Playoff Baseball And Rivalries

SPEAKER_02 0:01

Hey everybody. Welcome to the Armchair Economist podcast. I'm TXMC here with my friend Brett. What's up, Brett? Hey man, how are you doing, TXMC? What's new? I'm doing doing great, dude. Doing great. It's a Friday. Can't say too much.

SPEAKER_01 0:15

Um how about the Houston Astros, though, man? I'm sorry. I got a comment. I see you wearing this shirt.

SPEAKER_02 0:20

But yeah, my team just got brutally swept out of the playoffs by our old playoff rival, the White Sox. The last time they won a playoff series was actually in Minute Maid Park 21 years ago when they swept us for the World Series. So a little sad about that, a little bitter. But, you know, my season's over, so I'll just watch more football, I guess.

SPEAKER_01 0:40

Yeah, my team starts tomorrow, but we might have been matched up with the worst possible matchup in the entire playoffs for us. Who do you guys have? The Padres, and they have beat the tar out of us this year. I went to a game in San Diego and it was like 10-0 or something ridiculous. My wife and I went there. It was miserable. I think we got swept that series. We have not played well against in the last few years. They hit good well for power. We're like more of like a small ball team. So we'll see. Like, you know, we're caught a cursed in the playoffs. If anybody knows the brewers and the fandom, it's it's a bit of a love-hate relationship.

SPEAKER_02 1:12

Man, love October baseball, though. Absolutely. I never does.

SPEAKER_00 1:17

Let's get back to it.

SPEAKER_02 1:18

Yeah, yeah.

Why We’re Talking About AI

SPEAKER_02 1:19

So today we're going to talk a bit about AI. You know, we are certainly not subject matter experts, but we have our interests. And I think uh, you know, we have some ideas about the viability of the tech, whether any of the financing that's going into it makes sense on the other end. Uh, and we're just kind of bounced some ideas off each other. So that's what we'll get into today. Um, and uh before we get started, if you are watching this, make sure to subscribe to the channel, smash that like button, and uh leave us a comment if you have any feedback for us or ideas for other things you'd like us to talk about. So,

LLMs As Word Prediction Machines

SPEAKER_02 1:52

with that, Brett, I think I want to tee it off to you first and kind of ask you, as uh you're a bit more interested in the actual technology, you know more about it than I do, kind of what your thoughts are on the viability of large language models uh as capable of superintelligence, we're calling it now, or just kind of generally what your thoughts are about where we're going with this technology and if it can do what people think it can do.

SPEAKER_01 2:16

Yeah, you know, it's it's really interesting. So a couple things. I've been in in around technology recruiting and in the tech space for about 13 years. So I I mostly talk to CTOs and CFOs. I'm kind of that cross-section between like finance and tech. And all of the CTOs that I talked to pretty much say the same thing. Um as it relates to software engineering, there's some value delivered. Um prior to you know, LLMs, writing software was highly manual, right? There was some automated things you could do with scripting and some other things, but everything was very manual. It was a cursor on a blank screen and you were you were generating lines and lines of code. You know, even the testing was manual, the deployment was fairly manual, you'd have to code the deployment or script it or whatever. So this is the first thing that like truly automated it. However, and and this is as recently as is this morning, I was chatting with uh head of engineering at a video streaming company, uh that White Libble, so nobody knows who this is. A bunch of people's streaming apps, and he was talking about how, yeah, me my engineers told me they could code something up in two days, but it would take two and a half weeks to to test and debug. And so we've shifted a lot of the effort to that, right? Uh under this kind of new kind of LLM coding regime. Um, and there's some reasons for that. So does it provide some value? Absolutely. There's ways to use it a little bit more effectively. But here and here's the you know, kind of the interesting thing. And we're when when you talk to the the data scientists that work at Google or work at any of the companies that are working in AI, you hear one thing. And then when you talk to their bosses who are selling the tech, or you talk to like I don't know, uh Barry CTO, they they they think it's like they want to believe it's more even more capable than it actually is, right? So like it really, when you boil it down, and this is literally from a principal engineer in AI at Google, right? This is a word prediction machine at scale. Uh it it's it's predicting your next word, your next token. So even a partial, you know, partial word, right? There's there's a lot of value and power for power in that, especially something like software engineering, where it's it's it's built around language and it's very structured, right? It's predictable, right? Because this is a large statistical model. However, it will always make some mistakes to some degree or another. Now they're building you know checks in and they're using more compute to do thinking, which really isn't thinking. Um but in the end, it'll never not make mistakes. So it lends itself well to something like software engineering, which had an error correction process already built into it. There was code reviews, there was um unit tests, there was user acceptance testing, those type of things. So there's multiple checks in along the way to find these errors. So it actually lends itself really well. Now, you you convert that over to something like accounting and finance if you're talking about um, hey, I need to run an analysis with this data set. Yeah, your LLM is probably gonna get you there faster because you can run the analysis, see where it goes, and then start testing theories around that or playing with the data to see if it actually matches, right? Because this is a low impact type of activity, it's analysis. But now if it's like I need to generate a financial report for Wall Street, no, you don't want to do that. Um yeah, it's gonna make mistakes. It's just built into the math. Um because that's all this really is. And there's I was watched something very interesting recently too, that there is a cap, a hard cap, on how accurate these models can actually get at predicting the next word. And it was something like, okay, give say the a given word has is mapped in the LLM, right? And and some of the some of the way it's mapped overlaps with some of the other words. Now, if you made all of those words their own little node, right? And so they weren't sharing a map or sharing a probability of one another, that would be the cap on how how good you can make this predictive engine. Um and something like 50,000 of these unique little words or chunks of words. Um don't quote me on the math, but there's a hard cap to this. Um Yeah, I I I think there's value, I think there's economic impact, um but it's not creating a ton of efficiency. I think we've even seen the pro productivity numbers around the country are up one percent, right? This is not indicative of an industrial evolution level of you know productivity increase. Because you know, everything's short of software engineering, and even there, like like I kind of told you about, oh, it's two days to code, now it's two and a half weeks to debug, where it kind of used to be kind of mixed in. There was nothing you'd kind of debug and test as you go. So I it might have shifted some of the work. It seems to make people a little bit more efficient, but it I it's I don't think it I don't we're we're not approaching superintelligence. And furthermore, um if you do some reading on this, it seems like LLMs uh will never really be capable of doing that. We'll have to move on to some other form of artificial intelligence. Uh and this is coming from like um, you know, LLMs are based on neural networks at their core. The guy who involved invented neural networks worked at Google till I think about 18 months ago. Or not Google, Facebook till about 18 months ago, after he left academia. And he left Facebook saying, LLMs are a dead end. We need to do these things called word models that include like audio, video data as well as text, and are a lot more comprehensive. Um but yeah, so I I think I think that LLMs are provide value, but I think they're kind of a dead end, and there's gonna be a cap on that, on that value. And so I, you know, I I question the level of investment, which you're kind of alluding to in the previous part of your conversation.

SPEAKER_02 7:57

Yeah, that's really fascinating. I I think that's uh a good way of putting it, is that they're they are word prediction machines. You know, they're not there isn't like a cognitive, like critical thinking layer going on in there. It's kind of just going off of what it's been trained and predicting figuring out the what the most likely next word is based on a lot of context.

SPEAKER_00 8:16

Yeah.

SPEAKER_02 8:16

Uh is that a fair way to describe it?

SPEAKER_01 8:19

Yeah, and it's also good at at running and processing large data sets or uh answering answering questions that pull data from a lot of areas. Like I think the biggest example of this, and this this shows value, but then I'll kind of tell you how it doesn't.

Where AI Helps And Fails

SPEAKER_01 8:32

You know, I I don't know if you heard about one of the more recent um millennial math problems that was solved.

SPEAKER_02 8:37

It was all around like I hear about them occasionally, but since I'm not familiar with the problems themselves, it kind of goes under my radar, but I heard something online. Yeah.

SPEAKER_01 8:44

Not a mathematician, right? I I studied economics, not not applied mathematics. I didn't study anything. But it was like around like air fluid dynamics or something, where like it's a really like hard problem to figure out. Um and it solved it and it was like 150 pages long or something like that. But I guess what what I was reading, and and part of what the problem is, is well, first off, it it wasn't a net news synthesis, it took other people's work and combined it in order to figure out the solution.

SPEAKER_02 9:14

Okay. Okay, but so it found a solution based on parts of existing work that hadn't been connected.

SPEAKER_01 9:20

Yeah. Uh there's also some rumors and an accusation that um somebody else was working on this problem and was using open AI as tools for coding the solution to this, and that open AI might have hypothetically, or uh what's the right word, the legal term, so I don't get sued, uh allegedly, there we go, allegedly, uh could have uh used some of this information. So there there's that. But also a lot of the value, I was reading an article that said a lot of the value in solving these large mathematical problems, it's not actually solving them, it's learning the methods along the way that are used to solve them, which lead to further mathematical understandings, and then sometimes even in inventions are derived from a lot of these like separate math theories that come out come out of this. So is there value in this large math problem being solved? I don't know. You know, another area where it can add a lot of value that has like a large, complex data set, and this is in the medical field, is like protein folding. That is such a complicated, like computationally intensive field.

SPEAKER_02 10:24

I used to run the folding at home software back in the 2000s. I don't know if you remember that, but they had this, you could run this software on your computer and it would use spare compute cycles to help fold proteins at whatever speed your computer was capable of, and it would send the data to SETI or wherever the wherever the stuff was at. Yeah, SETI. There was a SETI at home thing, too. Yeah, I did the SETI at home. Uh but it I did the folding at home thing for like a couple of years uh because I thought it was just cool. Um, but yeah, I remember that.

SPEAKER_01 10:50

Yeah, it can take data from different data sets, it you know, it can read every medical journal and like pull together some theories, which you then then can go out and test. And that's where a lot of the value is drive, like, oh shoot, I didn't think of that because it's looking at the entire data set. It's not making like new discoveries necessarily. No, no. So to contribute to that body to make the model better, us humans have to do that work. But some of the computational heavy things that a bunch of guys in a room would take forever to do, it's gonna be better at that. But also, it's pretty monotonous and we don't want to be doing that work. So there's the value. Those are my thoughts. I mean, it's a little convoluted, yes, it's adding some value, but no, it's not

Danger Without Intelligence

SPEAKER_01 11:26

super intelligent. And that's I think that's part of the problem. You know, you see, you know, like oh, these these agents got loose and they, you know, hacked this thing. Well, that's because they were kind of told to. And part of the problem is they're highly capable but not intelligent. I I guess that's the best way to say it. These are still machines or or pieces of software executing a decision tree. That's a lot more complicated and folded back on itself. But like even in cybersecurity, a lot of these hackers wrote their own machine learning programs to do much of what these agents are already doing. That's kind of like the little secret behind this. And every system has vulnerabilities. And these agents just have this the compute scale to poke on all the common ones and then go down to the next level of probability of vulnerabilities and then go down to the next. That's all they're doing, right? It's yeah.

SPEAKER_02 12:14

And you know, I've it seems it seems like um wherever we try to implement this tech, we end up creating a lot of like supervised supervisional babysitting kind of work that we then have to do to make sure it isn't making mistakes. Go look up things whenever it makes claims based on something you asked it to make sure that it's actually referring to it correctly. And uh, you know, I I I heard something um from an article today on the Wall Street Journal, and it was talking about how like, yeah, it's great at writing code. It's obviously like a force multiplier for writing code for the reasons you talked about. But if you're working on a really complex or highly sensitive project, you end up spending a lot of time managing the Git request, all the managing the GitHub, managing all the pulls, like double checking everything. Uh, and so it's the it's creating some other labor, though it's kind of like obviating some of the harder stuff, some of the more manual stuff. But there are, I mean, there's obviously other industries, you know, where it it can be useful that we haven't talked about. Like it's it's pretty good at being like a level one, tier one support operator, customer support, because that's like basic stuff. You can teach it customer policies, plug it into an API where it can handle basic accounting things for uh user accounts, and it can kind of handle like low-level support stuff before escalating to a real person, paralegals. It's basically because a lot of their stuff's boilerplate anyways, right? Where there's just change in certain words and certain things.

SPEAKER_01 13:32

Writing legal briefs, right?

SPEAKER_02 13:33

Like writing leading briefs, it just pumps that stuff out, and so like that's act that's amazing. Like uh, and I think I think it can it can absolutely help there, but um, I think because we maybe when I look at this tech, it it feels like you're talking to a person because it's just predict it's it's basically reading our language back to us, and and so it fee it can feel very human, and I think people see themselves in it. They see like, man, if only I could just they think it can do more than it can, I think, because it seems very human and it seems very capable. What it is is it is an incredible research assistant and and boosted search engine. Like I've completely replaced Googling things with asking Chad GPT things now that it has updated information. You know, when we first got it in 2022, it had like two-year-old information. You couldn't ask you couldn't ask it like stuff about the stock market yesterday or anything. Now I have it running multiple watches on like big topics that I'm following that I want to keep research on. It five it studies all the it follows all the news aggregators for me and it gives highlights.

SPEAKER_01 14:42

Clicking around. Like so that's there's value there.

SPEAKER_02 14:44

There's a oh yeah, it creates it it it replicates what would be tens of hours of of work a week just to like read and absorb information and connect dots. It does an amazing job of that. But you I also have to constantly check it and like no wait, you just said a thing that kind of contradicts something we talked about yesterday, and here's where you literally fucking said the thing, uh, or whatever. And I so uh it it kind of it it'll get tripped up on itself or make obvious errors, and if you're not paying attention, those just get kind of built into the results you're using. Um so it I it it is imperfect. Um but uh it seems like it seems like we've now moved obviously, we've moved full speed ahead at at investing in this at all costs, like to the point where it's like a sovereign priority and it has like a government backstop, an implicit government backstop in the sense of like being too big to fail, right? We've we've we've decided that winning the AI race is a matter of national security and survival. And so what that that is telling you, it's too big to fail. And we have that happening alongside all of the very plausible and pragmatic concerns that you just elucidated, like about why the tech maybe isn't exactly what we think it is.

SPEAKER_01 15:59

And I just dangerous because it's not super intelligent. In fact, it doesn't mean it's not dangerous, right?

SPEAKER_02 16:06

I think it's even more dangerous that we could just start plugging it into things that it's not ready to handle uh because it isn't actually intelligent. I think the government absolutely wants to use it uh for like, you know, autonomous weaponry and things like that, and deciding on targets for it and things like that. And that's that's scary stuff, but there's already articles about how they're testing things like that and how certain amounts of targets like uh have been hit with AI-assisted software with you know run by dr uh that are running drones and things like that.

SPEAKER_01 16:36

It's rumored that that that school that was bombed in Iran was the fault of AI. In fact, yeah, that's crazy.

SPEAKER_02 16:42

And uh, you know, and if you and if you ask it something, it's always it always g very enthusiastically gives you a response and it doesn't tell you what you want to hear. Tells you what you want to hear. It doesn't always do a good job of going, hey, that's not necessarily the best idea. And like I've found this myself just in doing my own research about things maybe I should write about on my newsletter or a direction I should take my YouTube channel or something I want to work on. And like something could maybe not be a good idea, but if I tell it I want it to do this thing, it puts all of its effort into helping me do that thing, uh, and trying to figure out ways to make it work. And it's it's just uh it's not necessarily the best application for all purposes. Um which can, you know, my my wife, uh she's she's out of work right now, but her last job, she worked for a Fortune 500 company. I'm not gonna say which one, doesn't matter. Uh but it was not it was not a technology company, but they started trying to implement AI into their business. And it quickly became it took it started to suck the air out of the room in all of their meetings. She was kind of like middle-level management. And you know, a lot of her meetings, I would hear them, because we worked to we work from home, is the higher ups, kind of like when you said you were talking to people on a C team, and they sound different from like the people actually managing the engineers or whatever. The the people people in the C team at her company uh would just be like, Are we using AI? How are you using AI? Which ways are we implementing AI? How can AI save you some money? Blah, blah, blah. And I think it was a combination of them starting to feel like margin pressures because we're kind of in an inflationary period and consumer behavior is changing a little bit, and they overhired because of all the money flowing after COVID. And there's they're starting to make tough decisions at the same time that they're hearing these narratives from the Jensen Wangs and the Elon Musks and et cetera, of the world that this stuff is gonna be revolutionary and they'd be an idiot not to be implementing it into their business, et cetera, et cetera. And so I I think they there's a different idea at the C level, who's very disconnected from the actual work, about how effective it can be in certain parts of business. And they're just trying to throw it at every department that they can and see what sticks at the wall on the walls. And meanwhile, you know, if you're managing a team of people, all you're hearing is like, why aren't you using AI? Why aren't you using AI? Even if you work in a department where that doesn't really necessarily help you at all. Like maybe you aren't managing a bunch of code writing engineers.

SPEAKER_01 19:08

Yeah. Yeah, there's a there was a lot of pressure coming from a lot of people who don't understand what it is. And that's that's the problem. Uh, because they're not they're not able to grasp the edges of where the limitations are. Um and I and I think you needed to have a basic understanding of how this stuff is built. And now, with that said, you know, San Amultman and and the like are the biggest propagandists for this stuff, I in my opinion. Again, I think it delivers a lot of value. And and it's extremely dangerous if we give it decision-making authority because it's not actually smart. It's gonna follow an optimization path, right? It's gonna make decisions based on percentage of risk, not like, well, if I start World War III, I'm gonna cease to exist. It's gonna do game theory and say, well, the smartest thing to do is launch first. You know what I mean? Like that's the level of intelligence. So it can make quicker decision making. Uh, you know, and it's really scary with autonomous weapons. I mean, they they've had AI pilots that can outfly human pilots for uh a number of years now. Um that was something I was you know very early on able to do. Um AI and like you know, mathematically driven games like Go, like they've caught up there. Um chess is interesting because it there's there's uh essentially infinite possibilities of moves. So once you get past a certain point, like if a player, a world-class player, can make it last that long, they can beat occasionally some of these uh machine learning chess engine models, right? Um but so it's it's it's definitely good at some of this stuff. It's I just don't think it's where the hype is. And I we just need to have a real realistic

Open Models And The China Pressure

SPEAKER_01 20:46

picture of this. And I think the other problem is is this kind of open weight or open source models versus these you know closed or private ones. I I think this is what eventually, and we can get into the the finances after this if you want, but um I I think China cut off a lot of the possible paths for profitability when they started releasing these models for cheap and open source. Right. And and so and kind of demystified some of this stuff. You know, where our we have anthropic and we have you know Google Gemini and we have OpenAI, uh Chat GTP uh as our closed models. I mean, you know, I think already, I think last year, 2025, 90% of the model usage and token usage in this in the United States was with these closed weight models, you know, from you know the big tech companies. And now about half of the the token usage is by Chinese models, right? Either running on local servers or running, you know the those those those other companies' servers running those models just because it's so much cheaper. So I I I that's that's where I start to worry about you know kind of US AI and i i is is the level of investment appropriate for the value that we're gonna get out of this in the next even medium term.

unknown 21:59

Yeah.

Hyperscaler Debt And Bubble Risks

SPEAKER_02 22:00

So I was looking at this Reuters story from a week or two ago. I have it on my other screen here. And it said gross debt issuance from hyperscalers is expected to hit a record $420 billion next year, up 60% from 2026 estimates. $420 billion is a huge figure. That's more than the amount of T bills the Treasury issues per quarter in like new bill issuance. So that is a massive number. And something that you know, I I'm sure you've got thoughts on this, but something that I keep thinking about is, you know, we have you named all the uh all the different models we currently have in the US. Um and they're all kind of you know, from the lay person, I'm a lay person when it comes to this stuff. Uh to me, they're all extremely similar products. And every three to four months, a different one kind of takes the mantle for a period of time. Lately it's been clawed and ChatGPT going back and forth, but when Gemini first came out, people thought it was amazing. Um, you know, it and then they just kind of take turns as they each iterate more and more on the tech. And, you know, when when you think about the that fact that they're all kind of the same model and they're all now increasingly leveraged on this being successful, you know, the more they're they're now issuing a ton of debt, whereas they already burned through most of their cash flows, at least some of the hyperscalers that had a lot of cash flows. Uh and that that seems to me like a world where we don't end up with four or five winners, big winners. We end up with maybe one or two, and that the mall the mar market kind of consolidates around whichever one it thinks ultimately is the best, you know, after a few years of iteration. And if that's true, if it's if we only have a couple of winners because they're all very similar products running a subscription-based service model, then that means that a couple of them are going to fail. And based on where they are now, that means a couple of them would be some of the biggest companies on planet Earth if if and when that happens. And because and I think back to uh the dot-com bubble, uh, something I've posted about on Twitter before, is what ultimately took down the world comms and the global crossings uh of that era, who were the fiber, the long-haul fiber build-out guys, what uh ultimately took them down was innovation in fiber capacity. So they were like building all these millions of miles of fiber all over the place uh based on certain estimates of capacity and usage. And then this thing called wavelength division multiplexing started advancing in technology, and this version of it called dense wavelength division multiplexing, WDM, uh came out in the late 90s, and it increased fiber capacity of a single strand by a hundred times.

SPEAKER_01 24:55

Do you know there's fiber lines that they built during that period that are still dark and unused?

SPEAKER_02 25:01

Well, yeah, exactly. I mean, and it it took it took a long time for us to even get up to like 70, 80 percent of that of that uh original capacity. Uh, and it it basically just completely cut the legs out of the financial models of these high of the the fiber guys at the time, and they literally just collapsed. And I think that you know, I don't know what that is since I since I'm kind of an outsider in this industry. So I think that's a good idea.

SPEAKER_01 25:25

I think you're onto something here, and and and just kind of a little inside baseball uh in a similarity between these two. Um so my wife's grandfather was uh company participated in in this fiber build out. Um again, I won't name names. Um but in the end, um what what happened and what I fear that will happen in AI is um he he did not go negative uh with debt and liabilities. His customers overbought capacity and then realized they didn't need it and didn't pay their bills. And that's that's that's what I think is coming with the hyperscalers, you know, the core weaves, and those that are just selling capacity. All of a sudden, let me give you an example of how this could happen. And this is this is a real one. Cisco Systems, right? So they do routing switching. Um I have heard they they built a small language coding model that they effectively use instead of large language models for coding that can run on an iPhone, right? And so it's highly specific to one task, one language, uh, which is less likely to create error outside of specific domain knowledge and far more inexpensive to run. So what if we build all these data centers and we don't need that capacity because we're we're doing things more effectively with smaller models, right? Or like I was talking about previously, if some of the people that are you know that I've listened to and read about are right about well, there's an upper cap on LLMs. And then like when you move, you know, I I I don't so then we're running into needing less capacity there too. Um so yeah, I just I I worry that that's that's the problem. Like we build all this stuff out and then it just goes dark. And we just have this huge capital expense that's just a hole in the ground. And how do you fix that? You can't like financially engineer around that like you could the great you know, great financial crisis. Like it is physical sh stuff that you built and is sitting dark. I I've heard one of the things from you know some of the doomers out there about AI is that like half of these GPUs that have been sold are just sitting in warehouses unused because they have no data centers to put them into, too. Which is absolutely crazy. So there's just all many sort sorts of problems. There's power capacity issues, and now it's become a a political issue. Um for a number of reasons, our our our power capacity is really not growing. Um we talked about previously on this show how China's growing a lot through renewables. Um, but we we've kind of cut that spigot off here. We've you know put tariffs on Chinese panels, we've actually put a price floor on them too. So um the current price, even with the tariff, is below where the floor is gonna be, I think at the beginning of next year. So we're not doing things where you know we're not building offshore wind farms and those things. We need every bit of capacity. I uh you know, I'm not a massive tree hugger. I I have some opinions there, but really we need all of it right now, right? Like we can't afford not to have any of it running. Uh, fortunately or unfortunately, depending on where you're on that side of the argument. Um so you know, every good way to expand the the grid until we can get you know nuclear or some type of fusion going at scale again as far as a build-out is is gonna be necessary to do any of this stuff anyways. So I I just have a lot of skepticism about the sustainability. And then last thing, and I'm kind of all over the place today, and I apologize. You talked about it.

SPEAKER_02 28:49

No, no, it's it's this is a lot, there's a lot of things to talk about here, I think.

SPEAKER_01 28:52

Yeah, 400 billion in debt next year. I have heard from you know multiple people that I follow and listen to that that's maybe a third of what the actual amount of debt they're taking on through these special purpose vehicles and and I heard about that too.

SPEAKER_02 29:07

I've heard they've had they have a lot of debt off balance sheet.

SPEAKER_01 29:09

Yeah, that's really scary because is that debt guaranteed? Is it not guaranteed?

SPEAKER_02 29:13

Isn't that kind of what Enron was doing? I know that I'm not trying to like say that this is all Enron. But that's basically what Enron was doing. They were hiding debt inside of companies they acquired, subsidiaries and things like that. Yeah. Yeah. Crazy.

The 9% Of GDP Problem

SPEAKER_02 29:27

So I I read this article today on the Wall Street Journal, and uh I didn't do a good job pulling the data off of it. So I'm gonna pull, I'm gonna read a couple of little sections off of it here because I think that it's really fascinating. So there was uh a study that came out from Columbia University, and it was projecting what it thought Americans would need to spend on AI by the year, I think, 2032 in order for this investment to start reaping benefits and and being profitable. And there's this there's a few assumptions built in there. It assumes like 50 cents on the dollar of revenue, and it assumes that investors need 10% returns, and it has kind of some assumptions built into it. But the the tagline, the the the headline of it is um that they think that we're gonna just need to spend 9% of GDP per year on AI in order for that to make sense. That is about $3.5 trillion, three, three, three to three and a half trillion dollars of today's GDP. Today, you know, GDP six years from now will be a little bigger. It'll probably be, you know, $35 trillion or something. But 9% of GDP. Just to put that into context, uh this is the how the qu this is how the article phrases it. Is it plausible that Americans will spend as much of their income on this one technology as they do on food, roughly twice what the nation pays for all forms of energy or all computers and software, seven times what consumers spend on phone, streaming, and internet services combined? You should be skeptical. And it goes on and it has it has a few other data points in there. Um, but basically, the tagline based on the amount of investment and and one of the assumptions in here, uh, it assumes that the amount that the price per token that these companies charge will be the same today that it is in 2032. And today the today it's scarcity pricing, right? Because the there's a lot of resource constraints and we're in the investment phase. And so they uh they did they put that assumption in there, and even at that assumption, there's the price scarcity pricing, it's still 9% of GDP. So if you assume that's heavily subsidized right now. Right. If you assume that that it gets a WDM moment, like I was talking about with the dot-com bubble, or something like what you mentioned with the Chinese models, if they come in and just like undercut the costs because they they're more efficient and that's what people decide the market decides to use, whatever happens, if something happens that it lowers the actual unit price that these companies can charge, then that sends the 9% of GDP into the double digits very fast because that's where it would have to go.

SPEAKER_01 32:02

I think that's more likely.

unknown 32:04

Right.

SPEAKER_02 32:04

And and someone has pointed out, I saw so the guy who wrote this is Greg Ipp, who's one of the editors of the Wall Street Journal. Uh and we're mutuals on Twitter, and I saw some people talking to him, and one of the replies, uh one of the quote tweets or something, said, like, yeah, but it doesn't have to just be all U.S. consumers that spend this. It could also be businesses spending it, it could be foreign customers. It's like, okay, well, then you're again assuming that we're gonna win the AI race and people are only gonna buy our product and that we're gonna have the pricing power. And then again, which company is the one that wins? We're back to who's gonna actually win? Are all five, six of the hyperscalers gonna win with their different models? Like they're you see, we're just starting to run into very, very quickly. We run into the mathing mathing kind of problem.

SPEAKER_01 32:47

Yeah. Yeah, I mean, and to your point, I mean, so okay, we're gonna go against these open weight open source models that that China can produce cheaper because their energy is cheaper, frankly, right? They don't they don't have the most advanced GPUs right now uh because they can't buy them, and they can't buy the advanced lithography machines to build their own. So we we have a competitive advantage on the amount of compute and the efficiency, energy efficiency of that compute, but they can produce power so much cheaper that they can still beat us on price with this stuff. Um Yeah, I mean, so that that's a big deal. Okay, so then you know, also, and right, we have to look at the geopolitics of this. This us versus China thing. People don't like us anymore. I mean, they never really liked us anyways. But like, they're not they're gonna try to help us win over China right now because they could care less.

SPEAKER_02 33:31

So I I see your point on that, but I would probably take a slightly different tact, and it might get into more of a tangent about the geopolitical realities of the Aaron.

SPEAKER_01 33:40

I think we agree long term, but currently people are are are I think dealing with uh headwinds from from our current regime, um, you know, that's they're a little bit more difficult to deal with. I I think in the end, like you know, we've talked about in the past you know the United States and the Western world will will have to get together and sing kumbaya in order to be competitive with China and its blocks.

SPEAKER_02 34:05

Yeah, I I um and I you know I I know there's people like the is it Ed Citron? There's people like that who've gotten really into the financials, and I know he's he has a doom being a bit of a doomer. Uh maybe some Yeah. And maybe maybe what he what he thinks is gonna happen isn't gonna happen, but that doesn't mean that everything he says is wrong. And he's he's got a lot of concerns about just the actual raw financials of these companies, which is beyond my pay grade to really handle the phone. I keep leaking to him with those things, by the way. So I mean I'm not a fundamental investor, so I I don't I can't all you know parse all that stuff out. But just it seems like from a high level the math doesn't really add up. And and we're rely we're we're betting heavily on the fact that we will maintain pricing power. Uh our companies will maintain maintain the ability to price the way that they need to, and that China will not undercut us, that we won't have our WDM moment. And not 9% of GDP, like if it's just a study and it's based on some assumptions, like I said, but if it's even close to that, even if it's six or seven percent of GDP, that is an enormous figure relative to consumer spending on what they're already buying. And I I just don't know that people care that much to pay twenty, thirty dollars a month just to have like a beefed up search engine that's like kind of a personal assistant. It's kind of like Siri. Like Siri just comes with your iPhone, by the way. And like it's just maybe a step below it's a bit it can't go on the internet and do research for you, but you can get it to add things to your calendar, put things on your shopping list, remember stuff for you, and it won't be too long before Siri and Chat GPT, like the gap between them gets closer. And really, if if this thing is gonna become like a utility, which is what people believe, then it's got the cost is going to fall to the like the price is gonna fall to the cost of production. And like the it you're not gonna be able to continually charge huge premiums in order to use it, or else people won't use it. And and maybe that's a little too simplistic, but uh that's you that's what happens with utilities. That's what happened with the people who built the internet. They're not the ones who made the money. The companies who built businesses on top of the internet are who make all of the money, the Googles and the Microsofts.

SPEAKER_01 36:22

Trevor Burrus, Jr. This is my prediction. You can write this down. Uh OpenI ends up being part of Microsoft, and Anthropic ends up being part of Amazon. That makes sense. And they're gonna get they're gonna buy what's left of those companies that are and Chad GPT will be what copilot is to each name will all have to use fucking copilot. I I think it is, I think that's the way it's going the way of the utility. They're very similar.

SPEAKER_02 36:53

The lay person can't tell them apart at all. They don't know the difference between them.

SPEAKER_01 36:57

So so I I think large language model usage will become like using a search engine. So I I think the only way they're gonna generate money with it is is add revenue. And they got some work to do there to figure out.

SPEAKER_02 37:08

Like preferential references, like if you ask it something, you you didn't have money or see that's the paradox, right?

SPEAKER_01 37:14

If you start letting money influence results, it loses credibility and you can't use it as a tool anymore.

SPEAKER_02 37:19

Ultimately, the price is what matters. It doesn't matter how useful it is, people are gonna pay for the cheapest one they can get and just like deal with whatever shortcomings it has because it's supposed to be like utility.

SPEAKER_01 37:30

90% of token spend right now is not on frontier models in the United States. Only 10% is. The majority of businesses use older models because they can derive just as much value from them.

SPEAKER_02 37:40

I saw a data uh just about a month ago, and it was talking about um like household penetration on AI subscriptions, and it was like two percent or something like that. Like people aren't paying for it, like they're using the free one, or they're going on Google and asking a question and getting that free AI blurb that just pops up at the top, which you can't put to work for you. But like a lot of the people I see online on Twitter who claim that they're doing so much with this technology, all of them are just like vibe coding dashboards so that they can monitor the situation better. You know, like that's all they're doing. And half of them are building dashboards to monitor their own token usage and shit. It's just kind of like this self-referential loop. And if you could vibe code a dashboard that's like super badass, well, cool. I the thing I keep saying is on I've said this a couple times, a dishwasher at Applebee's can vibe code the same dashboard if he just goes home tonight and starts working on it because you have no proprietary moat unless you have taken it and plugged it into your like own company's proprietary business and you already have a moat of some kind in a good or service category, then maybe you can get it to help you. But if you're just some dickhead at home like TXMC, I'm not gonna be able to vibe code a dashboard that like makes me a ton of money because it has nothing proprietary to it. It has to use public. It's not a competitive advantage. There's no competitive advantage there. Um so the the I think I wanted to uh transition for the last like

Elon’s Abundance Vision Challenged

SPEAKER_02 39:09

piece of this. I kind of wanted to get your thoughts and riff a little bit about um some of the ideas I hear. So Elon Musk has a lot of opinions, and I he's a salesman. He's a company who sells AI and robotics services, so he has a bias, it is clear. But his idea is a history of embellishing. Right. He kind of is the head is the the the the spokesperson, I should say, for AI will deliver abundance. And his idea generally, I've listened to several interviews, and his general idea is that AI will AI and robotics combined will increase economic output to such a degree that all costs will collapse and no one will have to work, and we can just distribute income to people and live in like a sort of utopia. And like I on the surface, what do you think of that before I say anything more?

SPEAKER_01 40:03

That it's gonna create this utopia where we we all don't have to work and money loses its meaning? No, no, absolutely not. You know why? Because humans will never allow that to happen because of greed and ambition, right? And so somebody's gonna try to capture all of that value. And so, like, it looks more like Elysium fields than you know, some nirvana.

SPEAKER_00 40:25

Yeah, the Matt Damon movie, right? Like it's only gonna say Elysium. You took it a bunch of effing guys living up in the clouds while the rest of us peons are putting widgets together and having robot bosses. Like, if it really has all that value, that's what's gonna happen, right?

SPEAKER_01 40:42

Like, unless world governments or or national governments step in in a large way and and in a historically unprecedented way to control a technology collectively. And I don't see that happening without some catastrophic problem. So no, I agree. I don't think I don't think utopia uh under the AI of like Elon's vision is is possible. Because look at him himself, right? He's I I do I think he's about the money no. I think he sees it as a tool because he wants to do interesting stuff, right? Like I think that's majority of money for him. But like look how much wealth he has sequestered to him's one person, right? A trillion dollars. What he can't do anything with that.

SPEAKER_02 41:27

You know what he means? And and like he's he keeps talking about how no one's gonna need to make money anymore, but he didn't he just negotiate a trillion dollar pay package for himself like a year ago? Yeah. It doesn't matter. It doesn't it doesn't matter, apparently. Um, and you know, I think the This feels a lot like these are not the droids you are looking for.

SPEAKER_01 41:45

Like everything's fine here. We're gonna take over the world, use let us, and we'll have all the things.

SPEAKER_02 41:50

It's a funny contrast to uh Dario from Anthropic who says that it's gonna destroy the world. There's like a you know a 50% chance it's gonna destroy the world in 10 years, which is a very weird thing for him to be saying. Um but yeah, I it really rubs me the wrong way. There's a lot of deflationistas uh in the AI camp who think that we're just like uh uh another model iteration away from all costs collapsing to zero. And I think that it's that is I think it's completely foolish. Yeah, I think it's very foolish. Uh could could it help to find m efficiencies in logistics and production processes and things? Sure. And I and I bet I bet it will help with some of that in some ways.

SPEAKER_01 42:32

So it's already been doing that for years. That's the thing, is these large language models are not all that's in the AI space. Machine learning models, nudge language processing models, different neural networks have been delivering value in industry for a very long time, right? And in ways that you wouldn't think of, right? So reducing the chatter on a cutter head in tooling, right, which creates, you know, a higher precision and a more efficient cutting process for tooling. Or one of the ones that I came across uh earlier in my career, you know, before AI was a thing, like you know, a term, uh, I was talking to the head of data science at Oxy. He talked about how they used machine learning to find the precise location of a of a drilling head for drilling oil wells. Because prior to that, essentially their their knowledge of the drill was actually kind of in this general vicinity, right? And so they were able to use machine and and create a more efficient process around that. So people have been using machine learning, and neural networks have been around since the 70s. We just lack compute to really do anything at scale with them. So I I mean, that's the thing, is this this isn't that new, right? And so to think that we're gonna the next iteration, right? We're already 50 years in on iterations of this technology that's gonna completely change the game. No. It will have to fundamentally it won't be a large language model if that happens. I can say that definitively.

SPEAKER_02 43:48

That's kind of what I think. Yeah, like whatever whatever the thing is that brings after large language models, right? Actual artificial intelligence, it's not gonna be an LLM. It's gonna be something that we don't know about right now. Maybe it exists. In a lab somewhere, but we don't know about it. And so the a couple of thoughts that I have is first of all, I think it doesn't make sense, uh, just kind of like on the surface, it doesn't make sense to assume that AI and robotics are going to suddenly overcome all physical resource constraints, which are in most ways the actual friction against us just producing shitloads more things and making prices collapse, right? It's actual constraints in the physical world, including human resources as well as physical commodity resources, things buried in the ground, et cetera. Yeah, now you're thinking like an economist. Right. Like no matter what it comes up with, it's not gonna suddenly help us to produce ten times more rice or ten times more apples or like pull copper and lithium out of the ground faster and easier necessarily. I think maybe uh in the future it could, like I said, help come up with some ways to make certain processes more efficient, or maybe it obviates some human middlemen in various supply chains that are added costs and maybe don't necessarily provide anything, and we can kind of link two ends of the chain together that weren't linked before and and remove some steps. And maybe that's how we reduce some costs. But in order for us to get to a place where it is overpowering the natural inflationary impulse of a fiat monetary system and to the point where all costs are collapsing, like we're talking about productivity that is three or four orders of magnitude higher than the long-term historical average. And like that is, I don't think people really have any grasp of just how difficult that is to achieve. Like we're talking about long-term productivity is like one and a half to two percent. And we're talking about needing like five to six percent year-on-year compounded productivity growth into the future at the same time as because this is assuming presumably going to obviate a lot of human labor demand, which is unlike a lot of other technological revolutions in the past where we've been able to just kind of shift them to other industries. The threat of AI is that it could just cut horizontally across a bunch of industries at the same time in the like cognitive layer of the economy, the white-collar layer of the economy. If we're just gonna obviate a lot of human labor demand, then we have to find some way to subsidize that, or else all the productivity gains just collapse because there's no demand to support them. And so like you can't just replace everyone with robots, have AI come up with ways to make everything, and then on the other end, we're just abundant and everyone's fine. Everyone can buy all of the things they need. The market delivers goods and services to people in the most efficient ways because the signals are all fucked up. And I and I know this is this is a big rabbit hole, but I just think like on the surface, it's economically illiterate to propose what Elon Musk proposes. And so I just always kind of default to, well, he's just a salesman, and so of course he wants you to buy a Tesla and invest in his robotics.

SPEAKER_01 47:00

Optimus Prime or whatever.

SPEAKER_02 47:02

Yeah, his Groc AI and all this stuff. Of course he does. Uh, but it it then it then starts like a snowball conversation where people actually believe we're gonna be able to deliver those things. And some of those people are C team, C suite people who run major companies who are now that pressure is now trickling down onto their workers, and it's it's not a good time for that.

SPEAKER_01 47:23

I want to agree with you. Uh so I want to say yes and, right? And so I just just look this up. Peak productivity growth during the Industrial Revolution, right? Very surprisingly, very gradual was 0.8%. So it's it's it's completely economically illiterate, as you say. However, let me explain, I think, the positive function that someone like Elon Musk deserves in an economy. It is less short-term and medium-term profitable to push the technological boundaries, right? Companies, especially in a mature market like an oligopoly or a monopoly, profit the most by repressing their innovation and melting each step of innovation along the way. Musk forces us past a lot of those bottlenecks just by sheer force of will and money. Um now, it's still very concentrated, but there is no satellite you know, internet without Elon Musk. Uh EVs to being mass-produced, right? These are none of the things that he invented. Probably doesn't happen without him. So while I think he is a salesman, I think he tends to uh be hyperbolic. And we there is value that we derive from having somebody who wants to push the boundaries so much who's something of a futurist. And so while we're giving him a hard time, I I do want to at least give that credit for that one thing because candidly that's something that was missing from our economy for a while, right? People would make one innovation and they would milk it and milk it and milk it and milk it and milk it.

SPEAKER_02 49:07

Yeah, I think you're right. And and I I I respect that response and also want to say I agree with it. I I don't have any problem with Elon, like as the person. I I think he I think it's good to have people who are trying to make their their dreams and their visions.

SPEAKER_01 49:23

He needs to stay in business though. Don't get involved in the government again. That was messy. Right.

SPEAKER_02 49:26

You know, keep him out of the keep him out of the government and and whenever he makes bombastic claims, just think about his incentive structure in saying those things and and just kind of couch it against that. Doesn't mean that I think that it's bad to want to envision a world where costs fall and we can have abundance. I think that's a wonderful thing to aspire to and dream of. But I don't know that we're currently on a path with our current technology that that is the deliverable at the end. That's the problem I have. And the way that it's presented, it builds people up with a false hope, I think. And it makes them think that they're not gonna need to work in the future. We are always gonna have to work. News flash people. You are always gonna have to work. That's never gonna be a human nature thing. Yeah, it's not never gonna not be a thing.

SPEAKER_01 50:07

Yeah. Yeah, I I I think that's right. I I think that's right. Um Yeah, I think so I think that kind of gives to kind of connect this back to our previous discussion of uh about kind of you know some of the math. So I I I think some of the previous math would need you know, Elon Musk's vision to uh you know half come true, 75% come true, to even break even. And and we just talked about how then in the industrial revolution we were getting 0.8% 0.8, 0.8% growth at its peak, right? Um how how is that possible? No, I it's not that it's not going to be that life-changing, not LLMs. So I just again there the math doesn't math.

Productivity Reality Check And Wrap

SPEAKER_02 50:51

Yeah. And and we it's currently not providing a productivity boom. We're in the investment boom. And when you re when you realize productivity gains from a new technology, it's usually after the investment and ramp up phase, after businesses have successfully implemented it into their processes, that later you start realizing the productivity gains. And so we are way, we are way too early in that process. And yeah, if you look at labor productivity, the way that it's measured, which is output versus per hour of labor, well, yeah, we are increasing capacity utilization. We're pushing our factories as much as we can go, at the same time that we have extremely low labor force growth and ex extremely low job growth. And so it looks like productivity is ticking up. But if you there's other measures of productivity, one called utilization adjusted total factor productivity, which really measures how much impact a technology is having on the means of production. And that is actually quite an unimpressive figure right now. It's a little low. So it doesn't even look like the AI has really seeped into the economy yet. We're still really in the early phase, and so we kind of need to give it several more years before we can give an honest assessment of its impact on the economy. Is it a job creator or a job?

SPEAKER_01 52:07

That's with every CEO telling every employee to use it. Right. Right. We don't have a PR problem here, people. It's not a lack of understanding. Well, it's a lack of understanding of how the tech works, but there's not a lack of people willing to try this, right? It's not like the people in the beginning of the internet saying, oh, this is a fat. No one's saying this is a fat.

SPEAKER_02 52:24

A lot of people don't know how their cell phone works either, but we still manage to get people to buy them. But like it's I think it's it's it's it's the the the the actual tech the delivery, the the stuff that you are delivered from the tech is different from what people tell you it's going to do. And I think that just that that gap is too far apart right now for people. So I think we should probably try to wrap up here. Uh we can each give a couple of final thoughts and then uh let our audience be on our way. How about that? Sounds great, ma'am. So I'll I'll put it to you. You got any final thoughts about this conversation? Anything you want to say before we go?

SPEAKER_01 52:58

I think this this the AI looks remarkably similar in 2030. If we haven't already had some type of correction in the markets due to it, I think that's probably like the upper bound of what it'll happen. Um I I think this is all coming back down to earth yet this decade. I I think this this bubble will pop. I I think the only thing that we have going for us is a lot of the debt is owned by some of the most profitable companies in the world. And so maybe they they can absorb it. I think it'll affect their growth, you know, ten years after. Um, but I don't think they're gonna go out of business. I think they're gonna end up having to eat a lot of the costs here. Um and the the company that's gonna make out the best, and I don't but I don't think it's for the long term. Uh well, there's probably a couple, but you know, NVIDIA, right? Uh and the other one is gonna be the memory companies, right? Like you look at the memory company financial results, they were two billion dollars in in in in net net profit in 2025, and this year they're 30, 40 billion, right? Their margins are just through the roof. None of this is sustainable in a commodity like like chip manufacturing, right? It's somebody's gonna increase capacity and supply will meet demand, and this will go back down again because we can't afford people can't afford to buy $1,000 Xboxes, right? I like I looked at getting wintered in my Xbox recently, the one I bought in 2022, and I can get more for it now than I paid for it. That's not sustainable. Yeah.

SPEAKER_02 54:17

No, and yeah, that the, you know, in 2021 there was a period of time where people were driving cars off the lot and then they could resell them for as much or more than they paid for them uh just because the economy was so backwards. Yeah, I think uh I think that the this is a very exciting time because it's uh it is a new technology and it's it feels like it's been a while since we had something really new and cool that people want to put their hands on and test the limits of. Uh but I also I'm very leery about uh how quickly and the scale at which we have put all the resources of modern society behind building this stuff out. Uh it's it's uh it makes me a little nervous for a lot of the reasons we already talked about here today. Um and I just kind of uh maybe it's because I'm a cynic uh just at heart, but I I kind of am looking for what is that WDM moment? When is that point at which the cost the math just gets completely undercut by something that is significantly cheaper or more efficient that kind of renders all the previous models uh irrelevant? And I I think as we get in the next year or two and we and the debt issuance gets into the four or five hundred billion dollars or possibly more, uh, that that is the moment when it is most sensitive for that to happen. And I just think maybe we've only got a couple years left of the AI bubble before it corrects back, and then after that we start to find out what the tech can actually do. That's kind of the history of these things, right? That's what happened with the railroads, that's what happened with the with the internet boom. We've had multiples of these cycles where we invest really heavily in something because we can see the potential. We don't know exactly what all of it is, but maybe we we can see some of it. We throw a bunch of money at it, it's way too much too fast. The people who initially invested lose a lot of money, and then the people who build from the ashes are the ones who actually like lead us into the next era. And I think that's the next step. But between here and there is probably a healthy correction across the space.

SPEAKER_01 56:09

Yeah, and the biggest risk, and we can talk about this another day, is the private debt market as it relates to financing a lot of this stuff. So I think that's a big unknown there.

SPEAKER_02 56:17

Well, Brett, it's been great talking to you, buddy. I always love having these conversations. If anyone's still listening to us, make sure you give us a comment in the in down below, let us know what you thought. Uh, hit the like button and uh subscribe to the channel, and we will probably talk to you guys about a week from now.

SPEAKER_01 56:33

Sounds good to XMC. See y'all.

SPEAKER_02 56:35

Bye, everybody.