Thank you, Michael. Great. Well, really good to see all of you here today and welcome to the panel discussion on artificial intelligence in investing. To Michael's introduction, we've already heard a couple of sessions allude to AI, and I think everybody in this room has probably not managed to escape competing the potential AI has to be transformative in varying degrees, whether it be from an investment perspective or looking at applications for your teams and your businesses. We're very excited to have the panel discussion today. I thought I'd start off just with a few points just to set the scene before introducing the panel we have here today. So we've heard a little bit about AI already this morning from our colleagues at MediaStats, particularly around interpreting language, but just to set the scene on why now is arguably a good time to have this conversation, if not before. Firstly, companies are laser-focused on AI. In fact, if you look at 2023, about 80 per cent of the Fortune 500 directly referenced artificial intelligence in their earnings calls, which represents a 50 per cent uplift from the prior year. Clearly, an area that firms are keen to avoid missing out on. State Street's also completed its own industry research and surveys on artificial intelligence. We asked more than 500 institutional investors the degree to which they implement AI techniques, and their vision on how it will be transformative in the future, and the industry overwhelming response there was really that it would help unlock competitive differentiation around their data assets and would be an area of further investment. In fact, a recent Stanford AI report showed that last year alone showed a record number of new AI models being developed. On top of that, a record exponential growth in AI-related patents. In fact, a 60 per cent rise again on the previous year and a more than 30-fold increase since 2010. So lots of really good promise on what AI could bring. Yet there do remain some questions and risks on the table. Again, the same survey showed that for global businesses, AI remains a key area for risk too, and in fact particular concerns around privacy, data security, and reliability still remain front of mind for global business. So I'd like to welcome the panel at this stage for the conversation. Hopefully, on the video screen, we'll see Aman Thind. Aman, can you hear us?

Hi there. Yes, I can.

Great. Hi there, Aman. Aman is joining us from the US today and is Executive Vice President for State Street and Global Chief Architect, where he is responsible for defining the strategy as well as driving adoption of potentially transformative technologies such as AI. Aman is also Chief Technology Officer of State Street Digital, which is a dedicated service and division of State Street targeting crypto technologies and web 3.0 technologies, and also serves as Chief Technology Officer for State Street Alpha, which is our industry-leading data management platform where you can manage your investments and businesses all in one place. I'd like to also welcome David and Marija to the stage. David Turkington is Senior Managing Director and our head of State Street Associates, our decades-long partnership with academia, dedicated to delivering research on markets and investment strategy. David has published more than 40 research articles spanning a number of different journals and serves on the editorial board of the 'Journal of Alternative Investments'. He's also the co-author of three books, one focussed on 'Asset Allocation: From Theory to Practice and Beyond', and another on 'Prediction Revisited: The Importance of Observation'. Many of Dave's research papers have indeed been award-winning. Then to my immediate left, we have Marija Veitmane, who is our Head of Equity Research at State Street Global Markets, and one, if not all of her prime responsibilities is focused on combining many of State Street's proprietary measures on investor behaviour, inflation, media, and market fundamentals, to form discretionary as well as quantitative models to support forecasts on market directions and regional preferences, sector preferences and the likes. I'm also well-informed that Marija has actually been one of the most bullish in the industry on the tech sector, particularly in the last five-to-ten years. So if that hasn't served you well in the past, maybe we'll come back to that during the panel. With that, let's start and perhaps Aman, I'd like to start with you on the screen, if that's okay, and get some of the vantage point you see in your role overseeing some of the industry trends for AI. Given some of the backdrop I've highlighted there, could you talk a little bit to why there's such extensive coverage on the AI theme in particular, now more than ever? What is it that's changed, perhaps more recently, really driving this renewed interest in the developments we're seeing?

So a couple of things have changed. One is compute. The amount of compute that is available now is unparalleled and continues to grow year on year. The chips that NVIDIA has launched with over two billion transistors, the ESG and environmental friendliness of the chips continues to grow. The ability to be able to stack them in data centres continues to be more and more available. So what that means is we now have more compute, but something else has changed as well. Of course, generative AI and ChatGPT becoming the fastest growing software in the history of software, what changed was in 2017, a couple of researchers from Google DeepMind published a paper called 'Attention Is All You Need', in which they introduce two concepts. One was the concept of self-attention, which is where you can understand through correlation, relationship between words, understand which words are important and understand how they relate to each other, which is the primary problem with the English language. I can say that yesterday I went to a restaurant and had a lasagne, and it was delicious. Just because of all the pre-training that we've had, we know that when we are talking about delicious, we are talking about food, lasagne is food, and through those classifications, we know that when I'm speaking about delicious, I'm speaking about the lasagne. But I could also say that I went to a restaurant, I had a lasagne, and it was beautiful, and now I would know that I'm actually talking about the interior decor. So the concept of self-attention has finally allowed us to actually speak to our data. Another concept was introduced, which is the concept of a transformer. So until then, we only relied on recurrent neural networks primarily, and they would only parse data in a single direction and that is how autocomplete was used. So in some ways, generative AI has been around for a long time because autocomplete and Google Translate has been around for a long time. What has changed is the concept of transformers. It allows you to tokenise words, transform them into numbers, and then process them together. What this means is now, instead of just parsing a sentence, word by word, left to right, you can process a paragraph, a book, or in the case of GPT, for 20 per cent of all knowledge available to mankind. So the availability of these technologies and the availability of compute, what that has allowed us to do is truly take another look at all the things that we've always wanted to achieve with AI and give it another stab. This time around, we are clearly much more successful than ever before.

Great. Thank you. In terms of, you know, clearly scalability and processing power to your points are all things that have evolved for some time. It seems to be that much of the recent buzz has been around evolutions around generative AI and more of the curated output that don't require human interference or human modelling. Could you talk to perhaps some of the perspective you see there around where you see the market getting most excited around some of the evolutions there in terms of where we've seen more recent advancements around, say, generative AI or other techniques?

Absolutely. So what generative AI has allowed us to do is to talk to our data, get research, get insights, and get value from our data finally. Historically we would say that we don't have enough data on something, but the reality was we would have too much data that we would pay too much for, but when we actually have to make our trading and portfolio decisions, we will actually end up just relying on our gut instinct. The gut instinct is actually just human. So being an IT guy, I actually have a better understanding of machines than I have of humans. So usually people would take AI and they will try to understand that as cognitive sciences, and how humans think, and trying to apply human behaviour on machines. I apply machine behaviour on humans. So everything that we learned in school was our programming. Then when we went to engineering degrees, that was our fine-tuning, and then everything that we've learned on the job and experience has been our pre-training with our data. The more data that we've had, the more data that we've seen. Data around COVID, data around the dot-com crash, data around the financial crisis, it has just improved our decision-making power. So effectively, humans have been operating like generative AI, wherein we've got all this training data through which we can actually give some results that, you know, with this kind of a market, these are the kinds of characteristics of stocks that we should be picking. Now because machines have begun to think like humans, wherein you can actually give it your characteristics, you can give it your context, and it can actually help do that for you, now, everything that we've been expecting from senior leaders within each of our organisations, now we could actually expect machines to do the same.

Very interesting. We'll come back to that question or comment around the interaction between humans and machines, and perhaps which entity are better understood. Perhaps one last question for you Aman, just building upon some of the more recent announcements. We've seen Apple and other companies join the list of companies looking to implement AI methods to really help to your point on personalisation and curating customer experience around different products and things like that. What would you highlight as maybe the one or two really more prominent trends that you're most excited about in terms of where AI could have a bigger potential across the wider market?

Yes, so hyper-personalisation is definitely one. We will see AI adoption across the board and that is because everyone has started to realise that it is not just important, but it is existential. If we look at the S&P 500, 52 per cent of the companies that existed in the year 2000, ceased to exist by 2020 because they were replaced by companies that used internet for reach, mobile for accessibility, cloud for compute and scalability. Over the next decade, who stays in the S&P 500 is going to be determined by companies that are agile, and nimble, and are able to personalise to their clients' needs. The only lever that we have now available is AI. So across the board we will see the adoption of AI. The biggest trend that we will see is the area of co-pilots wherein irrespective of what industry we are in and what we do, we will be able to do that job better, more scalable-y, and with much more resilience than ever before, just like it did for co-pilots. So co-pilots did not take away the job of the pilot, but it finally allowed the pilot to actually go to the bathroom without the plane taking a nosedive. That is what will be deployed on each of our machines and just to help us do our jobs better. That is the trend that we will see. It is truly exciting because it will allow all of us to be more productive and it is estimated that there will be $15 trillion that would be added to the overall GDP, just through the productivity that will be added through AI over the next few years.

Great. Thank you. A couple of other great points there which we'll come back to. Thank you, Aman. In fact, just to add to that, a recent survey from Mercer showed that actually 90 per cent of managers within the investment sphere, plan to either use currently or in the near future, AI tools in their investment research. So way beyond just, say, the quant groups that I think many started with. Perhaps, Dave, I can turn to you at this stage to zoom in on some of the investing use cases around AI and what you see there. So perhaps just comparing and contrasting how we're seeing AI being implemented within the investment process, are there any particular similarities or differences that you might highlight generally in terms of what you've seen in terms of client trends and how they're thinking about AI?

Yes. One thing that's important to keep in mind is that AI is a very blanket, broad term and encompasses many different things which can drive value in different ways to the investment process. I do think that broadly, AI and machine learning, if you want to use that term, can make us smarter. It can also probably make us faster. Those concepts may overlap sometimes; I think they don't necessarily overlap. Part of what I mean by smarter is can we make better predictions about the uncertain future. At the end of the day, that's what we need to do as investors. There are a bunch of things that we currently do. We can probably do more of it, cover more securities, make the decisions faster. That's some edge, but we can also make better use of the information that we already have. So is there alpha there? Is there actually the smarter component, is an interesting question. I think that financial markets and investing as a use case has some important differences versus the general use case for other AI applications. For one thing, markets are by design competitive, and that is a feature and not a bug, but it means that there's an extra challenge because in some sense, you always have a small data problem for the investment application, because you need to consider that the market, if there's an edge or an opportunity to be had, can start pricing that in and adapting to it, and therefore the data that existed when that was an opportunity is no longer valid in some sense. So the ability to process huge amounts of data is very powerful but we need to be humble about the relevance of the data and maybe it's in a short window. I think also because of the competitive nature of things, we have to be humble about our understanding of it, and we need to seek an actual, transparent understanding because just the fact that something worked in the past does not mean it will continue to work for this same adaptive reason. I think another difference is the tail risk that exists in markets when you're making financial bets. There's a certain penalty of being wrong when you're classifying an image, or of giving a non-eloquent description of something in text, or even an incorrect one, but you could lose all your gains and then some and go bust from a wrong bet in the markets. So there's this asymmetry in this risk profile that I think demands an understanding of the underlying drivers. We all have stakeholders, and very rarely do we not need to explain our motives to the next person down the line.

Yes. A great couple of points there. I mean, the asymmetry on the cost side of being wrong, going back to some of the earlier image generation you showed there with the typo, Michael, very different in an investment setting perhaps to others in the more creative industries and things like that for sure. Maybe just elaborating a little bit further on maybe some of the use cases that you've seen or been involved with where you think AI is working quite well, you know, really starting to show early promise in its applications. Could you maybe elaborate and share a couple of thoughts you had within that sphere, where the use case seems to be quite suitable for AI?

Yes, and I would say that having gone to a lot of quant investing type conferences for many years, even just five years ago, everyone was talking about AI. I think it was already a theme, but everyone admitted to not actually using it in their investment process and few were testing it. Today I think the picture is quite different. People seem to be picking their spots, and while they're certainly not turning over the keys to a machine, they are integrating genuine use cases. One of those is to analyse the text that is widely available as unstructured data out there, in the way that Ronnie showed this morning with analysing media-fed sentiment. Obviously, you need to be able to scale to do that, and the advent of large language models makes it possible to interpret text in that way. So that unlocks a whole level of understanding from jumbled utterances online that we didn't have before. Another aspect I think is more of a bottom-up approach where we do see that investors who did not pay any attention to data-driven inputs or quantitative techniques are often integrating them in some fashion, whether it's just a scorecard or something of that nature. But those who have used it for some time are going out the complexity curve or the sophistication curve to use machine learning and AI. It's very challenging to do that. I think whenever you're going to make models more complex, you have to justify that decision. So we see a lot of interest in applying models on actual structured data, whether it's random forest models, boosted machine, which is a tree-based decision process, neural network algorithms. Not necessarily large language models but coming up the curve of making better decisions. We've been involved in that for some time. I remember there was a breakfast that Michael and I had in Taipei, of all places, where we came up with the idea to create a publication every month that would predict currency returns from a machine learning approach, or actually three machine learning approaches, and compare and contrast that with your own brain and some brains of your colleagues. We called it a mind and machine approach. Really that the two can complement each other, and in fact, the combination of the two has outperformed either one independently in this case. I think that makes good sense. I would say that generally we are obsessed with interpretability, and I think it's a healthy obsession to have. Not everyone agrees with this. There are some people that think this is obvious. There's some people that take the complete opposite side of this argument. It is an interesting one philosophically, but we've found that we can learn a lot by deciphering what a model is saying, and we've come up with our own approaches to get a read on that, something called model fingerprint, which gives the logic of a model. An example would be even if you ask a person for their own logic, usually, there's a first-order effect, there's some subtlety around that, and then there's conditional if factors. So if you think about the carry trade, okay, on average I like the carry trade. Then I might say well what matters more is the direction of the carry and less the extent of it - this is actually what our machine learning models find and think. Then you might say, but there's conditional factors. If it's aligned with valuations, I like it more. If it's a quiet market instead of a turbulent market, I like it more, and I might actually flip the bet. So there's a little story. The machines can tell us that story if we ask for it. It may or may not align with what Michael and others think. That's been a healthy process. I would also add that we have actually found new relationships from using this kind of machine learning model that we didn't know about before, and I would rather probably claim that I've discovered some of these things, but in fact, we put variables into machine learning and found that there were durable effects that we could then go back and understand and document on their own right. An example is the power of equities to predict currency moves, it was not well-documented or relied upon, but was the shining example that came out through some of our machine learning models. Just to close out my thought on this, we've also built other models that are based on a whole different premise for predicting from data, which is to find which historical experiences are relevant to the task at hand, and then summoning them to predict that what happened back then might happen again. This approach is quite different than putting a model out there and saying, I have a neural network with a thousand or a million parameters, and I want to solve for those parameters and then I'm happy, to saying what are the precedents, which are the actual events that I should rely upon, and what level of conviction do I have if I scrutinise those events? We have this whole approach called relevance-based prediction, which starts from a different premise and leads to this transparent and adaptive alternative to machine learning. So I think we're at early days as far as coming up with more sophisticated, more complex models. Now, we know it took decades for people to adopt traditional quantitative techniques like linear regression models and portfolio optimisation, and there are still naysayers. I think it's not going to be instant that people adopt these techniques, but the more that we run these models, and look at them, and try to explain what they're doing, the more comfortable we get, the same way we get comfortable talking to a person about something.

Lots of great threads to that. Clearly, some good use cases of how AI is being implemented already. Certainly predictive modelling, complex relationship detection, all examples I think many of us would perhaps have an appetite for that in the room. Also, concerns around lack of transparency. Often black box concerns, types of behaviour I think can be ameliorated by many of these transparency methods, such as the model fingerprint. So it's great to hear that. I think you closed out Aman's point quite nicely there on the role of AI and how that works alongside human practitioners. I think it's an interesting use case with the mind and machine, which clearly has paid dividends to your point so far. So thank you for showing those. In fact, Marija, at this point I might come back to you and think about, well, we've shared some thoughts around the bigger picture trends on AI, what's happening within the investment market in particular. It'd be great to get your thoughts on what this means for the market valuation. Of course, US Tech being the obvious starting point and we heard a bit about whether US exceptionalism has peaked earlier this morning, but the tech sector seems to just keep on going. So given everything we've discussed so far today, what's your view on the tech sector at the moment based on where valuations are? Are you concerned yet? Are there markers you're looking at?

Well, probably no prizes for guessing that I'm going to champion tech again, but what really, really excites me about - I mean in general investing is… I think I was very fortunate a good 15, 20 years ago when some savvy investors told me always buy good companies that grow. That's a good starting point. Don't try to buy something cheap. Buy something good. I think tech really, really fits into that style. Michael showed the chart of US earnings so much exceeding earnings anywhere else. All of it is thanks to technology sector. What is really interesting, technology sector, kind of what Aman was talking about, continues to reinvent itself. With AI, what I feel increasingly, it's found the source of revenue that is almost independent of business cycle. So Michael's talk was about is business cycle peaking, is economy slowing, when VAT is going to cut, and that has profound implications pretty much on any sector. If you're a cyclical sector, yes of course economy slows that's bad for you. If you're defensive sector you probably have very low margins and margins are getting squeezed, can you pass cost to consumer? Technology sector is almost immune from that. Now we created this technology sector, created this infrastructure that big software companies want to train those very advanced models. They need to have the best chips. They need the best chip designers. You need to get those machines that build chips. So it's kind of infrastructure, but then you have independent input, what Aman was saying again, is that it's transforming every business, it's not just technology. If you're not investing into technology, into AI, you're very likely that your competitors are, so you're potentially lagging behind. So almost any company needs to invest. I know lots of us have been throwing around statistics on IT adoption. What piqued my interest in our Q1 earnings reports, the most commonly mentioned term was AI. Think about that's CEOs reporting on their earnings. It's not profits, not margins, not clients - maybe that's because it was mostly bad news - but it was about AI, how excited they are, how transformative. I think it's still fair to say that most CEOs outside technology sector, probably in technology sector as well, have not yet figured out how to make money of AI. What they have figured out is that if they don't invest, they'll be left behind. So this investment comes to technology sectors. That's a massive inflow of revenue, as I said, almost at the time when we're concerned about the economy slowing. So that's massively helping technology sector. So their profitability is likely to be a lot more stable, a lot easier for them to defend their profits when overall economy is slowing. So that's a really, really strong supportive point. I mean, we can go back to valuations positioning. Yes, tech stocks are not cheap. You don't need a PhD in anything to figure that. Valuations are comparable to what they were during financial crisis. Michael showed us charts that institutional investors holdings are quite crowded in tech sector. Again that's the biggest position, I think overall equity holdings are the highest since GFC. I think tech sector holdings have only been higher during TMT bubble, so we can go back in another decade. So yes, investors are excited. Investors are positioned for it. It's crowded and expensive, but it does give you something special when lots of other sectors not giving. So yes, we are susceptible to potential shocks. If you remember back about a months ago when NVIDIA was reporting earnings, all Bloomberg wires was, I mean, people were watching it as if it's a macro event. Which is kind of a bit silly because NVIDIA earnings are somebody else's CAPEX, can't be macro. Market is obsessed with it. So there are some volatility points, but as long as those earnings are coming through as we see them coming through, as adoption of AI continues, excitement about AI exist, it supports. One thing that I'm really excited about - of course, tech is the best-performing sector over the last however many years - what is interesting recently is that if you say last year, all of those gains came from multiple expansion. So people getting more excited about tech. What we're seeing now, about half of the gains come from multiples. People still excited, but the other half is from actual earnings growth. That's really different from any other sector. Actually S&P Ex-Tech have seen earnings decline. S&P Intech have seen earnings growth. That's really why I'm still very much excited about it.

Very good. Clearly a big disconnect in tech to the rest of the market to your point. Obviously, it'd be wrong to assume that all of the opportunity for AI exists within tech to your point. There's clearly lots of value creation to be had in other sectors of the market or other countries. Where do you see the biggest opportunities lying for those excited about investing and getting exposure to AI outside of, say, US Tech? Where would you start?

Well, that's a lot trickier. I think the idea is AI is complicated and I'm sure Aman can talk a lot about it. It takes a lot of investment. It takes a lot of understanding of the process. So there are some sectors, some companies, sectors, parts of the market that are far ahead of that curve that have been investing in AI. I guess all of us have experienced the surged pricing of Uber during strikes or dynamic pricing of airline tickets during school holidays and things like that. So AI has existed for some time. Some sectors, some companies have already invested into them. They are far ahead. That's generally not true about every market. I think an interesting example recently, market got really, really excited about electric utilities because we know that demand for electricity is going to skyrocket because we need to power up all those data centres, and servers, and it's huge, and potential is almost limitless, and people got really, really excited. Utility stock performed very strongly. Then actually, if you peel one step back, yes, there is strong demand, but actually, have those utility companies invested into infrastructure? Actually, no. So yes, there is a huge potential, but you need to do a lot of investments. I think that's really the big difference between winners and losers. Who has invested, who has put up money there, who is ahead of the game. It's almost like a race, so if you're being ahead of the race, then there are a lot more potential. So interestingly, and we probably can go back to Aman is that financial company is probably one of the few where there are still quite a lot of inefficiency, but companies have a lot of money, a lot of capital, they're really focused on it, and investing quite a lot.

Great. Thank you, Marija, and in fact, Ronnie from MediaStats, earlier in the session on narratives, talked about how they can use NLP to measure narratives. I asked for an update before this session on the AI narrative, and although, as he mentioned in his session, it has come down in terms of media interest globally on AI in terms of its peak being mid-last year, the utility sector itself has had a resurgence in interest on AI. So that definitely backs up that point. Following Marija, Aman I might actually do that and come back to you. It would be great to get your perspective on what we're doing at State Street around AI. What are our ambitions within the space, generally speaking? If you could perhaps elaborate on what this could mean for our clients and the wider industries benefits.

Absolutely. So at State Street, our ambition is to infuse AI into our DNA across the firm. What that means is, instead of looking at specific applications of AI or specific use cases, we are building foundational capabilities that will allow us to build and deliver whatever we need to do, faster and cheaper. So for that, we've picked out four different key areas. One being document intelligence. So there is a lot of value that is locked away in unstructured data in the form of PDFs or screenshots. What we are trying to do is come up with a lab through which we can parse, process, and analyse any document of any type. The problem is actually quite hard. The minute your data gets structured as forms, you need to start doing multiple parses. Sometimes you will realise that a normal document parser is not as good as an OCR when your data is structured in a PDF, so you end up taking screenshots and then running those screenshots through OCR and then reprocessing them. So we've come up with a lab which is parsing and processing lots of data. We've started with private markets, K-1s, fund prospectuses, and trying to do things like risk and controls within those documents. So if you hold multiple funds and those funds all hold one same underlying asset, so making sure that it is being priced consistently, are the kinds of things that we are doing. The other thing that we are focused on is anomaly detection. That is an extremely important piece, especially within AI, because AI is garbage in, garbage out. The only difference is, in this case, if it is ChatGPT that has read more books than we all combined have any chance of reading in our lifetime, means whatever it will say, it will be said with such verbosity that you would assume that to be the truth. So making sure that your data quality is pristine through anomaly detection is extremely important for us. We actually even won a research award from CA100+ last year for that research. The reason why anomaly detection is so important to us is because it is a gift that keeps on giving. Once you are really good with anomaly detection, if you show it how to find bad data, it will find bad data. So it will help you find bad market data. It will help you find bad transaction information. But then you can feed it, for example, samples from money laundering. That could be real-world examples or examples that you could create through synthetic data generation. Then the same anomaly detection algorithm becomes a great AML algorithm. So then co-pilots, being able to talk to our data, creating an NLP layer across the board at State Street such that no matter what business you do with the firm, being able to ask a question and get a response directly is extremely important to us. Then finally, client onboarding. So a lot of errors stem from just our inability to match, for example, the schema in which our clients expect the data versus the schema in which we store the data. For example, you could have around count of 2000 in one and a blob in another, and as long as your blob is less than 2000 characters, it will continue to work fine. The minute it is more, all the ETL tools, they will still just simply concatenate, and you will find problems when you are actually trying to parse and process that data. So being able to come up with automated matching and reconciliation of schemas that allow us to onboard clients in a more resilient manner is the fourth capability. So documents, NLP, data quality, and better schema matching and onboarding are the foundational capabilities that we're working on.

Excellent. Thank you. Well, we did promise variety and we did promise optimism. So there's a few flavours across the panel. So thank you all for those contributions. At this stage I'd just like to ask the audience, please feel free to ask questions. We can do that again raising a hand and we can get the mic over. For any of you using Slido, I've got the tablet here, so I'll see those come through. We'd love to open it up to the floor. We've got one here.

In Aman's opening comments, he made the point about computing. We've heard a lot about investing in computing. To what extent is this an arms race for people to develop computing power? To what extent do you need unbelievable computing power in order to be able to do the things that you've been talking about here? Is it a case that the companies that succeed will be those that invest the most?

I mean, looking at market performance, I think that's definitely market belief. We see very substantial outperformance of large-cap stocks versus small-cap stocks. A lot of it boils down to investments. Can you invest? Can you invest more? I think the interesting parallel is what happened with the financial sector, maybe five, ten years ago when there was a big threat of fintech. Can fintech be more nimble, more exciting, more faster moving? It started to win market share from banks but then banks figured out that actually we are bigger, we have more capital, we can invest more, we can do everything fintech can do better, faster, more efficiently. I think now we don't talk about fintech very much and I think that seemed to be the technology. I mean, it's probably fair that you need to ask right questions, not just to have the massive computing power, but seems like those things come hand in hand. So I think at this stage there is a quite a strong belief that you do need to invest more and that obviously comes to bottom line of a lot of technology companies.

When it comes to investing in compute, I think we should leave it to the people who are investing in compute. So the way compute and GPU will go is the same way cloud went. So if you look at the early 2000s, every bank was trying to create their own cloud. There was always these instances of private cloud. Then it became private public cloud. Then all of it became public cloud, and people are shutting down their data centres, and people are selling all their servers back to Amazon and leasing them back. That is the same thing that will happen to compute as well. I don't think we should be in the business of building large language models from scratch, which is where you need all of this compute, because that would be almost like I need a data analyst, and for that, I decided to give birth to a baby so that he can grow up to be a data analyst. I'll just hire a data analyst.

Great. Thank you.

So that's where it would lie. So there are companies which are spending billions of dollars in compute, their companies that are making GPUs available at scale. At the moment I'm in San Francisco at the Databricks conference. There was the Snowflake conference last week, and all of them are making GPU available on demand and at scale. So I don't think we should be purchasing GPUs. In order to build these large language models, we should be focused on fine-tuning them for our own needs and then subsequently inferencing. All of that is available on public cloud and that is what we hope to simply use.

Great. Super. I think we have another question at the end of the room there.

Yes, hi. If you think about the history of innovation in financial markets, it originates typically in the US. The innovation cycle becomes very strong. It draws outsized investor attention. That means excess capital flows a very strong dollar. It grows up and as the innovation cycle matures, you then have diffusion of the technology. The benefits start to spread out across the globe. Productivity gains start to catch up in the rest of the world, capital starts to flow out, the dollar weakens, and so forth. What I want to ask all three of you a little bit conceptually, if you look at AI, and we're hearing kind of mixed messages here about the power of compute, the centrality of cloud GPU processing, is there anything in here that changes that normal dynamic, where actually the bulk of the gains, whether that's competitiveness, productivity and so forth, actually stay concentrated in the centres of innovation, particularly the Anglosphere and perhaps China, and don't diffuse out to the rest of the world in the same way? It's a question. I don't have an answer, but Marija you sit from a financial angle. Aman, I'd like to hear from a tech angle, whether you think there's the same dynamic or a different one at play here.

Sure. Aman, did you want to start with that one?

It is definitely the same dynamic wherein… Because everything in the US is outsized, right? So there is outsized investment in the space, and there is money really being thrown at the problem. There are companies which just have decks right now that are turning into unicorns. Sam Altman had tweeted that the world of a one-person unicorn is now near using all the technologies that are at our disposal. I think, there will be a few things that will determine the winners and losers and the overall diffusion of this technology across the board. One would be regulations. There will be some areas which will almost regulate these technologies out, and they will then subsequently suffer and be playing catch up later. The second thing is the overall ubiquity. So even countries that don't really manufacture mobile phones still use mobile phones. So what will happen is eventually all the AI technologies, they'll become commodities, they'll become McDonald's, and they'll be available everywhere. It's just that there will be companies that had the first mover advantage will be taking most of the benefit, not as the manufacturers of these technologies, but these technologies will be available for use across the board, depending upon the regulatory regime that runs those countries.

I know we're a bit over time, but I'll just add one comment to that which gets to your question too. If you look at the AI-related start-ups over the last ten years, about 5500 of those are in the US, around 1500 in China, and more than 700 or so here in the UK. But if you were to flip that and look at AI-related patents that have been granted over the same period, 60 per cent of those reside in China, 20 per cent in the US, and less than 2 per cent in the UK and Europe combined. So your metrics will definitely give you a slightly different picture there, but it definitely seems to be centred around the same centres from at least those perspectives. Great. Well, we're out of time, but thank you so much for the questions and please join me in thanking Aman, Dave, and Marija for an excellent panel. Thank you.