty, the fear of giving it access to information. Will it train on your data, right? I think, you know, a lot of investment professionals that are probably listening to this, you know, they all work at firms that have really strict, the regulated, they have really strict compliance requirements,

right? And and that limits their ability to like use a lot of these tools the way that they can because they would either have to take it offline. I mean, look, and to be fair, like, you know, I've been pushing this internally a hedge. I in a really big way.

Like earlier in the year, I basically told I had a, you know, meeting and I was like, look like this is existential to what we do. You know, we have to figure this out and we have to be riding the wave and we're small, nimble organization and we have to innovate and we have to embrace this because our larger competitors are going to be on a

lag. And even our clients, like a lot of them through conversations, they are not using AI the way that they as much as they should because their compliance departments court, which of course are risk reverse are telling them that they can't or they're putting shackles on them in in many ways. But that's I guess that's the advantage of, you know, people that are independent, that can stay nimble, that can do all these things.

But at the same time, you know, I think it, it from a research standpoint, I think it's makes it harder for people to actually understand like what the future state of the world is. Because if you're not using these things like day-to-day, you know, I basically been using AI 3 to 4 hours a day for the last year and a half. It's like, I almost like, like they're like, they're, it's great. And they're, it's like an analyst, right?

And yeah. And you have to learn how to use them the right way and understand their their nuances and you know their strengths, their weaknesses, again, just like you would with an intern or an analyst. But the whole information verification thing is it is really critical. But I don't really see it being much different than kind of when the Internet came out, right? You probably remember this as well. I mean, you're a little bit older than me, but we're not

that far apart. Like in the, I remember growing up in elementary school and grade school, you know, when the computer labs first came out and they were teaching how to type and all the stuff, but the Internet was still pretty nascent. But they would still teach you how to run research by going to the library, the Dewey Decimal system, getting the encyclopedia out, doing all these things. And you were told to be very wary of the Internet. Don't trust Wikipedia, You have to verify.

Remember microfiche? That was crazy. What a different time. Yeah, you know, verify this, verify that with actual hard copy. And that was a form of like parallelism where you were doing both for a period of time. And now Fast forward to 2025,

and it's very similar. It's like, OK, use AI for research, check your sources, but instead of using encyclopedia, you're actually using the Internet like Google, you know, to, to, to anchor your, your truth and verify to make sure that the model is not hallucinating. So it's just very interesting how, you know, every generation thinks they've kind of found the real source of truth, but only to inevitably abandon it for something better.

And, you know, we're not really getting closer to the truth. We're just getting, you know, more efficient at feeling confident about uncertainty, right? Like that's ultimately what it comes down to. And you know, kind of one of the things that I've thought about, and I'll pause after this, is just like you look at human history and technology and just behavior. And I was almost, I almost meant to become a philosophy major. Thank God I didn't. But it's super interesting.

But, you know, the only concept in human history is I think, you know, is using the thing we just learned to trust to verify the thing we're learning to trust, right? And it's that constant cycle. And if you understand that it actually, and you study the different phases from a psychological perspective and behavioral perspective, like the world actually becomes a lot simpler.

Decisions become a lot easier to make the higher degree of confidence because it's all pattern recognition, you know? Yeah, we're all individuals and we all make individual decisions and like to think we're special. But at least on kind of a macro level, you can really identify like very similar behavior

patterns. And because we live in the Internet, the age of information and Internet with social media, a lot of those patterns become a lot easier to recognize because they all kind of surface themselves through trending topics. You know, what you see on X, while people will say that's not the real world, that actually is kind of, you know, much closer to the real world than, you know, psych and and people's thoughts and psychology that I think what people actually like to admit.

Yeah, no doubt. So anyway, sorry I just went on like no a philosophical rant, but that's it's, it's all I think it's all really important in the long term. Well, the thing that when you were talking about when the Internet first came out, one of the I, I've been looking at Mary Meeker's recent AI adoption, you know, slide deck or whatever.

And the thing I, I may misquote this, so Fact Check it for yourself, but I think she said it took 12 years for the desktop Internet to get to 50% adoption and the projection is 3 years for AI agents to get to 50% adoption. And like when you think about the difference in the rate of change there, I mean, that is crazy. Now, I don't know. I don't know what the definition

of adoption is, right? I mean, I don't know if that's just people playing around with it or, or like it's truly being used in their, you know, integrated in their life. But yeah, it's super interesting, man. I have not. I mean I I don't want to over exaggerate but I have dropped my my use of Google search.

Now, like I said, I use Gemini, so I have, I'm not off of Google properties, but but the traditional search is like really really more or less gone out of my life unless I'm looking for something specific, which is crazy. Yeah, I mean, I think that's yeah. So, yeah. So I mean, I, it's and I, and just for the record, like I'm, you know, I'm, I'm pretty

bullish on Google long term. And that's a lot of the cut like, and I've just like I've been on the road speaking with investors and it's always like a great, it's always just an awesome experience, right? Because the long as you do this, like you just have really like thought provoking conversations with folks and you know, these are people managing billions of dollars of capital. And you know, this has been a

huge topic conversation. I guess like embedded in your comment is a question about Google's business model, the sustainability of search. What does it look like what happens from margin perspective, all the all these other things, right? Yeah. I think again to kind of look at this from a historical perspective and framing this between a, you know, like framing it between like that using that scarcity and abundance framework, right? Is that from a scarcity

perspective? You look at what you just described and you extrapolate that to Google search as a whole and you say you would come to the conclusion that Google's a 0 or you know, like search. Is search. Search is definitely going to have it. And I want to say Google is 0 for those of you listening. It's like a thing I do on Twitter. It's related to Fuba. I just like it's, it's just funny. I think it's funny.

But anyway, sorry, but the I don't obviously think Google is a 0. Even if search goes to 0, Google is not a zero. But the point is it's worth something. I'm sorry, I just got myself totally. No, you know what I think is interesting because I think about this, where I go with it is I look at the value that Gemini creates and I and I think, OK, well theoretically what could they charge me for this? And I and I think subscription revenue has a massive upside from this.

Totally. There's so I guess at its core. I don't own it for the record, but I just, you know, I spend time thinking about it and I, and I and I look at my search usage and then I open up the financials and I look at the growth in search and other and I'm like, how in the heck does what I think the world looks like? Like this is one of those Adam Robinson's things that make no sense. He he likes to say that you can make a lot of money and things that are very obvious and things

that make no sense. Google's search results make no sense to me given what I think is going on, but that probably means that I don't have the right model of the world, right? It turns out. It turns out that I am not the world, which is sad, but. True. No, but it's true. I mean, look, I think that's a very important point. There's everything. There's always a Wall Street bias, right? You know, if you go and ask, you know, 90% of people, and we did

a, a survey on this too. You know, most people are still using Google search and they still plan on using Google search, right? Your use case, my use case, really, it's a search, especially for informational queries and research projects, is much more advanced and nuanced than the use case for a lot of people. If you just, if you don't believe me on that, just go to, I'm not saying you, I'm just saying I'm listening. Like go to Google's trending results page, like look at the

most trending topics. Like, it's not anything that you would probably be, you know, I'm sure it's not Bill Brewster's. Google. Search history is coming up as most trending, let's just put it that way. Or or mine either. I think that if you look at it from the abundance mindset, right? And look at through history, it kind of comes down to a company's willingness to disrupt themselves, right?

And looking at their core technology and what they have in terms of assets to help them make that transition. So, you know, tying it back to the innovators dilemma, it's like, OK, you know, even if you do have the assets, what's your actually willingness to to disrupt yourself? And if that's the case, at what speed are you willing to do that? And what does it look like on

the other side? You know, in the case of looking at things like, you know, when the when the mobile phone, when the when cell phones first came out, right, it was a yeah, when cell phones came out, I was like, OK, this is going to be a replacement for the landline, right? Fast forward 1520 years and it's like, OK, this is actually, you know, an entire computing platform at on your hand, like

in your hand, right? And then and then the ecosystem develops off that through the App Store that creates billions and billions and billions of dollars of worth of value. Where if you just looked at it from when the phone first came out and you just looked at it from a OK, this is going to replace landlines. Like, yeah, like landlines ended up getting like going away, sure, but it ended up being. Just more. You can go back to look at Kodak, right, is another example.

Like they actually kind of they invented like a digital camera and then they were like, OK, well, we have to preserve film. So they buried it, right, because they didn't want to self their self themselves. So there's just so many examples of that in history. I can keep going on. But in the case of Google, I think that look, search as we know knew it is gone like that is I agree with that 100%. Now Google is disrupting itself currently for the use of AI

overuse rolling that out, right? They're they're currently going to the run phase and now they're also rolling out AI mode. So the all the reason why the only reason why Bill that every you and everybody else and myself initially was so attracted to ChatGPT and all these other bots that are giving all this information is because they provided users with a new discovery factor, right to that replaced search.

And otherwise, if you look at Google, given their billions of users, if they had just disrupted themselves sooner, right? It would have made it so much more difficult for an open AI and ChatGPT to actually come into the market. Because instead, when ChatGPT came out and it would have been like, oh, what is this like, Oh yeah, this is just Google. Google already did this. Now the question becomes, is it too late for Google?

And I think a year ago that was in my mind, especially in the beginning of 2024 and a year ago, it was more of a possibility, right? Like I had serious questions about management's ability to execute Gemini in terms of the, you know, 1/5. It wasn't in a good model. I wasn't using Gemini. I tried using Gemini.

So I've, you know, been doing all this coding and, and, and building applications and all these things that I naturally just came back to Claude because I tried everything else and it was giving me bad code and it wasn't working and it was breaking. And I was just like, well, This is why, like, why am I ever going to use this if it's giving

me bad results? And so I went and went back to Anthropic, which owns Claude. And then that shifted earlier in the year when 2.5 came out and everything that I thought about Google and their infrastructure advantage related to a six TP use, having all the data, the tools to be able to give like have the best model out there were 2.5 pro. That was a huge unlock for them. Like massive unlock. You cannot be bare. You cannot be bullish on on Google. If they were not, they didn't

have the best models. Now they're leading the frontier in terms of. Lowest token cost, highest quality and as you know and I know you're like. That then you get the developers using it, yeah. Absolutely. And by the way, like you know how this works too, because you studied like network effects and scale networks and and even kind of like can come back down to a cable like or like a fiber like I'm not. Why are you trying to bring up like old PTSD man?

No, you know what I mean, Like or, or or DSL like the whole, the whole like, I think it's actually like a really fascinating parallel because you know, the transition from DSL to cable to fiber, right? The whole premise of that is around new technologies coming in and being able to provide the fastest service at the lowest cost. Yeah. And there's really no room for any other alternative. Like if you can't, if you don't have that structural cost advantage, right, then you're

always going to be behind. So in the case of, you know, fiber versus cable, they can, cable can do whatever they want with DOCSIS, right? But the reality is that because fiber is just structurally faster and it's a soft and it's largely software upgrade. As soon as cable gets to 2 gig symmetrical, they can come out and you know, go to 10 gig symmetrical, 15 gig symmetrical because the marginal cost right to deliver that extra bit for that is like 0.

Yeah, right. And while the marginal cost to deliver like the extra level of compute or inference capacity for an LLM isn't 0. But if you're going to actually develop something to and, and make this a mass market product at scale and incorporate your AI across billions of users to drive adoption like Google has across all their applications, you better have a massive cost advantage. And they have that with TPU. So the inference capacity,

right? And, and inference is basically the every single time you run a query, right? It fires off it's tokens, text get translated into series of tokens. And then once you submit the query, the GPU fires and it runs the compute and gives you back the results, right? It's like the tokenizing like I when I was hosting Llama or whenever, whenever you do it, like you hear the GPU fire, it's like it's like an engine like it's just like. Interesting.

Yeah, it's really cool. And then it goes up and it goes down so you can hear it actually happening. It is pretty cool. But but you know, like, like the point is like you need for Google, you need to have that, that scale advantage is huge, right? Because that's how you can, they can disrupt themselves. And while there's probably going to be margin implications in the short term, that's how you get AI adoption to scale.

And so now what you're, if you use Google, you have AI mode, like they're embedding AI mode, which is based, you know, kind of deep search across the board right now, deep research if you want to use that, which is really cool by the way, you know. Yeah, deep research is awesome, so. Cool, like I and but you need a paid subscription for that, right? Yes, and I would pay more for the record. Yeah. And it's why is that? Because it's expensive now. Yeah, to to run it up.

Yeah, that makes sense. And now Google can just is incorporate can incorporate AI mode with deep search. And because they have by the way, like the best search index and all the history in the history of the like history of the world, they have the best data, right? They can they already know what users are looking for. So they can pre train a lot of the models they can index, create multiple indexes to better surface information

faster. And then because they have that structural cost advantage with TP US and ASICS, which are like probably anywhere from the research says they're like 5 to almost, I've seen as low as 5 to 100 times more efficient than the GPUs, right, when it comes to inference. And even if it's like not 100 times, but it's 10 times or if it's five times, it's still huge. It's still a huge cost, cost advantage that they have.

Then it's, it's they can embed this technology across all their applications, drive usage and do things that the competitors simply can't for a much lower cost.

And so my kind of hypothesis is that it's not too late for Google that they are going to roll all this AI out across every single part of the surface, try further adoption of it. And while, and that in a few years from now, and hopefully sooner, if we're having a follow up conversation, Bill, you're saying that you cancelled your perplexity subscription, some of these other things because Google, you know, just made all these things available through it's, you know, a on mode interface.

And that there is really no longer traditional search as you know it. Instead, it's going to be, you know, a chat bot, right? But it's going to be free and they're going to have the best information. And I think that's, that's my bet. That's what I think is going to happen in the long run. Now obviously in the short term, you know, 6 to 9 months, there's going to be some disruption related to that.

But in the long term, they had every single asset for them to succeed and they've been shipping product like crazies. Anyway, that's that's kind of my quick or not so quick thoughts on it. No, that's interesting. And the other, the other thing that you got me thinking about is I remember talking to people about AWS when it was starting and someone much smarter than me said, you're not thinking enough about the, the cost going down and, and what that's going to

look like in five years. And to the extent that, that the costs decline like every other technological revolution has, if they, if they do keep me paying whatever I'm paying, they don't, they don't need to worry about necessarily raising the price that I'm paying, right? They could just squeeze margin out of efficiency gains, which is maybe what I should be thinking more along the lines of. Yeah and yeah. No, absolutely. I'll tell you who would scare me if I was an exec of a

traditional media company. You know, I think. The amount of content that anyone can make now is. Yeah, it is wild. I don't know what's more scary to be as a exec at a traditional media company because, I mean, that's just been scary forever. Yeah, that's true.

It's just been awful. And actually I can make an argument, I could make an argument that AI actually might be better for them, like could help to them if, if there is that structural, if it lowers the cost to producing content and they can produce more content faster while maintaining quality or improving quality. And they have all the IP that they already know is pretty bankable, right? In terms of what people care about.

Then in theory, you could actually maybe see kind of that value capture shift maybe back more directly towards the studios. Yeah. Where historically in the last five years, the the value ship capture has been going more towards like talent and the consumer. But but you're right, like the proliferation of content is going to just be massive. And so competition for eyeballs is only going to go up.

And on the flip side to your, I think to your, to your, you know, the point you're you're alluding to before is that, you know, if I can create cinematic quality or like very high production value content at very low costs, then the barriers to entry get reduced. Then how can Paramount or Warner Brothers and all these studios like compete effectively? And that's, that's, that's also true. I mean, I mean, all of it's bullish for YouTube, right?

I mean, I think. That's well, that that's kind of where I get with it, right? That and and Instagram right. To the extent that people are looking at reels. I just I don't know. It's dude. It's an interesting world, man. When one thing that I do worry about is the amount of power that this is all requires. It sort of seems to me that we've thrown the planet on the back burner in this race. But but maybe I'm wrong. I hope I'm wrong. Yeah.

I mean, there's been our energy analyst has done a lot of work on this and a lot of the analysis out there says that, you know, we just don't have the amount of, you know, at the the grid just can't handle it, right. And so that's why we have like Meta and Google building nuclear power plants, which is also terrifying. Like, like we're going to have pictures of like Mark Zuckerberg sitting there, like growing out in front of like, you know,

nuclear power plant. I don't know, it's, it's going to be it's going to be interesting. So, yeah, I mean there's, there's all these implications and and that's also why like for Meta, you know, they're they're front loading a lot of the CapEx spend related to data center build outs, right. And the infrastructure piece, because they know that the lead time to get a data center up is like five years, five to seven years permitting, getting the

power set up, all these things. It's like they can they can get GP us a lot faster then they can actually build out the data centers and NVIDIA won't deliver GP us to their customers. They don't have the ability to receive them. And if you don't have the data center, that makes sense, you don't have the ability to receive them. So that's kind of I think a big part of where we are on that front. It's it's just crazy.

I mean, I don't know what's were like to your point before, like I don't know what's scarier, who should be more scared like a traditional media company or a a traditional software? Traditional software is something, right?

We're like a SAS company, like a, a large incumbent, like a sales force, if you think about it, because the whole barriers to entry, barriers to development, lower cost argument is, is even more true, I think on the software side, because with video content generation, even with Google's VO3, you can still only like stitch together 8 second clips, right? And it's actually way more expensive to produce that type of content than it's to reduce the lines of code.

And so for example, like I've been building software applications just as a hobby, nights and weekends, just part of the learning process. Part of it's just, you know, like my family has a small vacation rental business out on The Cave, right? And you know, one of the huge issues we've been always had is around creating kind of a customer database, right, to track everything, but also a booking system. That is what we need, right?

We don't need anything crazy. It can be simple because we only have like, you know, it's like 3 properties. So it's not like it's massive. So basically, you know, coded out like our own little Airbnb app, right? And it's all hosted front end develops. It's a next JS application. It's all back ended on Postgres with the database and you know, using Stripe for payment

integration. We actually, I also like developed a rag off of it. So you can, we were able to take all of like all the paper and my dad's old school, right? He's like, like he was still using AOL press and like from late 90s to do his website. Like they deprecated that thing 15 years ago. Like I, I think he's like the last customer on it. And, but we basically digitized everything, put it was able to using AI, was able to get it

into a database. And then we created a, a, a rag with a vector database, which basically allows us to run natural language queries against all the data, which is like what it's like, OK, we have this availability for the third week in June. What rate should we charge or who should we? Can you give us a list of people that we should reach out to that are the most likely to, you know, book that week? Interesting.

And it goes back through 30 years of information and returns us a source document with going back to the actual data, right, that we have in the database of who we should reach out to and why. And it's, you know, like that took that was probably took me like 50 hours of like just really. Yeah. And. That's pretty quick.

Yeah, it actually is right. Like, yeah, for like, and that's me. Like not having and then by in Full disclosure, like I didn't know how to code like up until like Python. Any I know how to do any of this. Like, I mean, I was always even technically savvy, but and, and curious, but I'd actually like up until like the very beginning part of 2024, I had literally just given up on coding. Like it's something I've always wanted to do. I everyone's like, do you do all

this stuff with data? Like you got to learn Python, right? And I was just like, look, I'm, I'm not old by any means, but I got two kids. I'm kind of like in the middle enter like that middle phase of my career. Like I don't have the like, it takes too much time to like, you know, to actually learn how to code, but not just the basics, like to learn how to code to actually be good enough to actually have it have a positive incremental impact from where I'm at.

So I kind of just like gave up on it. I was like, yeah, I just accepted the fact that I'm never going to be able to do this. And then AI came along and holy crap, it's just been life changing. And so you know, yeah. So like 50 hours. And is it, could it be, is it is what I built like technically like production ready, secure and scalable to like, you know, thousands and millions of people?

Of course not, right. Like you still need, like if that were ever going to be the case, you need a senior engineer, you need to do all those things. But you know, for our purpose, it, it works just fine. I think, you know, the next step of it is we're working on creating a true agent, not just a workflow, but like an actual agent that can help source bookings and actually book for us. So like we don't.

So, you know, they don't have to necessarily call and speak to somebody, I mean, me or my sister or my dad, we can just, you know, they can just use an agent and they can handle a lot of the complexities associated with the booking process. And it can just be more of a stand alone system. And it can, it can handle all of the manual processes like getting, you know, leases signed, coordinating with cleaners to have them come in, all these things.

And so, you know, it, it, it's just, it's amazing in terms of like what a lot of this stuff can do. And yeah, I think the biggest. And so, so, so OK, so why did I say all that and said it because I think if you're, you know, an enterprise software company, right? Yeah, like, you know, like like a lot of these things are enterprise grade yet, but the lower cost to entry to develop

software is a really big deal. And it's going to matter over time because you take something like a monday.com, which is a task at its core is a basic task management solution. I could, I could just I could code out something with probably 80% functionality of that in a weekend. But literally now I can't get it to scale because, you know, because I don't have the sales team and all this stuff, but the

concept is kind of the same. So on the one hand, the barriers to scale are still there just like everything else, right? But on the other hand, there will be a next wave of billion dollar companies, multi billion dollar companies that have been built using AI, you know, 110th of the budget that a lot of these predecessor software companies have.

And so just like we've seen with disruption before, right, scale, lower cost advantage, they're going to translate that structurally lower cost advantage into lower pricing because they can and the existing cut, you know, the incumbents are not going to be able to compete on that, right? They're not a they're not going to be flexible enough, most likely to self disrupt and move faster and beat the unit

economics. You know, like basically they're, you know, the cost of capital is going to be so much higher for them. I mean, plus capital is not the right word, right? Just because they, you know, their, their break even is going to be so much lower for some of these newer companies that they can afford to price at like 3040% that. So it's going to be like, well, look, if I can get 80% of the functionality for 50% less, that's a pretty good deal, right?

And I, and I think that I think that's like the big part with software companies and like, I think about like a sales force and, and specifically and been thinking a lot about that with like ServiceNow, especially because AI is going to disintermediate a lot of software companies as we know it. And this whole concept of system of record where you want to be the system of record, because that gave you pricing power, it would give you stickiness that

is going to erode massively in, in AI and you know, the future because you're going to have interoperability. And the premise of being a system of record is that you kind of aren't interoperable. You are you have to be the platform and you don't talk well with a lot of other software companies. And so you need to buy the module from the current platform that you're on.

And with AI agents, interoperability, MCP servers, all these things, the entire thing is breaking open where really they just kind of become glorified back end databases, right? And whoever can, you know, handle the front end AI orchestration is going to really win today. And I think, you know, that's the service now thesis. And I think that's actually the right thesis. So it's going to be really,

really fascinating. And, and my, and I will make one last point on Google is, is that, you know, Google with all their assets in GCP, we're going to get to a place where what I just described that I built, like I built that through individual prompts using Claude. And it was very, you know, I want to say inefficient, but it was just told me what to do. I downloaded all the modules. It was all command line

interface. But eventually, and we're seeing this now with kind of like the rise of like lovable replit all these smaller platforms, like with Google, you're going to be able to, I'll be able to at one point in the future to say, like, look, here's my small business, right? We have no technology. This is what I want to do, right? It will prototype it out and then it will build that application for me and it will build the front end and it will build the back end like a full

stack engineer. And it's all going to be integrated and built on Google's cloud platform. And if you really think about that, like the implications of

that, you know, that's huge. And because they have Gemini and it's fully integrated, you know, across the entire staff, they have owners economics, they have TP us, then you're really looking at a situation where Google is going to be in a great spot where the 1st place where any entrepreneur or any company that's going to want to build a new application is actually

going to be on Google, right? Because Microsoft with Azure don't actually own their, you know, open AI while they have a stake in it, it's also a risk, right? And then Amazon, they have a stake in Anthropic, but they also don't really have strong models as well. So as far as all the hyper scalers go, Google's in like a really unique space. And I would say that that value

long term is enormous, right? Like just to basically be like all the, you know, just to be able to go to a company and just build a application. It's not. And I know it sounds like it sounds simple, but it's, it's the vision longer term, like that's what this is going to turn into. And that's why I think a huge part of this too is like people are over focusing on search as AI. You're not looking enough on, you know, the GCP component and what this can mean long term.

So. Reminds me of when people were focused on video and not broadband margins. A little. A little, yeah. A little not. Not totally, but a little, yeah. That's wild, man. I, you know, the we, you would have mentioned something about your small business.

You want to create an agent. The other slide that like pops out in my head from that Mary Meeker deck is I think they said 73% of people can't tell the difference when they're talking to an AI agent versus a human, like in customer service. That's crazy. Yeah, it's only going to get even. That numbers going to go even higher. Yeah, Akram, he said this a long time ago, but I think he, I think it was, it was prescient

and correct. But like he was like, man, I could see a firm like American Express really benefiting from AI. That was like one of his early things that he said. And I was like, yeah, that does kind of make a. Lot of sense.

What was the rationale on that? I, I mean, I don't want to speak for them, but when I think about it, the, the things that make sense to me is to the extent that you're analyzing something like credit risk and different data signals, I could see AI possibly helping early warning systems on when to maybe reach

out to people. And to the extent that you're dealing with customer service, I could see, you know, I mean, look, if 73% of people can't tell the difference right now, you probably need a lot fewer customer service people. Yeah. But. No, it's he would need to. Yeah, well, I mean, let's let's think about it in terms of our jobs, right?

Like, so I, so I kind of, I say this jokingly, but I'm kind of serious about it or become more serious about it is like, you know, like I I'm actively trying to replace myself with AI. Like it's a it's a challenge. And a lot of people say to kind of like they don't want it, no resistance to that, right? Because it's like, how could you possibly do that? I think it's like an inevitability, right?

Yeah, in many ways. And I think that it's not ultimately like, again, like I don't think we should think about this or people should necessarily think about this as being replaced. I think it's like augmented, right? Like you're using AI to make yourself better, faster, cheaper, right? And when I say cheaper, I mean like to be able to do like lower your own, like to be able to do things at a lower cost, to do

more of it, right? And therefore that's a huge competitive advantage if done properly. And so you're enhancing your own productivity. You're building things that you couldn't do previously or that you were never able to do before. You know, I'm doing, you know, my own workflow, you know, analysis, like for, for example, the FCC has a bunch of great data on the cable space, right? So those are very, very large files. And you know, in a world of Excel, like you can't analyze them.

But in a world of AI where you understand how databases work and you understand how to ask questions, which I think is actually probably the most important thing out of all of this is just, you know, if you're if you're not good at asking questions like you're not going to be good at. Yeah, right. Comp engineering is real. Just like you could actually be a really good Googler and people would laugh at you when you said that, but it's actually true.

When you ask it, will you ask the the model that you're working with if there's a better way to ask the question? Like will you? Will you use it to help you reason through things or do you have to learn how to prompt it? Yeah. So I So what I'll do is I will.

So usually what will happen, or This is why I started doing more recently, is you can create a project in any of these apps, but really like I use thought as my go to and you create a project and you can unsert project knowledge and think about this is like a knowledge graph. Because what ends up happening is that a lot of these AIS, they have context limit window limits. You hit your limitations for the chats.

And so the project knowledge base is actually really helpful way to kind of get around some of that. And you give it like core information data. So maybe you would, in this case, I would upload, you know, a sample file of like 1000 different rows of data, 15 columns, right? And and then at the same time, I would upload any type of developer docs or like the schema definitions that come from the data provider.

And I put that in the project knowledge space and then I would create a chat and I would ask it, I would give it my use case, be specific, what I'm trying to achieve, what my outcome is. And I'll prompt it and saying like you are an expert, you know, you are a data engineer, right? You are a data engineer, but you've also, you know, helped build the FCC's, you know, national Broadband System and you have first hand experience working at charter comp, you

know, like something like that. And, and all those prompts actually matter. And I'll ask it for a prompt than what I'm trying to achieve. It will give me a prompt. I'll review the prompt, I'll other the prompt, and then I will take that prompt and I'll give it to probably two other a is and ask it to tell me what it thinks about the prompt. Interesting it will. Give me other edits and insights that I wouldn't have thought previously.

Almost like a peer review is kind of the way to think about it. And then once I have that prompt and I feel like I haven't missed anything, I will take it, put it into the instructions section of the of the project chat. And then I'll just start asking for step by step instructions on how to do things. And I think the way, the best way to do it is like, you can't, like AI doesn't do one shots very well. AI is not just going to be give you the answer right away, but you have to chunk it.

You have to think in steps and sometimes very small steps. And so if you retrain your thinking and expectations, almost like you're putting together like, you know, like I, I've put together tons of like research process notes because we had analysts that come in and I'm like, Hey, I need you to update this. And I've done a million times, but it's nuanced, right? You got to go to the government website, put this in Excel, update this chart to all these

things. So I've written like very detailed step by And you know, there's some things there's always some first hand knowledge of like little quirks that you have to deal with, but it's very similar in the sense of AI. And then it gives you steps and then you just follow those steps. And then if you have a problem with something or you get stuck, you know, you could actually take screenshots of your, of your page, right?

Or in the case of like AI Dev studio from Google, they have this really cool functionality where you can do screen share. So literally it's you're sitting there with a Gemini open screen sharing and it can see everything you're doing. Now again, compliance like this is not stuff I'm doing. You know, this is all this is stuff I'm doing as like projects, just my own. But you can use these tool. Look at, you know, it's basically like, see what you're

doing. And then that case they're like, OK, I see on your screen you're trying to do this. I see this code and it can give you insights as if you're sitting there with a senior software developer, an engineer, like over your shoulder. And that's something that I would have to do. Previously we've had the higher talent to do that. I would have to get on their schedule. We'd only be able to do like one hour sprints. It would not be very efficient probably. Then the meeting ends.

You hope you both remember what you talked about. You. Get back together, you redo it. It's, yeah. And it's just like, you know, and it's like, and then and it's also like a lot of this stuff is just like, I need to sleep more. So it's like I'm up. I have an idea. I'm doing it like 1:00 AM. It's like not calling a software engineer do that. Like, so it's always on that kind of a thing. I mean, the biggest issue is just burning through credits and

all these things, right? That's why I have like, that's why I'm like spending $1000 a month on like all these different subscriptions because I'm literally just like burning through tokens even after becoming more efficient. Like I'm hitting these limitations like very, very quickly. But but yeah, that's, that's kind of like, I mean, and so, OK, so sorry, my point before, I

just keep going off. But point before is like with the some of this data set now it's like, OK, I couldn't do this in Excel. So, you know, let me have the data in there, put it into a let me get into a database. Let's write a Python script on it, right. And it's like, wow, this is like took took me, you know, 4 hours where before it would have taken me like, you know, 20 to 30 hours working with the software engineer.

Yeah, it's amazing. It's really just amazing what it can do as long as you're kind of, you know, willing to put in the effort and, again, understand what it is and what it isn't. Like it's not omnipotent. It's not going to give you the right answer right away. It's going to make mistakes. Just like to your point earlier in the conversation, Bill, it's like an intern, an analyst. People are going to make mistakes. The question is, is the margin of error better or worse?

Yeah, Yeah, that's right. Do you think so? I had a guy that was was at my house. It was one of the first times I've actually gotten to just sit through a software implementation. And he was working with a large firm and data quality really mattered, right. And like they were going through all of the checks. It was all weekend long. They were just finding errors and then fixing errors. And it was three of them on the phone.

And I thought to myself, I was like, why in the future, why can't a computer just like, do this and run through all the iterations and find the errors and fix itself? Now I understand you've got it, you've got to train it and you've got to know the answers or you've got to know the questions to ask and whatnot. But man, like, to the extent that you start reducing that kind of friction of a, my first thought was, boy, that's a sticky relationship.

And my second thought was, I wonder if it'll ever get less sticky over time. Yeah, right. If computers can kind of. Like.

Troubleshoot all that. So I mean, coding has been like the killer app so far with AI. But if you look at what Google just announced with their jewels coding agent open AI release something to forget the what their what their forget the name of what they called it. But basically you can take your entire code base and you can have an agent, a coding agent that's does just that and it can scan through up and down back, you know, up, down, sideways, your entire code base and look

for errors and they can change things in real time. And that's so that that part of it is going to get changed, is going to change too. So you're going to need as far as like engineers go, you're going to need a lot fewer to do things, but you're still going to need the top tier, right, Because you're still going to need some type of human

validation, human verification. But that's where the that's actually probably going to be like an entirely like, you know how there's like ad verification that will verify like all these things to see if an ad is fraudulent, if there's a watermark, all these things, it's going to be some right. You're going to have entire companies whose, you know, focus is on creating like a verification system for AI to make sure that it's doing what it's supposed to be doing and

that it's it's accurate. So and then that kind of goes into the whole quality component of software. So that's yeah, I mean, it's, it's, it's pretty remarkable. And the speed at which you can develop software and fix code bases is, is really, it's really something else Like it's, you know, who's going to who's going to be writing lines of code. You know, it's just not going to happen. You're going to be tweaking code.

And and then also for our job, man, like have you like the way I the way I think it's kind of going to play out right is I think that you're basically going to have like model curators or model builders. So every single hedge fund by side. So like eventually you're going to need somebody on the team whose sole job is to basically fine tune models and build agents that specifically for, you know, individual teams or asset managers, processes and

work flows. You're going to basically try to, and then the agents are going to be competing with each other just like you would hire someone for a specific role for an analyst role. But you're going to be doing that using AI and you're going to be training it on your own data, but you're also going to be probably have, you know, security specific agents or security specific models, right?

So that's the other thing when I said before, like I'm trying to replace myself with AI, it's because I realized that I think, you know, like on the South side today, you're like, you have people that are like the axe right? In certain, in certain, certain sectors. And it's like, I, OK, I go to this person for this task. I go to that person for that task.

And, you know, and I, I think that for better or worse for me, right, like I'm get like Roku's already on my gravestone, like better or worse, like that's, that's what's happening. But I, I do think I know that company better than probably anyone else does. Again, for better for worse. And good models spend a ton of time on it. So people come to me, right, and they pay me because they want to know more about the company and they want to know what we think

about certain things. Well, I can create a model and fine tune an LLM that's basically a reflection of all my knowledge and data and insights on a specific company, not just a rag, which is like databasing, querying for real time information and, and, you know, retrieving and providing like a text, a text response with sources, like actually train a model to think like me and with its knowledge.

And because if you think of like the way and, and there's more that goes into it, but like the way that you fine tune a model is just the Q&A. So you, you give it a bunch of Q&A, like thousands and thousands of examples and it just, well, and in the model, you know, depending on how good the model is and the weights you give, it will train. And so then in theory, you could have like Andrew Friedman's Roku, lol, or Andrew Friedman's Roku agent, right?

And that's infinitely scalable to the extent that if you want information from it then and you know, you can just chat with it whenever you want. You pay a certain amount for that, right? But I trained it, right. It's so therefore it's unique. And to my point before about like the efficiencies of certain models, like compounding inefficiencies, like you're probably going to want to use that model instead of somebody else's model on the street, right? Because it was trained by me.

That's how you can extend your own knowledge and your own expertise right in, in, in an AI gentic type of world and, or, and that's also how you can easily be replaced, right? Because then all of a sudden, you know, you basically are breaking down like time, like like their time doesn't become a friction. It's like when, you know, it's like when content, when Netflix came in, like video went just

streaming, right? There's no more prime time five to seven, like you're no longer bounded by my schedule, right? You want to access information. And so you are reducing that friction. You're increasing information liquidity, you're getting to scale faster. And then I can monetize that and that's why. And that's kind of a version of the future.

So I think you're going to have like, you know, someone who specialized in AI that understands prompt engineering, that can actually train models, but we're not there today yet because training is expensive and it's hard to do or sorry, fine tuning is expensive and it's hard to do.

Eventually, I think like the next, it's either going to be a hyper scale or somebody's going to come in and just make it so much easier for somebody to fine tune a model based on their own expertise and their own knowledge base and scale that out that you're going to have like, you know, and set like there's marketplaces that emerge for agents.

There's already that case, but it's going to be more like open, right, where you could have the Bill Brewster, you know, business group podcast, LLF, right? And it's basically proprietary. It's fine-tuned on every single transcript, every single question that you've ever asked anybody, right? And you can use that two fold. You could use that one either for internal purposes. So when you're preparing for the next podcast, right?

You can ask it for a series of questions and you can give it information on your guest, right? And you can do deep research on a topic, but it's also going to know your style, how you think, types of questions that you ask, because all even in the world of AI, authenticity is going to be even more important, right? And so. Yeah, although people won't know if it's actually me. If it gets too good, they'll just they'll just have videos.

Yeah, that's totally true. But I mean, it could help you like, you know, in that sense become more efficient, ask better questions faster. And then, you know, on the other hand, too, it could be a form of a product, right, where it's like this knowledge base, you can ask, ask questions to it and you know, it will give you answers.

And then, and actually you could use a rag for this today if you really wanted to. But but I think like the future is going to be like fine tune models because that's just otherwise everything just becomes a wrapper, right? And everyone just has a different UXUI on the same model. And I think that, you know, while I think everyone's going to want to use the same model to fine tune, but everyone's going to use the best model to fine

tune, right? Because there's a cost implications and quality implications. But I think the fine tune element is something that a lot of people aren't talking about today, which I think to me at least seems like an inevitability. Because you're also, what that also does is it breaks down the

problem of context. Because right now, you know, with like Claude, like you can only have like 100,200 thousand tokens, which means that just like a conversation over time, the longest conversation goes kind of either maybe you get tired, right? So your energy levels drop over the course of the podcast or conversation, or you've kind of forget what you're initially talking about. You go off topic, right? And or you can't recall what

you're initially talking about. The models work the same way in the form of context windows. And, and that's why there's limits on the chats, right? And that's why, you know, yeah, that's why there's limits on the length of the chats that you hit a lot of time. That's also why, like, I like to use sometimes, like multiple models because if I get stuck on one, then I get the Gemini, right? And then Gemini's like, looks at it differently with the different model, with different context.

And it can sometimes solve problems that Claude got stuck on vice versa. And I think that's, you know, going to be a really big deal. We're not there today, but that's like, I think it would be a pretty cool product, right, for like media for, you know, information in general and accessing it. It's going to be really transformative to a lot of people's experiences. So going back to a previous question, what, what do we think

this does to employment? Are we are we going to 20% unemployment or do we create enough new work and efficiency? I think that my perspective is really hard because I everyone internalizes it through their own lens.

And I've come to realize, and I know I'm not trying to sound arrogant at this and it's actually probably A and, and this, but I've just come to realize that like not everyone kind of thinks like me in terms of like the ability to self disrupt and like find new technologies and you know, approach things from like this like total abundance mindset.

So I try to. And the reason why I'm saying that is because like if, if, if I, if my base case was that everyone was like that right, then probably, yeah, like they're probably be like 30% unemployment. Yeah, Yeah. Well, everybody would be fine disintermediating their own job, right? Yeah, exactly. Just, I mean, productivity would go up massively, right? So there's that offset from an

economic growth perspective. Yeah. Like, you know, everyone would be figuring out how to do like 3-4 jobs and using AI agents to replace themselves. So for that reason though, I think like it's going to be a slower rate of adoption. But I do think that the core of it's still the same in the sense that a lot of jobs today will go wet just like they've gone away in past technology cycles.

Now. I think what makes this a little bit different is that, and you, and you called this out too, it's like the speed at by which everything's happening is a lot different. Oh, as technology moves faster and, and it disrupts faster, you still have the same impact on unemployment, right? In the sense of companies deciding that they can automate things and they're trying to drive margins in the slower growth environment and they're

getting efficiency gains. And therefore, instead of 10 engineers, they only need 2 engineers and they're going to pay, you know, those two engineers may be a little bit more, but maybe they keep 5, right, and then pay all 5 less. Like there's going to be kind of a deflationary shock that comes from all this. In the long run, it's good for corporate margins. But what I think the problem is that since the technology is moving so quickly, the ability

to retrain people is going. I don't think that is necessarily going to speed up as quickly, right as the rate of AI adoption and the impact on business models and layoffs happened. So for that sense, you could have kind of have like an asset liability mismatch. Yeah, maybe one way to think about it. Yeah. Whereas if someone was like. Call it a societal asset liability mismatch. Exactly, yeah. And you know, the risk is that if you do too much of that, do you just risk?

Man, the politics could get crazy. Yeah, crazier or unions like, I mean, that's why it's like, you know, it's like the concept of like AI, right? It's isn't like it's interesting, but it's also not crazy, right? Yeah. You know, this idea that you're going to have, you know, way more regulation around where AI can be embedded and what it can't be and who can actually make decisions.

It's kind of not, It's, it's like, I mean the unions with this, one of the things I've learned researching like the port strike that happened more recently is that a lot of the a lot of those crane operators that are taking the container ships from the boats as part of their contracts, They actually have to like a robot can't do basically everything, like take it off whole thing can be automated, but someone has to be

in the seat. And then the person operating the crane has to actually, per terms of the union contract, annually lower the lower the container for the last 10 feet. So maybe that's like something similar. I mean, you can also like, really go dystopian and think about it in terms of like, does this open up the potential for like, universal basic income? To the.

Extent that, you know, the economy and all these functions are going to be able to be operated using AI, but you know, and, and it's going to be so accretive to corporate profit margins, but at some point lose the demand function, right? The demand works against you because AI doesn't consume in a

similar way. And so in order to sustain levels of demand, you need to have some type of form of like universal basic income, which probably comes in the form of like a higher tax on corporate profits because the share of corporate profits as percentage GDP goes so high because productivity is through the roof, right.

But even though labor demands down and so eventually it kind of comes full circle and you have to close the loop and basically give people money so they can go out and maintain their lifestyle and also buy things to keep everything going. Again, kind of dystopian, but also not crazy. But there will be like a Charlie and the Chocolate Factory, the Johnny Depp version part of this too, where Charlie's dad's working at the Chocolate Factory.

I'm sorry, working at a factory, putting squirt on the tops of toothpaste, right? And then all of a sudden machine comes in to replace him, he gets laid off and then at the end of the movie, he comes back, gets retrained as like a robot operator, right? Fix the robots. Like there's going to be, you know, an element of that for sure. But I think this concept of, you know, labor to capital shifting and is, is very, very real. And it's it's going to, you know, impact.

Absolutely. Every part of the economy, yeah. I don't think many jobs won't be some more. And I do think it will probably be resulting structurally higher unemployment because of that. And as soon as we have like the worst thing that could happen is let's say like, let's say in like a couple years when this technology is like really pervasive is if you have like a recession or some type of event that causes corporations to cyclically lay people off

because the. Then they maybe lean into the efficiency and then the jobs don't come back. Exactly. That's exactly right. So yeah, man, it's, I think the, the core of it is like, and my approach is like, you just have to like, you can't dismiss it. It's real, it's coming, it's scary, it's threatening, it's

exciting. It's all these things you depending on where you are on your career and what you're, you know, like what your net worth is, right, and like what your incentives are, you're going to be more your your priorities are probably going to be different, right, in many ways. And but if I was like, you know, I think the biggest thing is like, and I see this happening with like a lot of analysts that are younger coming out of college, like an existential

crisis around knowledge workers. And what's the value of somebody, a four year education if somebody is coming out of school, coming an analyst, right? If I 20 dollar $200 a month subscription to a chat AI service is so much more efficient, so much better. And it is like, I mean, I honestly, it is like, I can't, it's, it's just the reality. You can't dismiss it that there's margin for error like everything else.

But I've been able to substitute like, you know, 5060% of a lot of the work, you know, to AI directly. Even I created a, you know, an application model that can create slide decks for us, right? Something that on on our knowledge base of analysis that would something that would take. God slide decks take forever. I probably built 10,000 slides like forever and now I can literally give it point to my knowledge base, which is basically a synthesis of all of our research that we've done.

And I can't say, and it can create an outline and I say create this slide and it's not perfect, but I make edits to it, Yeah. And it saves me a ton of time. And so I think. And if you don't have to worry about like do the numbers tie and you just know they do and you can trust it, that's. Nice. I mean, like, you have to still give it like conclusions, but you know, we'll get there eventually where it can do a lot of this, like some of these other things as well, like more.

Well, I just mean as as easy as like, you know, you get typos and stuff when you do things manually to the extent that something and, and, and those kind of those things, when the numbers don't tie, that's the stuff that undermines your fundamental like everything else, right? That's the stuff people, at least in my career, it is. That's what people notice and then you start getting. So if you don't have to worry about that, it frees your mind

up to be creative in other ways. Yeah, I think the the biggest part, the hardest part of this all is that you and I got to where we are. Like we're sitting here and I spent 10 years hunting and pecking, reading every single filing, building models from scratch, like all the, all the learnings that come that associated with that, that are very, very like beneficial. Because for me. And then like with AI, it's like you already know this stuff.

You know what kind of questions to ask, you know what to look for. You can ask better questions because you've gone through that whole kind of experience. You have experience, right? So you can just ask better questions, tighter questions, you can learn how to think critically. You've learned how to think critically more because you've made mistakes.

That's the biggest issue that I see with AI, especially if you're like, you know, younger analysts, is that on the one hand, I want to say I've been telling people and I agree. And I, it's like you have to figure out how to use this to make yourself more efficient, right? Because this is the future. But on the other hand, there's like this paradox of you have to, you know, learn how to become more efficient. Like you have to learn how to learn. You have to learn how to ask the

right questions. You have to learn how to think critically. It's not something that's net doesn't necessarily come natural to some people that you can't, you know, because otherwise, you know, just like, just like if a junior analyst goes on, does research and gives you back work, you look at it, you're like, you know, that's wrong, that's wrong, that's wrong, that's wrong. Or you look at a slide deck, you look at it for 2 seconds, right, That's wrong. And the analyst is like, how did

you figure that out? How did you see that so fast? It's like, well, it's not because AI, it's just because it's this experience. I've done this a million times, right? And that's, you know, I think that's that element is still really important and AI can't solve for that yet. But but The thing is like, if you the whole fine tuning element that I mentioned before, because we're not at that stage, like in theory, you can actually fine tune a model for just that, right?

So like pick up on those nuances. But for younger people, it's a it is a it is stuff, right? Because you're competing against AI in many ways, there's going to be just less demand for labor because of it, like fewer new job openings, because so much of the basic grunt work is, can just be, you know, replaced. So it's kind of, they're going to have to use AI to accelerate their learning curve, right? Like I said, everything's moving so much faster.

So you're going to have to have to ask questions, you're going to be have to be intellectually curious and you're going to have to use AI to make you learn skills that you previously couldn't learn. You got there's you're going to have to create roles for yourself that didn't previously

exist. And that is something that I'm not even sure like our educational system is readily equipped to adapt because while the Internet was disruptive in the sense of it was an information, you know, of that like information, just the rate of information flow just broke open right now. But you can still anchor on like analytical rigor and critical like learning skills. You know, now it's AI because it's going to be prevalent

everywhere. I think a lot of the biases that we talked about in terms of this preservation mindset and come and see mindset versus abundance mindset is going to be endemic to a lot of institutions that they're going to view it equally as a threat. Professors are going to view it as a threat, right? And therefore they're not going to teach people how to use the tools, right to actually, because you can use AI to help you with critical thinking to solve problems. You just.

Yeah, I was, I was thinking since you've been saying this and my wife said this, so I should give her credit for it. But the the way I think that effective teaching may occur is, is more is closer to the law school experience where you're teaching people how to think and asking like, like more Socratic method type stuff, I think could prove a good way to educate

people. Because working on people's logic and working on people's ability, I mean, you said it earlier, your ability to ask good questions is going to be a good or a key differentiator, right. So I think maybe that's, that's how the professors that that want to do a service, that's that's how maybe they can lean into their role. Yeah. Because, yeah, I mean, like, I don't know, obviously there's still going to be a need for research professors.

I don't want you know, but a lot of the research might be able to be outsourced. Yeah, and, and, and then, you know, it's like your ability to learn, think it's infinite. It's everything that you could possibly really want to know at your fingertips, right. I mean, the danger of like we talked about before is just verification. Like, are you, you know, equipped and able to call BS if it's wrong, right. Like that's a thing.

But in theory, you know, if you anchor in Google search, if you anchor in other things or if you create agents that are basically like designed like I once had a, one of my best meetings that I've ever had was it was a little awkward at first, but it ended up being a really good meetings. It's really seasoned PM like former like one of the biggest hedge funds out there and follow our work.

They were we, we came independently to a different research conclusion on a stock and the PM like set up a meeting and it was like his analyst who I've talked to a lot and me and he's like you're saying this, you're saying that, but you're both are looking at the same data. Yeah. Figure it out, right? And it was like a very interest. It was a great it was. I mean, it was definitely a little terrifying and

intimidating, right? Yeah. But I kind of do the same thing with like a lot of AI is that I will train agents or I'll put the different models against each other and say like, look, you're saying this, you're saying that or you know, what do you think? Like what are the pros and cons? You basically do the same type of optimization, just an optimization and you have them break it down in steps and you can get to, you know, at a minimum what it does is it for me at least, it helps me.

It expands like my capacity for analysis because I have all these questions in my head all the time, right? But I, I can't, my ability to execute all of it is not there. I do have a learning disability that I've struggled with my entire life to help. It makes writing harder, makes kind of managing things harder. And AI has been incredibly helpful in that because it's just like an ability to it. It actually, it can solve for a lot of that in terms of writing,

synthesizing information. And so that's I think that's probably and, and for me, that's been a, a, a huge value add and that optimization piece, I think is something that maybe people aren't leaning to as much. Again, this isn't just like summary, right? This isn't just like, hey, summarize this, help me write an e-mail faster. A lot of the applications now and a lot of people are using the free models, like the free versions. It's like you can't, like you just got to go AI first.

All in experiment with the different models. Ask the same question to different models, even within the. Same. I find that very interesting and then comparing the answers and trying to figure out. I like to read the reasoning steps too and and see you know how are they thinking about things. Totally. And, and it, it makes me think harder, right?

It makes me think differently. And I'm OK if it frames my thinking differently, if it's influencing me because I just at the end of the day, I just know that this is where the world's going. Like it is like it's, this is where it's going. So it's either get on board and you're either going to ride the waves or you're going to be, you know, run over by it. And I fully intend to stay on top of the waves, not fall off the surfboard. Well, I appreciate you coming on and talking about it.

It's funny, my dad asked me at dinner just this past week what I think about AI And I, I gave him about a 10 minute discussion and I'm just going to send him this and I'm going to be like, I think you should listen to this. And I I think he will enjoy it very much. I appreciate that and thanks. Yeah, thanks for, I mean for reaching out and we should definitely do more follow-ups like as things progress because it's moving so much fat, it's moving so quickly, right.

I know you said earlier that you're glad you didn't do any recordings in last month because it would proven a lot of things wrong. And but look, I'm totally comfortable and OK with the fact that maybe we'll look back six months a year from now, two months from now, and everything has come out of my mouth and everything just talked about just just proven utterly and totally wrong. And that's all this. This is different.

So I, I think one of the issues as, as a podcast at this stage is like I've noticed with my own listening, I don't listen to everybody's episodes anymore. There's so many podcasts out there that it's like I almost don't even there are some that I want a weekly, you know, sort of like something put in my head. My commitment to my listeners, at least going forward is I'm going to try to make sure that if something's in their ear, it's because I want it to be and it's not because I'm trying to

Juke some weekly release. So Spotify recommends me, right? And I just, there was so much macro news flow and I found it so interesting and Anna, it was moving the market so much. It was what I was focused on. So I, I don't know, man, I probably would have ended up chasing some sort of, you know,

macro guru of the of the day. I mean, I had I had Andy constant on that was kind of macro E, but I'd followed him for a while and it wasn't it was inspired by what was going on, but it wasn't chasing it. And I just I think it had I been booking a lot, I just think I would have probably gone after the wrong content. So I'm glad I didn't, if that makes sense. I say plenty of stuff that ends up being wrong is not concerning.

Yeah, that's why I like always when you reach out, it's always love coming on here because like I can't, it's like feels authentic and just, you know, genuine. And that's that's important. And yeah, I think, you know, we're going to, you know, maybe maybe the next step at some point is like we talked on like kind of high level across the board and a lot of these different like AI topics.

But yeah, I feel like a lot of them each and itself, like even like the Google discussion by itself could be like on search and what's happening could be like it's own separate things. Yeah, well, let's do a follow up and like, I don't know, like 2 months or so. Let's let people listen to this, digest it and hopefully get, you know, start playing with the tools and then come back on. Yeah, I I love talking.

I. Think I think it's great listening to this like it's all one big experiment, right. So like if you have any questions or thoughts or anything that like as you're listening to this, any inspiration, any ideas? Any questions like don't keep them. I mean, you know, don't keep them to yourself. Like if you want like throw them out there like on the X or send them in to Bill and we can talk

about next time. Or you know, it's I think part of this is just that sharing learnings for all of us is something that in a world where informational advantages and edge is so closely held close to the best, This is this is a technology, this is a phase that we're all entering together that I think we actually all benefit tremendously from sharing our experiences and our learnings. Because at the end of the day, you know, this does impact all of us.

And that's foster community of intellectual curiosity related to that I think is really cool. And I know your audience is like the perfect audience for that. And hopefully we get some good feedback and thoughts from it as well. Indeed. Well, have a good one. I heard your kids screaming in the background. How's that going? It's going well, man. Like he's he's just the three and a half 18 months the two boys are beating the shit out of each other like as they.

Shocking to stop, man. Yeah, they're trying. They're trying to. Keeping them alive is proving to be an interesting daily task in his first BMX biking race on the starter bikes. The other. Day which? Is pretty cool and you know some he's just a total. He's just a killer. He's just like running around getting into everything. But dude, they're happy, man. And that's like the only thing that matters, right? Like honestly, it's that's my

goal, right? My goal is to make sure that they're just happy and have a fulfilled life and being present as a father or something. That's really top of mind for me too. And it's hard to do. But I don't think any, I don't think any type of AI is going to replace that any type any. No, I don't think so either. The mine are, are now, you know, getting older and starting to

think about teenage years. And one of the things that I've been pushing myself to do is make sure that, you know, I'm leading by example and, and making sure that it's not just talk because, you know, they, they watch and they learn from from who we are. And I don't know, it's been, I want to, to the point of this whole conversation, right? I I want to show them. A father that's able to adapt and able to embrace new things and not be so scared.

But it has been a little bit difficult, man. It's been, it's been an interesting six months or a year, but I'm mostly through it. We'll see. Yeah, you know, it's life's a journey. No doubt, no doubt. All right, my man. We'll talk soon. I look forward to the follow. Up. Yeah. Thanks, Bill. Appreciate. It all right, take care. None. The.