View transcript
<p><strong>As with any investment, your capital may be at risk.</strong></p>
<p> </p>
<p><strong>Robert Natzler (RN):</strong> You'll have sat through quite a lot of AI presentations in your time. When I first made this presentation, there was a very active debate in the market last year as to whether large language models were here to stay or a bit of a spoof. And I think over the last five to six months, everyone's realised that this is a transformative technology. All the same, that hasn't liberated us from slides like this one, which predict that whilst we might be on the verge of world-changing superintelligence or of the collapse of the whole human economy, we're probably looking at trend GDP growth plus 2 percent. I'm not going to give you another version of this presentation. What I want to focus on in the half an hour we have together is the question about what artificial intelligence means for us in our industry.</p>
<p>And when I say us and our industry, I don't just mean Baillie Gifford. I mean everybody here in this room. Because regardless of whether your job is picking individual investments, managing portfolios, making high-level asset allocations or selecting managers, all of us are impacted by the changes large language models are beginning to make to what computers are capable of doing. If you think about what we all do in our lives, we map opportunities. We gather data on the opportunities we're most interested in. We synthesise that data into argumentation.</p>
<p>We analyse the worthiness of those arguments, trying them on for different sizes. We discuss them with colleagues whose points of view we respect. Off the back of that, we come to decisions. And then having made decisions, we have to manage the consequences of those decisions. In that sense, all of our roles are the same. And in that sense, all of our roles are equally impacted by the rise of artificial intelligence.</p>
<p>So in this presentation, my aim is to cover three broad beats. In the first beat, I want to quickly review the speed with which artificial intelligence is arriving in the world, because as much as it's discussed in the press, I still think people underestimate just how transformative this moment is. Secondly, I'm going to try and get us all onto the same page in terms of understanding this technology. It is all too simple when talking about artificial intelligence to fall back on analogies and buzzwords. And I think if we are to think calmly about what the technology means for our industry, we have to begin by actually understanding its capabilities, how it works and, by extension, its limitations. Once we've established a shared understanding, I'm then going to end in the third act by applying that understanding to our industry and pulling out a few areas where I believe there remains a key role for human beings in making investment decisions at any level of the investment industry stack.</p>
<p>Hopefully, that doesn't sound boring or repetitious from past conferences you've been to. If it does, you're welcome to heckle. In the meantime, this is always the chart that I like to start with. It shows annual grants in the US computer science field. And the point I want to make right at the start, it's not showing you the stock, it's showing you the flow. So that flat line you see running along the bottom isn't a constant level of patents, it's just a relatively constant rate of addition of new patents into the US patent pool.</p>
<p>And so you see there the IBM and personal computer eras. Computing advanced at a pretty steady rate year after year. This was then the internet boom. We joined together all of the different information silos of computers across the world for the first time ever in our civilisation. Engineers all around the world were able to collaborate and share ideas. And this was the result, a massive increase in the annual addition of computing patents into the US pool of patents.</p>
<p>You then see the majority of the last 20 years. Now that flat or slowly rising line doesn't seem very impressive, but what that actually represents is every year following the introduction of the internet, a roughly comparable number of new computing patents getting added to the system. This is the era where with mobile, we put computing into everyone's pockets. With the cloud, we made computing flexible, extensible and cheap. This was a revolutionary period. I think it's fair to say that at the beginning of this period, computing and tech was a dedicated sector in most people's asset allocations, and by the end of it, there was scarcely a business on Earth that had gone undisrupted by the computer.</p>
<p>That flat and slowly rising line was an incredible period to be a tech investor. I'm about to show you what's happened since the beginning of large language models arriving at scale. And before I do, I just want to remind you, this is not showing you stocks, this is showing you flows. This was the first 12 months of large language models. That is an equivalent uptick as it took the internet eight years to produce in 12 months. We have now, with artificial intelligence and large language models, delivered a tool that synthesises enormous amounts of complicated information from a wide range of different highly technical silos and makes them understandable faster than ever.</p>
<p>And we're going to be feeling the second-order effects of this transformative increase in innovation for decades to come. We can talk a little bit about the way that's already begun to manifest in the applications of AI in areas like coding. This chart shows that the length of time you can leave an AI unsupervised before needing to intervene doubles roughly every seven months. You already see in the way that computer engineers work, people working in four-hour shifts, sleeping for four hours and then working in four-hour shifts because Anthropic cuts off your tokens after four hours, and there's no point being awake when Anthropic's offline, and so people are adapting. We can also talk about how this has gone from being something that we're wondering whether there will be an ROI for businesses to seeing now over 50 percent of US businesses, according to data from Ramp, paying for AI and increasingly, forgive me for talking our own book, increasingly doing it from our portfolio company, Anthropic, which is the nice little uptick you're seeing right at the end of that. We can talk about all of that, and as growth investors, we can feel very positive about that.</p>
<p>There's a certain sense where Baillie Gifford, as an organisation, fundamentally invests in understanding long-term change. And if you think about the charts I've shown you, what they show you is we're at the beginning of a whole new period of radical technology change. That's fantastic for business disruption, that's fantastic for new business creation, and that's also fantastic for what it means for global productivity and therefore global wealth generation. And so there's consumer businesses like Hermès who stand to benefit from ever more surplus being created by these machines into the future. It's an exciting time to be alive and investing in growth markets. But it is also a terrifying time to be alive and to be a human, wondering what this means for you, particularly if, like me, you're a human who does a lot of these motions that the AI seems able to address.</p>
<p>We're often reassured that the future belongs to people who learn how to use AI. I don't know if you're familiar with this joke, but I sometimes wonder if I'm the farmer or the horse. I really hope I'm the horse. I'm sorry, I really hope I'm the farmer and not the horse. But I worry that I might be the horse, and I'll show you why. I pulled up the different beats of all of our jobs earlier.</p>
<p>And if you think about it going back 36 months, unquestionably, you would have wanted a highly educated credentialed analyst to perform each of these functions. Think just for a moment where we are today already. Today, I can get a large language model to scour the internet and produce a long list of potential opportunities. Today, I can get a large language model to call through APIs from any digital database I have access to, public or proprietary, data on those opportunities. I can ask the LLM to rank those opportunities based off the back of that data and recommend a shortlist for further work. I can ask it to produce arguments based off that data, suggesting what the future might hold.</p>
<p>I can even, and I recommend this to people who haven't tried it. It is quite fun. I can even get the AI to argue back with me about those points and do it in the style of any of my colleagues whose work it has also read. Create a little synthetic board of senior colleagues to quarrel with. So on the face of things, it looks like already, for myself, but also for all of us here, a large part of our job can already be done by the machine. And yet, and yet, we all know that these machines are not perfect.</p>
<p>Flaws remain. Specifically, I'm talking about what we call hallucinations. The machines appear to lie to us. They confidently tell us things that are not in the data, that are not true. That can even manifest in some cases as mendacious sycophancy. It makes it very hard to know where and how we should be deploying this technology in our process.</p>
<p>What I would like to do over the course of the next 10 minutes is explore how much of these hallucinations arise out of a feature of the underlying maths of AI, and how much represents merely a bug that we're going to be able to brush past in the course of the next 12 to 24 months. Now, to do that, I want to explore two different concepts with you. This is the concept of, on the one hand, rules-based reasoning machines that use symbolic logic to actually think and deduce answers, versus machines that are merely statistically predictive. It was mentioned in the introduction, one of my passions is history, so I hope it's not a surprise that I think the best way to introduce these topics and let us understand them and wield them in plain English with one another is to come at you through the lens of how these technologies evolved historically. So, with your forbearance for the next 10 minutes, we're going to speed-run the history of computing. And we're going to do it in the simplest English I know how to speak.</p>
<p>And I want to start just with some questions, because it's often surprising all the things that we think we know that actually we don't know about the history of AI and the history of computing. So, first question. Does anyone here know the first time an artificial intelligence beat doctors at medical diagnosis? You don't need to be shy. You do need to shout something out, otherwise we're not going to get to dinner. </p>
<p><strong>Audience member:</strong> 2018?</p>
<p><strong>RN:</strong> 2018, no. It was actually 1976. This is the Mycin machine at Stanford. Does anyone know the first time an artificial intelligence chatbot passed the Turing test? </p>
<p><strong>Audience member:</strong> [unclear]</p>
<p><strong>RN: </strong>It's a good guess, but it was actually 1966. This is Eliza at MIT. Finally, does anyone know the first time the world got excited about a computer being able to beat humans at chess?</p>
<p><strong>Audience member:</strong> 1990?</p>
<p><strong>RN:</strong> That's a very well-educated guess. It's also completely wrong. It was actually 1770. So this here is the Mechanical Turk. The way the Mechanical Turk worked in Vienna back in 1770, you got a very small human chess master and put them in a box. In my defence, I did ask what was the first time the world got excited, not what was the first time it was actually achieved.</p>
<p>The Mechanical Turk perplexed the biggest minds of the time. It even managed to beat Benjamin Franklin, who couldn't work out why it was this machine had beaten him. It's an important moment in the history of computers because it's the first time the idea goes viral that there might be a machine that can beat humans at cognitive tasks. And it so happened that even as this machine was making Benjamin Franklin sweat, the first actual computer was being manufactured. This is a Jacquard loom. It was being made in 1804 in the French silk-weaving city of Lyon.</p>
<p>Monsieur Jacquard, pictured here, created a machine that could automate the production of complex images onto fabric. When I say complex images, this portrait of him was actually created by his automated loom. Now, I bet you didn't think you were going to get a little 101 on 19th-century weaving, but the way weaving works is the image that appears on the fabric comes about by the way you lift and lower different threads running on the loom. And so if you control which threads get lifted and which threads get lowered, you control the image that emerges. Jacquard's genius was creating a single wooden board in which he punched holes. And as you pass that wooden board through the machine, the placement of the holes determined which threads would lift and which threads would stay lowered.</p>
<p>That system, that simple hole, no hole, one and zero binary, represented a transformation in our ability to build programmable machines. In the short term, Jacquard became extremely rich. He transformed the weaving industry for good. But in the longer term, he inspired an American. This is Mr Herman Hollerith. Mr Herman Hollerith, in 1890, took the Jacquard binary punch card system and used it to start recording information other than the patterns that we wanted to appear on cloth.</p>
<p>Specifically, he started to partner with the US Census Bureau in 1890 to automate the collection of census data. Now, like many inventors before and since, Herman was an absolutely terrible entrepreneur and quickly went bust. However, his patents were acquired out of bankruptcy by this gentleman, Tom Watson. He was running a private equity roll-up of business machinery. It was called International Business Machines, now known as IBM. And under his leadership, the Jacquard punch system, innovated by Hollerith in America, got distributed into offices all across the USA.</p>
<p>And in time, that mechanical system of recording information with zeros and ones from lifting and lowering of threads and later of pins, got replaced with a zero and one system recorded inside electronic relays. And the modern computer was born. So that got us to the point where we were able to use machines to record information. But to go a step further and conceive of thinking machines, it wasn't enough just to let a machine record information. You needed a machine that was able to logically deduce new information off the back of the information that you'd given it. Does anyone here know who this gentleman is?</p>
<p>So this gentleman is Claude Shannon. He's actually the person after whom Anthropic's Claude app is named. And Claude Shannon in the 1930s was the guy who worked out how you could make a machine think. He conceived of how we could use the electronic relays that were previously just used to record information to simulate the processes of human thought. Prior to Claude, the philosopher Boole had shown that you could simplify all human thought statements down to and, or, if and negation. Now what Claude Shannon did was show how you could reproduce those thought statements in electronic relay circuits.</p>
<p>That meant that you could build electronically a machine that was capable of replicating a human thought process. And that got people like the much more famous Alan Turing, who was essentially a journalist at the time, really, really excited about the potential of creating a machine that could actually think. Come the 1950s, the machinery is good enough to really build that machine, and the US government gets very excited. It's the Cold War, and so the idea of building a thinking machine that can outthink the enemy gets an awful lot of military funding. That takes us to Eliza in 1966. Eliza pretends to its users that it's a psychiatrist, and the majority of them believe they are actually speaking with a human psychiatrist.</p>
<p>That gets us in the 1970s to Mycin. So Mycin manages to outcompete junior doctors in diagnosing blood diseases and then in prescribing the correct antibiotics to deal with those diseases. And it also takes us to Carnegie Mellon's XCON machine. This was built for the hardware giant of the day, Digital Equipment Corporation, also known as DEC, and is actually the first machine that was ever made that would design the next machine. It is actually incredible for us to think that all the way back in the 1970s, the machine was already designing the machine. Given this history, you'd be forgiven for saying, “Hold your horses, Rob.”</p>
<p>Why are we getting so excited about AI today if we could do all of these things back in the 1970s? And the issue was that these projects ultimately proved unscalable. And the reason for that is the intrinsic complexity of our world. We live in a chaotic world, a world where any rule we define is prone to being changed down the line. So it wasn't enough for the engineering teams to write and define logical rules for their machines to follow. They had to constantly be updating those rules.</p>
<p>That takes time, that takes people, that takes money. DEC's XCON machine got to tens of thousands of procedural rules before eventually it became too expensive to maintain and the project shuttered. This fate applied to all of these different so-called expert systems machines. Ultimately, the US government gave up on the idea that a machine would be able to outcompete humans at thought. Funding got pulled. And if you've heard of the phrase in the AI industry of the first AI winter, well, that happened.</p>
<p>People backed off on the idea that AI was going to be possible. But it was in the depths of this winter that the approach we are now using began to develop. This gentleman is Geoffrey Hinton, shown in what was a very good haircut in the 1980s and his most fetching T-shirt. And he, as a grad student in the United States, developed a statistical technique called backpropagation. Now, if the words backpropagation ring bells, it's because that is still the technique we are using today inside large language models. What backpropagation does is abstract patterns from enormous amounts of data.</p>
<p>However, and this is an absolutely critical bit for everyone to grasp and understand, what backpropagation is not is a logical process. It's not a deductive process. It's not A, B to C. It's not solving for truth. It is simply, and I hate to say simply because it is amazing, but it is simply a statistical question of given X, what is the most likely thing that follows X? Given what we've seen in the pattern, what's the most likely next piece of the puzzle to fall into place?</p>
<p>Now as data sets got bigger, and as with the cloud, our computers got larger, we became able to run backpropagation at larger and larger scales. In the 1990s, it started getting deployed into consumer finance with more advanced credit-scoring technology. In the 2000s, it came to you in the form of autocomplete, whether that was in search bars or on your phones, given these letters, what are the most likely following set of letters to complete the word. In the 2010s, it would turn up in the form of voice recognition with devices like Alexa. In every case, it is the same simple statistical technique getting run, given these noises, given these patterns, what is the most likely completion of that pattern? In 2017, the now famous Transformer paper landed.</p>
<p>This was the paper that said our machines and our data sets were now large enough and impressive enough that we could run backpropagation, not just looking for single-pair probabilities, given these letters, what's the most likely word, but to map possibilities across the entire probability space. Given these words, what's the most likely sentence? Given these sentences, what's the most likely next paragraph? That is the process you see running inside large language models every time you log into ChatGPT, or as I hope you all do, Anthropic Claude. So, OK, that's been a bit of a speed run. I'm going to pause for breath.</p>
<p>The point that I'm trying to bring out here is that there is an enormous difference between the rules-based reasoning, logical, deductive, deterministic, truth-seeking of classical AI as conceived by Claude Shannon, and the statistically predictive, statistical autocomplete, if you like, of large language models today. And I hope that this can help us all understand that when we talk about hallucinations, we're not talking about a bug that is going to get ironed out as we make our large language models larger and larger and larger. We're talking about a fundamental part of the process. The AI is not lying to us. It's not misleading us because it doesn't have intent, because it is not trying to solve for truth. All it is doing is simply saying, given the words you've shown me, given the data I've ingested, this is the most likely completion for those words.</p>
<p>And that's profoundly dangerous. You can write into an AI, what are the impacts on global supply chains of the Gulf of Hormuz closing? And the AI will be able to give you a very compelling-sounding answer. But it does not understand what the Gulf of Hormuz is. It doesn't understand what oil companies are. It doesn't understand how supply chains work.</p>
<p>And it is not modelling any of those things. All it is doing is, from the enormous amount of writing it has ingested, giving you statistically the most probable conclusion for your question. And that really messes with our understanding of what thinking is. At Baillie Gifford, and I'm sure at all of your firms as well, you like to hire people who can think for themselves. And when you hire people who can think, the way you make sure they're really thinking is you force them to write. For that reason, as a civilisation, we've come to view writing as proof of thought.</p>
<p>Half the time, you don't even need to read the memo. The fact that someone you trust has written a long memo shows they've really thought about it, and if they still stand by the conclusion, then you're probably going to back them to some degree. What we have now built with statistics is a machine that can faultlessly write. And we have to be really careful about distinguishing where that's useful and where that's dangerous in our practices. Now, some people would use this, and I think this is way less common now than it was 12 months ago, but some people use this to dismiss AI outright. It's been called a stochastic parrot in some quarters, which is really, really unfair.</p>
<p>I actually think we all need to be experimenting with the tool because it is clearly wildly useful to have something that can effectively summarise the information that we have chosen to feed to it. And I think that experimentation is something we can't afford just to trust to our IT departments, who are using it to write enormous amounts of code, and it's very helpful for them there, too. It also needs to be something that those of us in the frontline roles in our organisations are doing, because it's only through experimenting with these tools that we're going to be able to understand for ourselves which parts of our jobs we can fully automate, which parts we can mostly automate but require a human fact-checker for, and which bits it's still most effective to do in the good, old-fashioned way. So as we move into the final part of this presentation, I want to take that hopefully clear understanding of the tool that we're dealing with. And just make a few examples of how I think we can apply it in investing. And hopefully it's not a surprise to anybody here that I'm going to use Baillie Gifford as my applied example because it is the investment firm I'm most familiar with out of everyone here.</p>
<p>I think there are three things that we can very clearly say about human beings in the process. The first is the human investor as an explorer. The second is the human investor as a judge. And the third is the human investor as a synthesiser, as an understander. I'm going to really rapidly charge through what I mean when I say each of those things. So first off, the human is an explorer.</p>
<p>Very simply put, the machine can only summarise data that's been fed into it. If the data has not been collected, if the conversation has not been had, if it's not been digitised, then the machine can't summarise it. And so in the clearest sense, even the most AI-pilled extremist would say, you need humans to go out into the world and be sort of fleshy sensors, gathering information and feeding it back into the central AI brain. Baillie Gifford, sitting on the private companies team, I have the privilege of seeing how that works in action. Long before 2012, Baillie Gifford Public Market Investors made a point of networking with private companies in order to understand the challenges to the public companies they held in portfolios. And since 2012, we have only built on that practice and on that access.</p>
<p>That's how, from a private market side, we're able to help our public market investors talk to the leading AI foundational labs and use that to understand which of the main semiconductor designers are getting the best traction. We're able to put people in touch with the local social media properties around the world to better understand how the global social media players are doing in those local markets. We're able to talk to the local logistics and automation people in private markets, again, to understand how the listed e-commerce players are adapting or failing to adapt to the new conditions of the world. Again and again and again, from my seat in private companies, I see how differentiated data can really feed into better public market decision-making. And I'm sure that everyone here will have their own analogy to this, the sources that they cultivate in order to get edge, whether it comes to asset allocation or to manager selection. The next bit then is the judge.</p>
<p>And when it comes to talking about the investor as the judge, it's important to consider for a moment how we all actually make conviction investment decisions. Does anyone here just sit at home in their office getting their juniors to write up reports that they read and make decisions solely off the back of that? I'm not even going to pretend to pause. That's a trick question. No one does, because you're all here. It is clearly wildly important not just to read information about the world, but actually to be in the room where it happens, to see where the enthusiasm is, to pick up the subtexts in human behaviour, to understand how people work.</p>
<p>You'll have heard your Baillie Gifford fund managers talk about culture, and that means something slightly different to every single investment team. But the commonality across Baillie Gifford is we really take the time to let the investor who's led the work explore with the team those slightly more tacit, hard-to-communicate parts of investment research. The bits which they just picked up about someone's vibe being off. The areas where they felt someone's actually much more enthusiastic about that than comes across in the written wording. These things are monumentally important, and because these things can only be picked up by the person who's on the ground, by the person who's actually in the room and seeing the people in the flesh, that means clearly humans can't just be reduced to being the sensors for the machine. They also need to be the processes, because only they can combine the information that is machine legible with the information that you needed to be in the room to fully understand and grok.</p>
<p>So that's the investor as the judge. Finally, we have the investor as the synthesiser. What do I mean by this? There is a huge difference between when an event happens in the world, whether it's in an individual company or it's at large in a technological field, feeling like you know what the most likely response is, and actually being able to simulate and understand that system in your own head. The best way to bring this to life is with a worked example. So this slide, it comes from our long-term global growth team, which has a very similar philosophy and approach to Plina and Lawrence's international concentrated growth approach.</p>
<p>And this is looking at their 10 biggest returners in the fund across holding period. And what we've done here is on the top, you can see the amount of upside that the fund gained by holding on to those companies. And on the downside, you can see the drawdowns that they had. And we discovered that across those top 10 contributors, more than 43 occasions, there were drawdowns of more than 30 percent. More than 30 percent drawdown. That is a horrifying thing to live through if you're running a concentrated fund.</p>
<p>When that happens, it's not enough just to be able to have the most probabilistic set of words describing what's likely happening. You need to, on a visceral level, be able to understand the new information and how it impacts the system that you have been studying. You need to be able to simulate that system, either on your machine or in your head, or ideally both, so that you can work out whether this is a moment where you have to sell before the thing goes to zero, or this is a chance to double down and generate fantastic returns for your client. This is what I mean when I talk about synthesis. It's the visceral understanding of the companies that you're holding that help you make the right decisions under pressure. And again, I don't see a way for AI to replace that part of the process.</p>
<p>So, explorer, judge and synthesiser. How do we think about doing those things at Baillie Gifford? And I'll speed through this bit. But we think about it in three ways. It's partly about where we are, it's partly about who we are, and it's partly about who we're becoming. Where we are, it's hopefully not a surprise to anyone at this point, we're based in Scotland.</p>
<p>That's not a place most financial firms choose to be. It's certainly a long way from the noisy chatter of the great financial centres of the world. That said, it is a fantastic place to sit in peace and reflect on the information that we've gathered from being on the road in those great clusters. We live and work in constant sight of the rivers, the mountains and the sea. That's not poetic analogy. I can literally see all three of those things from my desk.</p>
<p>And because the world loves castles and golf and whisky, we don't suffer for the lack of direct international flights to come and go with the tourists from the great centres of the world. We think if there's anywhere where you can sit at a distance and synthesise and reflect, Scotland is probably it. Secondly, who we are. We are an unlimited liability partnership. Unlimited liability for many of the folks in the room who are living it at Baillie Gifford. Those are two of the scariest words in the world.</p>
<p>They are terrifying. But what they do is not only drive alignment with our clients on the other end, they also force a certain sincerity when it comes to the way that we think about evolving. Baillie Gifford is never going to be the first firm rushing in to adopt a bold new idea that has a chance of really blowing up in our faces, but we are absolutely committed to doing the long-term correct thing properly, safely and seriously. That then shows up in who we're becoming, the way that we approach hiring. We're not just thinking about hiring the people we need in order to cover a certain amount of stocks over the next 12-month period. We're thinking about bringing people into the organisation who can be trusted to continue working in that organisation under conditions of unlimited liability.</p>
<p>That means people who can think long term. If you think about hiring, there is this old adage that you can choose to hire for intelligence, to hire for aggression, to hire for integrity. And maybe if you're good, you can get two of those three, but you certainly can't get all three because that's too good to be true. I think most financial firms around the world absolutely say we need intelligence. And I wouldn't say Baillie Gifford's any different. Many of my colleagues are some of the most intelligent people I know.</p>
<p>I don't want to pretend that that's where we differentiate. There's always going to be someone cleverer in a New York hedge fund. But they are really, really bright. So I think we're going toe-to-toe on intelligence. Most financial firms, however, will choose very deliberately to go overweight on aggression. They want teams that will fight tooth and nail for every single piece of short-term alpha.</p>
<p>And they will remunerate people aggressively based on that short-term competitive, which stocks did you add to the portfolio recently basis. And that's one of the reasons why the finance industry has relatively high employee churn and lots of people complaining about toxic workplace environments. Because the cost of going for intelligence and high aggression is you necessarily have to under-index on integrity. That's simply the way the trade-off works. With Baillie Gifford, I think they take the opposite approach. It's not as aggressive as some of the other places.</p>
<p>We don't go for people who are desperate to get rich quick over the next five years or 10 years. Sometimes talk about it as long-term greedy. We want people who really are genuinely motivated by the desire to beat the market through the cycle for decade after decade after decade. And because we go for those less aggressive, more thoughtful people, we're able to absolutely maximise the way we go after integrity. Because the truth is we have to. If you're running an unlimited liability firm, doing anything else is candidly disastrous.</p>
<p>I would put to you all that the result of this particular ingredient in hiring is a place that is incredibly well focused on the idea of developing strategic patience, on the idea of being able to look through what is merely short-term profitable and long-term unwise towards the places where the market is structurally wrong about the directions where the world is going. And I like to think that this differentiated approach to hiring isn't just something that served us well in the 1990s and the 2000s and the 2010s, but it's something that is going to continue to serve us well as we think about investing in the age of AI. We now have the tools and we're experimenting with the tools all across the floor for using AI to do everything from rapidly building us new vibe-coded tools to understand the world better through to simply using it to summarise information in more efficient and more effective ways. But what we're not getting rid of is what is fundamentally human about investing. That's exploration, that's understanding, that's genuine synthesis. Thank you.</p>
<h3>Risk Factors</h3>
<p>The views expressed should not be considered as advice or a recommendation to buy, sell or hold a particular investment. They reflect opinion and should not be taken as statements of fact nor should any reliance be placed on them when making investment decisions.</p>
<p>This communication was produced and approved in May 2026 and has not been updated subsequently. It represents views held at the time of writing and may not reflect current thinking.</p>
<p><strong>Potential for Profit and Loss </strong></p>
<p>All investment strategies have the potential for profit and loss, your or your clients’ capital may be at risk. Past performance is not a guide to future returns. </p>
<p>This communication contains information on investments which does not constitute independent research. Accordingly, it is not subject to the protections afforded to independent research, but is classified as advertising under Art 68 of the Financial Services Act (‘FinSA’) and Baillie Gifford and its staff may have dealt in the investments concerned.</p>
<p>All information is sourced from Baillie Gifford & Co and is current unless otherwise stated. </p>
<p>The images used in this communication are for illustrative purposes only.</p>
<p><strong>Important Information</strong></p>
<p>Baillie Gifford & Co and Baillie Gifford & Co Limited are authorised and regulated by the Financial Conduct Authority (FCA). Baillie Gifford & Co Limited is an Authorised Corporate Director of OEICs.</p>
<p>Baillie Gifford Overseas Limited provides investment management and advisory services to non-UK Professional/Institutional clients only. Baillie Gifford Overseas Limited is wholly owned by Baillie Gifford & Co. Baillie Gifford & Co and Baillie Gifford Overseas Limited are authorised and regulated by the FCA in the UK.</p>
<p>Persons resident or domiciled outside the UK should consult with their professional advisers as to whether they require any governmental or other consents in order to enable them to invest, and with their tax advisers for advice relevant to their own particular circumstances.</p>
<p><strong>Financial Intermediaries</strong></p>
<p>This communication is suitable for use of financial intermediaries. Financial intermediaries are solely responsible for any further distribution and Baillie Gifford takes no responsibility for the reliance on this document by any other person who did not receive this document directly from Baillie Gifford.</p>
<p> </p>
