Thank you so much for having me today to talk to you about the technology space, which you've described as euphoric. I hope I can hold back my enthusiasm, because the risk is that I'll end up supporting that idea. Actually, I don't think we're in a euphoric phase for tech investing and for the AI theme yet, but this has all the hallmarks that we might find ourselves in that euphoric phase before very long. So let me, with no further ado—well, it's worrying—move on to who we are very quickly. If you don't know Polar Capital, we are a, I don't know, the term is boutique, independent, smaller, brilliant fund management group, based well, all around the world now. But we were formed by technology investors a long time ago, one of which used to run the trust that I now run, Guy Russell. That was a long time ago. I've now been at Polar for 22 years.
I've run the trust since 2006, but I don't do it alone. We're one of the largest tech teams in the world. We're 11 people based in London, and all we do is look at tech. I'm slightly being unfair; we actually have an AI fund as well, which has roughly half of the fund in technology stocks and half of the fund in companies outside of the technology world that we think will benefit from AI, companies like Walmart, Slack, whatever it might be. We can maybe talk about some of those later. But the key here is that, well, we think you need a very large team in order for us to really tackle a sector that has only grown in importance over the years. And you know, again, tech being a very dominant theme or part of the market. One of the reasons why the US, the US stock market has been so good over the years is because of how central tech is to the US stock market, unlike, say, the UK, where there's very little tech.
You only have to make a few good decisions in your life. I was asked if I wanted to be a financials analyst in the mid-90s or a tech analyst in the mid-90s. I chose tech. Thank goodness for that, because I've been there for Apple and Nvidia and for a whole bunch of things. Anyway, this is the team that supports Nick and I, who run the team at Polar. And as I say, we run about, what does that add up to, about 13 billion dollars of dedicated technology money. The way that we invest, I haven't got an awful lot of time obviously today to cover this, so ask me questions afterwards or over drinks. But we are focused really on growth stocks, and we try to build portfolios that are diversified. And the reason for that, you've heard already, is in the end trying to not fall foul of hubris is one of the things I've learned over the 20 odd years that I've been doing this. And you know, nothing like a bear market to realize how little you know. And so I've tried to build portfolios with the avoidance of hubris, which means in the end, for me at least, focusing on companies that grow, well capitalized, and I'll have exposure to the most important themes, the stocks and themes that if you tapped me on the shoulder 10 years later and said, why didn't you own Apple, you silly sod, I would be able to—I just don't want to be feeling that I missed a stock like Apple. Same applies for Tesla, for AI, so on. And it's really served us very well.
Most recently, we've changed or added one thing to our investment process. We now think through, or we consider, this idea of an AI lens, which really is a metaphor for saying that we think that AI is so important that all of the various technology themes that we've been excited about really for the last 20 years are now almost second to this, what we think is the next general-purpose technology, which is AI. And so that's been critical. We moved a lot of the portfolio over the last 18 months following the launch of ChatGPT. Again, I won't spend too much time on this, but the key here is that professional investors are good at selling things. We know when things have gone wrong and we exit. And honestly, most people don't talk about sell disciplines in this world. We tend to come on stages and tell you about how we bought Apple on the 1st of May 2003, which by the way I did. But we don't spend very much time talking about the mistakes we make. And one of the critical things that professional investors do is recycle their losses, put them in the top of the funnel, and find new winners. That's really one of the most important reasons why we exist.
So this is our investment process. We try to avoid really long duration, we call it, but ultimately the sort of binary outcomes that people are excited about when they go horse racing. This is investment; we're not in the horse racing business. We're trying to invest in companies where we have a good idea of a range of outcomes, not yes or no, 50/50. We're not interested in those stocks; we leave those to other people, usually outside of listed equity markets. We focus on real companies with growth that have exposure to key themes. Now, the most important theme beyond any in the tech market right now, of course, is AI. And in time, this will become the most important theme across all markets, because in the end, this is like steel or like electricity, like the most important things that have happened to humankind. So of course, it starts with the people that build the rail, the steel mills, and the factories in the industrialization periods. But today, it's really about tech because we are the epicenter of it. And as a result of that, we have moved very materially to get as much exposure as we can to the theme, not just in companies that sell semiconductor chips or cloud computing companies, but also beneficiaries. We'll talk about some of those. But this is really one of the most important takeaways from my conversation tonight.
You can't see the quote, which is a shame. That's what you know. But anyway, the quote was something along the lines of, you know, as important as fire or electricity or some other superlative. So again, we're going to struggle here, aren't we? But the idea of this slide was supposed to show you that for most of human history, people lived in buildings that were about five stories tall. And then Mr. Bessemer came up with a new way of making steel. And post Mr. Bessemer's breakthrough, we started building structures with steel frames, steel frame structures, rather than just building with bricks or with cement. And what's so fascinating about this parallel is that had you been alive in 1885, after the Bessemer process, after the first skyscraper was built, you might have been underwhelmed. The Home Insurance Building of Chicago was basically about, I think, six or seven stories, or maybe eight, and it looked very much like the building next door. But then, of course, what happens next is an explosion in building height, and very quickly we transform the landscape. Actually, cities are very different today, of course, than where they were in 1885. The reason for this obviously very rapid change of the landscape of New York, the reason for this is ultimately to try and provide you with a parallel for what we think is happening with AI right now.
So the history of AI, very horrible, horrible slide. The history of AI is that it started in 1956 when six blokes got together over a summer and thought they could solve AI in a matter of weeks. Of course, they didn't. But over time, we can see this as a log scale, there has been improvement in the complexity of AI systems. And it was following a linear chart, wasn't it? And then you can see very recently, and we've just literally off the chart, this is OpenAI's ChatGPT-4, and it's rumored to have one trillion parameters. And that's of course why ChatGPT is as phenomenal as it is, because, well, because lots of things. But the point we're trying to make here is nothing happens, nothing happens, and then it all happens in five minutes. And the best of the moment is the transformer model, which is a horrible chart. What happens in 2017, Google comes up with a new idea of how to build AI models, and it's built on something called a transformer. And these transformers are basically able to take more information than ever and be able to compute against it very rapidly. And then something miraculous happens, which is that these AI models stop being—they start to have emergent qualities. And someone gave me the parallel: imagine when you're talking to a child and you're trying to teach them language, and then at some point there is a spark where there's understanding. Now, again, I don't want to be too ahead of my skis here, but the transformers that have enabled, are the foundation for ChatGPT and other large language models today, have some—not sentience—but they understand, they comprehend, and they're moving very fast. And as a result of these transformers, you can see this chart here. People in my industry talk about inflection points. Before the transformer, AI models were basically getting eight times better every two years. Now they're getting 275 times better every two years. And what that means is—and just to be clear, this is like 1885. The people alive at the time couldn't understand the significance of what was happening about them. The AI models are happening and improving so fast that it's very, very hard to keep up with them. And again, I will try and bring that to life in a minute.
What does this really mean? Well, it means firstly that this is a really big deal. This isn't, I don't know, headsets that you're going to put on and maybe wander around in the metaverse or whatever. It isn't about that. This is about a technology that has the potential to impact very large markets already. So think about internet advertising. I think Meta was talking about how using AI has enabled them to get 6% more time spent on Instagram, I think it was the stat just today. AI is more engaging, gets people to stay, and of course they get to monetize it. These markets are very big, and the incremental opportunities to AI—imagine the healthcare market. If AI could penetrate a market that's worth 10 trillion, these are massive market opportunities, which of course is why companies like Meta and others are investing as aggressively as they are. But it gets much cleverer than that, which is that, and I'm sure if you've been using ChatGPT and other AI models, you can already begin to feel the risk it poses to human work. And now again, this won't happen overnight, and I'm very bullish on what actually AI will do for humankind and for creation of new industries and new jobs. But let's park that for a second. My industry is a $5 trillion industry, four, five, six, whatever it might be, ultimately that's the size of it spending. But if you think about the global wage bill, it's nearly 10x, and we're going to start moving from that world to that world. And one example of this is Klarna, which many of you might know. It's a buy now, pay later company that basically decided to get rid of 700 people working in call centers that cost them $2 to $3 million, and that is now saving them $40 million a year. It turns out the people on the end of a telephone don't know anything, and actually AI models that have access to all the catalogs of all the information you've ever needed actually know more. And it turns out that the deflection rates are really high, and people are much happier talking to an AI agent than they are actually talking to someone who doesn't know what they're talking about. I can't remember the last conversation I had—we've all been there. So this is why this matters. This is unique to AI: the ability to transcend a capex item into an item that ultimately can arbitrage—not today, but beginning to arbitrage global labor.
I won't do this slide, but I wanted to focus on this. History tells us that we can see very significant gains from new tools, and obviously people know about horsepower with agriculture and tractors. Obviously, and again, I haven't really got time to go into it. My favorite is the sewing machine. The sewing machine transforms our relationship with clothing. People used to—in fact, quick story: my grandfather, when I first came to the city in the '90s, gave me a shirt of his, and it had detachable collars and cuffs. And he told me the backstory and all of that stuff. It was horrible and very uncomfortable. I chucked it in the dustbin. But the point was that before the sewing machine, people did have a shirt for the week and a shirt for Sunday for church or wherever you were going on a Sunday. The sewing machine changes that. We end up basically reducing the time—I think it's something like 14 hours to make a shirt went to one hour, 22 times output. But people went on strike. People were upset about it. Seamstresses went on strike in America. But actually, now the apparel market, clothing globally, is a $2 trillion industry made possible by the sewing machine and obviously technology that followed.
This again, not the prettiest chart. This is a company, Axon. They make tasers, which isn't a lot of fun if you get, you know, tasered, but it's better than being shot with a bullet. I've owned this stock for, I'm going to say, 10 years. It's been one of my best investments for reasons that actually have changed now, which is that the company has body cameras. You know, you see them everywhere, don't you? You see them at railway stations, you see them, you know, anywhere. Those cameras are picking up audio and video signals. These large language models are so clever that they can now take the audio content and turn it into a police report. Police reports take 30 to 40% of police time in America. They're the number one reason why police officers retire. It's not because they're scared of being shot; it's because they're scared of dying of boredom at the end of the day writing reports. So to think about this in a very positive light, this is why we're bullish on what AI means for humankind, because it means—my brother's a teacher, he, you know, six weeks off for summer. He's like, 'I haven't got six weeks off, I've got to do marking.' Why, why, why, why are people doing marking? Doesn't make any sense. Why are police officers writing reports? And these technologies, this AI, is now able to take that audio content alone and basically do 80% of the police report in six minutes. Again, very early application. We're going to see a lot more of them.
If you haven't seen these videos, these are mind-boggling. This, I don't think we have time to really show it, but video: type in a text prompt into a model called Sora from OpenAI, and it will produce you a Hollywood-style video. Be very nervous if you're an extra; there won't be any jobs for you in the future if you do that on the side. Sorry about that. And so on. This is the video that's prompted by saying, you know, 'lady walking down street in Tokyo, wet, blah blah blah.' This is all produced by machines with no human input. Again, look it up online. Sora is the name of the model. I think it means endless or something in Japanese, or 'the sky's the limit' or something. But anyway, this is Figure AI. I can't—unfortunately there's no audio, I don't think, but again, go have a look at Figure AI. This is the idea of embodied AI. If the machine can see and recognize what's on the table, it can actually—I mean, anyway, the optionality associated with this is very significant. There's lots more to come. We are—this is the worst AI will ever look, and it's only going to get better. And so we have new chips coming from Nvidia, we have new models coming from OpenAI, Perplexity, and others, and we're moving—I've just come back from a conference where the hottest topic is agentic AI. The idea that, you know, I've got 19 things I've got to do today, wouldn't it be great if I had a, I don't know, a little monkey on my shoulder that could do that work for me? Well, you will—you will be called an agentic AI, and it will run through what it is you need to do. You've got to pay this, the gas bill, find new insurance—that's what an agentic AI will look like. It will be acting on our behalf.
I haven't got time to talk about capex, but I think what I—I ran out of time already, I was just firing. But what I would say is that the opportunity is very quickly—we're building new infrastructure for AI. AI can't be done with cloud computing infrastructure; it needs new compute. And the CEO of Nvidia—Nvidia is 12.5% of our portfolio, biggest position, has been for a long time—has said a $1 trillion opportunity to build new infrastructure, because in the end, AI is hugely performant. It's like a Formula One car compared to the cloud, which is unfortunately a bit like a Toyota Prius. Sort of can do everything, but none of it very well, unless you're a driver in PR. And I haven't got time to go through it, but some of the names, some of the ways we're playing in the portfolio—the key here is, for the first time, we have found really exciting opportunities in Asia, and not just in Nvidia, which everybody knows, and not just in TSMC, that everybody knows, but in a whole bunch of companies that do other stuff that's really important in a Formula One car. So before compute, and then maybe I'll just finish on this very quickly, which is again a horrible chart, but the point here is that this is called scaling laws, and I will leave it on scaling laws, I promise. So the really exciting thing is, like, why are these companies—why are they spending $150 billion a year on building out AI capex? Why would they do that? 'It's a bubble, it's a bubble, it's all crazy.' The reason that they're doing it is because of scaling laws. And scaling laws are something that OpenAI talked about in 2020. Have a look online if you're interested. But the idea basically says—and this is crazy—which is that at the moment we're still in a position where scaling laws are holding. Scaling laws say that as I add more compute and more data and more parameters—parameters would be like the moving part in an engine, words maybe in text, but not—and pixels in an image—as I add more of those things, it turns out that the AI models are getting better. And so if you go back 18 months ago, people said, 'chat rubbish, makes up stuff,' and it doesn't make up stuff now. The hallucinations are much lower than they were before. We now have models that can reason. We have models—if you haven't tried OpenAI's o1 model, you haven't lived. That's an exaggeration. But reasoning models. Anyway, the point here is that as long as these scaling laws hold, people are going to invest, because in the end, if I have an AI model that is 99.99% accurate, I'm well through human levels of capability.
So again, I've run out of time, apologize for that. I think there's some crazy stuff happening that I have every chance to talk about, but I think that the key takeaway is that the path that we're on, the proof points that we see, the model progression that's happening, chips that are coming, says that the chance of us being alive—this is crazy—the chance of us being alive at the point where computers are more capable than human beings is very real. And in that scenario, going back to my tap on the shoulder test, you will feel sick if you didn't have exposure to that. I will feel sick if I haven't delivered you great returns. Now, of course, there's going to be bumps in the road and all that, for sure. But this idea—the timeline just 18 months ago, people thought that there was a 50% chance of us getting to that point by 2050, and the gains that we've made to the technology in the last 18 months have brought that forward by 20 years. I think with that, I should shut up. There were about a thousand more slides here that I didn't have time to talk about, but hopefully we can answer some of them in Q&A, and I'll just click through. Thank you very much.
