The Capital Cycle Podcast - AI: Scaling Flaw

Episode Date: November 29, 2024

The risks behind the biggest gamble in business history. Presented by Edward Chancellor with Kai Chen, Analyst Emerging Markets.For more information, or to access select articles from Marathon’s Global Investment Review publications which accompany this podcast series, please visit www.thecapitalcycle.co.uk Hosted on Acast. See acast.com/privacy for more information.

Hello, this is Edward Chancellor with another episode of the Capital Cycle podcast. I've got Kai Chen, who's an emerging analyst with Marison Asit Management, and Kai has written a piece for the Global Investment Review about the subject that everyone's been writing pieces about this year, namely artificial intelligence, but Kai's got an interesting angle. As we all know, the advent of ChatGBTGT has sparked an unprecedented surge of interest surrounding generative AI. Do you want to talk a bit about what you're seeing? Yeah, so when ChatGBTGBT was first launched on November 22, it really captures the markets and public's imagination because it was the first time that we were able to interact with a machine that felt like human.

It was like talking to a real person on the other end of the conversation. And so this transformative experience has really led to the whole industry thinking that maybe artificial intelligence or AGI is within reach. So as a result, that led to an arm's ways to see who gets to AGI first. But you're saying that the market is missing some perhaps sort of investment fundamentals, let's say. Yeah. So I think you're referring to the lack of return on investment. I think that's often the bare argument for these AI models, because hundreds of billions of dollars have really gone into building the AI infrastructure,

but very few revenue has actually come out of it. And I would say a lot of this investment, if you look at the logic behind it, it's not completely irrational per se, because it's supported by this idea called a scaling law, which suggests that as you throw more and more computational power into training these models, then the model will just keep on getting better and better. And maybe at some point the model will reach human intelligence level. And I think it's this excitement that have really propelled the AI industry.

But obviously, this argument, there is some drawback in the sense that because it's basically looking at the value that AI is generating today and implicitly assumes that the model just won't get that much better in the future. And if we look at the development of these AI models in the past, they have been improving at an exponential rate. Very soon, if we extrapolate this trend, which is dangerous, I think. But if we do extrapolate, then these AI models can maybe so smart that they can start replacing human jobs. And each office job costs tens of thousands of dollars and some jobs costing a lot more. If AI is smart enough to replace a lot, lot of these jobs, then there's a lot of value that can be created.

So why might this be an element of wishful thinking? Now, you one sees many pieces about how these AI models are hallucinating, how they're making basic cognitive errors. Will you talk us through KSU site of the ARC Prize Challenge, which I hadn't heard about? Yeah, sure. Before we get into the ARC prize challenge, maybe it's, It's a good idea to take a step back and understand the limitations of these large language models. So all of these large language models are based on an architecture called the Transformer,

which was introduced by Google in 2017 in a paper called Attention is All You Need. And since that publication of the paper, the whole AI industry has really propelled in a meaningful way. The transformer architecture is very powerful, but it also suffers major drawback. And all of the transformer-based models, all the LLMs that we see today, Chachibb-T, Anthropic, and all of these models all suffer from the same drawback. And basically, you can think of these models as a giant pattern matching machine. They perform extremely well on tasks that they have seen before because they were able to do the pattern matching and then locate a memorized answer that is sitting deep inside the annual network.

and then basically give you the memorized answer or memorized solution. But the opposite side of that is that they tend to do extremely poorly on a problem that is not familiar because they are unable to do the pattern matching. They cannot find a memorize solution. This reminds me a bit of the quant risk models that were much in use before 2008 that hadn't seen a financial crisis for 30 years and therefore imagine there was no risk taking place. And how that's irrelevant. But I can see your point that if you're not.

you don't have the data and large amounts of data on which to draw from to find a pattern, then actually the machine is going to do less well than the individual. That's exactly right. This may sound really obvious in the sense that for a machine to acquire a new capability, it needs massive amount of data to train on that specific ability. So if it's something that the machine have never seen before, then they just don't have that massive amount of data. It's massive amount of data to do the pattern matching, to draw the conclusion from. And in a way, basically, when a chat GPD gives you an answer on a very complex problem, you might see really complex reasoning on that answer. But actually, they are not doing the real reasoning. They're

just doing the pattern matching to a memorize answer. Like undergraduates who are copying their essays of the internet. Exactly. So tell us a bit about the Ark Prize challenge, because that sort of shows this type of limitation that you're talking about. Yeah. So this year there's quite a lot of research, so just basically pointing to this particular limitation. And the art prize challenge is an interesting one because it's designed to be very straightforward for humans,

but very complex for the machines to solve. Essentially, you are given a few examples of how to solve a particular puzzle. Then you, as opposed to use that learning, use the same logical reasoning to solve it. a new piece of puzzle. Humans surprisingly have done really well. The average human score about more than 70%, but machine do terribly on this because they have never seen this novel problem before. And to be clear, the example you'll give is a primary school pattern recognition problem that I think even I could get. So let's talk a bit about the AI development course.

So, you know, marathon with concern about capital cycle and about when businesses ramp up CAP-X, that returns might be expected for everything else being equal. But what I hadn't, we all hear about these, you know, how big tech is scaling up its investments. But what I hadn't realized is what you're saying here is that actually the costs of developing new AI models is itself rising at a, at a, sort of super exponential, or least an exponential pace. So will you talk through that a tiny bit? So that topic is, I would say, there's a lot of different facets in there. One facet is exactly

what you mentioned. So each generation of the model costs 10x increase in terms of amount of money you have to put in. And also 10x increase in terms of the computational power. This is all based on this term that we referred to earlier called a scaling law. For each new generation, you just need an exponential increase in computational power. So just to put some numbers into perspective, GPT3 costs about $10 million to train. You start your first one because it didn't your, doesn't your first one only cost a few hundred dollars, your first? Yeah, so that's basically the transformer model that Google used back in 2017 and that only

cost $900. And then since then, since the publication of that research paper from Google, some really smart people, especially those from Open AI, realized that you can really scale this up by just throwing a lot more compute than the model capability would dramatically increase. And that's basically what the scaling law is. But what's interesting is that when people talk about the scaling law, they talk about computational power. You increase amount of compute, the model would improve. But in fact, actually, the original paper from Open AI suggests that there are actually three factors that drive the scaling law, not just computational power. So it includes computational power, size of the model, and amount of data. So maybe in the earlier years, for example, when you refer to the 2017 Google example, there are just an abundance of data that you can scrape from the internet.

but there was a lack of computational resource. So as you scale up computational resource at 10x for each generation, you get into this exponential return in terms of the model performance. But now you can argue we enter an era where the amount of compute has been increasing exponentially, but the amount of data hasn't increased in nearly the same way. So data kind of becomes a bottleneck. So if you take into that account of the three elements, you increase a lot computational power, but you're not increasing as much data.

Then it's natural to think that at some point we are going to get into diminishing return. And you refer to the earlier exponential increase in the model. So each generational model requires 10x increase in cost. So the next generation. And huge amounts of energy. Yeah, exactly. People talking about AI debt centers use as much. energy as Japan. And yet the heads of the big tech companies are locked in this AI arms race,

that they feel they can't lose. There's this sense that if one company gets decisively ahead and establishes a competitive advantage, other businesses they might encroach upon their business domains. And there's this comment attributed to Larry Page. Google, where he's supposedly saying that he'd preferred to the business burnt down that he should lose the race to artificial general intelligence. And that may not be an accurate comment, but there's clearly an element of this prisoner's dilemma, which we've seen in the past using capital cycle analysis

is a very, very common feature of, if you will, sort of negative capital cycles. huge KAPX spending, but everyone feeling they have to play the game because if they don't play the game, they stand to lose out. Yeah, so if you look at all these hyperscalers, all the big tech companies, one thing is clear, they are generating so much cash flow. They can afford to spend tens of hundreds of billions into building these models, but they just cannot afford to lose this war.

Let's say if Google becomes the first one to do, get to AGI and all the other ones just never get there, then AI, then Google will have this monopoly over AI and basically would be a huge threat to the business models of say meta or all the other big tech companies. So they just, it's a prisoner's dilemma you, you can afford to spend, but you cannot afford to lose. Except in those prisoners' dilemmas, the shareholders are the ones who often lose out. Now you're, but you're also suggesting that, actually the way AI is developing nowadays is that the product, these large language models are becoming commoditized. I think this is important because we have become so used to thinking

of big tech operating in its own, each in its own discrete area, Google and search, Facebook, in social networks and social media and Amazon retail cloud, and so on. And now we may be, be moving into an era in which actually the services they provide are perhaps not that differentiated and therefore don't earn such premium returns on investment. Yeah, that's right. So basically, if we look at the landscape today, there are actually quite a lot of players. And I was looking up, among the tier one players, you have OpenAI, Google, XAI, Anthropic and meta.

and for Amundatier, two players, you might say there's Ms. Charles, Microsoft, Amazon, and even Nvidia has recently joined the race of producing a very capable model of its own. And then I would say don't count Apple out. It's got a lot of resource, a lot of unique user data. So if you want to scale up, it can. And then on top of that, you have a lot of smaller players and also a lot of players from China, which actually have done exceedingly well, considering they are restricted from access. the latest Nvidia GPU chips.

So if you look at this industry structure, the commoditization is bound to happen. For example, OpenAI recently launched this model called 01, which is their latest model is called reasoning model. And even though the name OpenAI says it's supposedly open, but they are very close about how they do the O1 reasoning model. But just within like two months, there's another Chinese company called Deep Seek,

just came up with a very similar type of model and that have performed really, really well on various benchmark tests. This is basically telling you that the differentiations amount all of these models are not that great. And in fact, each model probably has a lead, the best model in the market probably has the lead of about three to six months before another more powerful comes along and overtake its place. So yeah, definitely seeing that commoditization, I think. So in the history of technology booms, I think where AI stands out is not in the amount of relative amount of capital that is being spent because you've had very capital-intensive technologies in the past. But in the fact that the proven usage for the technology is not quite there yet. The first great tech boom was railways.

By the time you had the railway booms, it was pretty clear. Railways moved people and goods from A to B. You had cars, so on. Even the internet boom, we could see at the time that the outlines of what the internet was going to deliver, although there were areas like social media that hadn't emerged at all at that time. So I think that this does make, you know, least should from the point of view of a skeptic, or intelligent investor, make one wary about the amount of money that's being thrown into AI at the moment. I always think in the technology world, in the cement conductor world, there's the Morse law,

and then in AI there is the scaling law. And at Marathon, we have the law of capital cycle, which suggests excessive investment is often followed by poor returns. And I think this is kind of what we are observing. There are huge amount of investment fueled by the idea that scaling law, will continue. And you refer to earlier, each of new generation of model requires 10x increase in terms of computational power. And I think that's where a lot of the risk is coming from. Because if you look at, say, the four hyperscalor plus Oracle, this year they are making a CAPEX of about $200 billion US dollars. Next year is estimated to be over 250 billion US dollars of CAPEX. And I think a lot of it is actually not going into training the next generation of models to be coming out.

They're actually, because you need the 10x increasing compute, you are actually preparing the infrastructure and preparing for the capacity to training the next, next generation of model. And all of this assumes the scale length law holds true and will continue to improve the models as you increase compute. But if the scaling law starts to slow down, then what we are going to end up with is massive amount. of overcapacity. And that's exactly where the marathon capital cycle comes in handy in the sense that excessive investment is followed by poor return as a result. Great. Thank you very much, Kai. Thank you, Edward. Thank you for your time today. I hope you will listen to the next edition of the capital cycle. This communication is provided for information purposes only. Please refer to Marathon's website

and the global investment reviews. for further information, including important disclosures.