Prepare for a deep dive into the intricate world of artificial intelligence with our esteemed guest Prag Sharma. As the Global Head of Artificial Intelligence at Citi's AI Centre of Excellence, Prag's forefront insights into the rapid integration of AI into the finance sector are truly riveting. Watch as Prag and host, Jim Jockle of Numerix, unpack the implications of AI in finance, from strategy to ethical usage, and the urgency for an expert ecosystem.
Prepare for a deep dive into the intricate world of artificial intelligence with our esteemed guest Prag Sharma. As the Global Head of Artificial Intelligence at Citi's AI Centre of Excellence, Prag's forefront insights into the rapid integration of AI into the finance sector are truly riveting. Watch as Prag and host, Jim Jockle of Numerix, unpack the implications of AI in finance, from strategy to ethical usage, and the urgency for an expert ecosystem.
Welcome to Trading Tomorrow, navigating trends in capital markets. I'm your host, jim Jockel. In my decade plus of working with Numeric's global leader in capital markets risk management technology, I have launched our Thought Leadership Division, a place where insights, innovation and expertise converge, just like this podcast. Through my journey in the financial realm, I've had the privilege of witnessing firsthand how the capital markets landscape has transformed. The complex dance of market trends and innovative technology has redefined how the finance industry operates. With game changing innovations just around the corner, we now stand at a crossroads, one where it is more crucial than ever to understand the interplay between these realms. That's what we do here. We talk about current and future processes and technologies you need to be aware of moving forward AI. It's the topic of the year. While the concept has been around since the 1950s, 2023 was truly the technology's breakout year, mostly because of the explosion in popularity of chat GPT. A recent McKinsey survey found a third of those they spoke with said their organization is using generative AI regularly at least one business function. The survey also found that many of those who aren't using it yet plan to explore using it in the future. It's a technology that has permeated every industry, including finance.
Speaker 1:Joining me to discuss AI and finance further is Prague Shama. He's the global head of artificial intelligence at cities AI Center of Excellence, which was created to provide trusted leadership and consistent guidance on AI across the entire organization. The center of excellence handles all of the organization's processes affected by AI, from client deliverables to internal processes. Prague works to set standards in policy and develops best practices, while ensuring that city always considers the impact of AI on its overall strategy. Welcome Prague, thank you for joining us today. I'm delighted to be here, james. So, prague, perhaps you can start off just by telling us about your journey with AI and what made you want to learn about it and work with this technology.
Speaker 2:I'd love to start off by saying that I knew this was going to happen, that this was coming in the future, and this is why, back when I was in college, I got really interested in AI. But the truth is obviously more nuanced and less exciting than that. When I was in college, I was very interested in stats, maths and computer science, and machine learning, and AI sort of brought all these different fields of interest for me together. Plus, I was always interested in pushing the boundaries of various things, including the computer that computers that we use. Machine learning and AI is a really amazing way of getting the maximum out of the technology that we have in hand, in this case computers to do things or think like humans themselves.
Speaker 2:So I went to college, did engineering. What followed was an interest in machine learning and AI. I was lucky enough to do a PhD in machine learning and computer vision back in the day, when it was not the coolest field on the planet like it is today. So I was very interested in image processing, looked at how we could use image processing to detect objects in not only still images but in video sequences as well. That piqued my interest. I then stayed on in this space to do data analytics, machine learning and, more recently, generate AI, the hottest single technology that I have come across in my entire career in the technology space and in financial services.
Speaker 1:I have degrees in politics. That wasn't very popular back then either, nor is it today.
Speaker 2:You may still get there. I mean, AI might help me get more popular.
Speaker 1:You know it's interesting on AI fraud. Apparently now there's new disclosures within the US that any political campaigns as it relates to the upcoming presidential election are going to have to disclose their use of AI. So you know, very timely topic, rightly so, I think.
Speaker 2:we're all prepping for that for sure, and not only in politics, but in many other areas. I think disclosure would be a good idea, given that they're so realistic nowadays.
Speaker 1:Oh, it's unbelievable. But you know to that point is, you know the technology itself for AI and machine learning. It's moving and changing so fast, so quickly. So you know for yourself. You know as well as your team. You know how do you keep up with the speed of growth, adaptation and innovation in this?
Speaker 2:space. Simply put, we do not keep up with it. There's too much going on to keep up with everything. But the idea here is not to know about everything, but to focus on the aspects of artificial intelligence or or generative AI, which is a subcomponent of that. To focus on the things that matter to you specifically for your job or the industry that you work in, and really spend some quality time getting under the hood, understanding what the technology does, but also understanding the use cases, so a number of ways to keep up to speed with, depending where you are in your life cycle, in your career life cycle.
Speaker 2:So I make sure that my team has a mix of technologists, business people, design thinking experts, project managers and other experts, in ethical considerations or otherwise. The younger generation that's coming out of college is pretty good at programming and know the latest technologies and algorithms, so stay close to them and make sure that we use their skill sets in that space. In the middle and senior management, my role is to ensure that I educate and provide thought leadership to the right folks at the right level so that we're all aligned with what this technology cannot do. And then talk to our second and third line of defense, our legal people, our internal audit, our regulatory folks, compliance risk, information security, cybersecurity. So the idea is to have an ecosystem, talk to the right people and then continue to keep up to speed by reading as much as possible and discussing the different viewpoints. It certainly isn't enough to have a whole bunch of data scientists in your team and think you're on top of everything.
Speaker 1:You know, I think it's interesting that the technologies come out of the 1950s. Clearly you and I were not popular at college drinking parties, sorry you're still not popular, but anyway, except this light, Fair enough, I'm in derivatives. No one wants to talk to me. Did you expect public interest in AI to kind of explode the way it has in this?
Speaker 2:past year. That's a good question, because nobody had expected public interest to explode like it did this year. But there was a clever ploy here. There's a really good reasoning why it exploded Usually and let's talk about OpenAI's chat GPT, which is what you mentioned. The reason we keep saying chat GPT is because it is the most popular tool and it has become synonymous with all large language models, Whereas we know that there are many different types. Bard from Google is one. Many other language models exist, but chat GPT has just become that term, like a Hoover, for example, for all vacuum cleaners. So chat GPT has become the Hoover for all large language models.
Speaker 2:Chat GPT became popular because OpenAI unknowingly or maybe knowingly, released it to the general public as a starting point. Before that, large language models weren't directly released to the public. They were released as research papers. They were released as technical sort of documentation's open source code, so not many people had access to it. I was reading up on them very excited to use some of them in my work for demo purposes. But as soon as the BICA became available to the audience at large, the general public, it exploded because it is so versatile, from writing poems in the style of Shakespeare all the way to explaining technical details, comparing regulatory documents. I mean, it's endless what these larger language models can do, and this is why it's become popular. Notice that the underlying details of the algorithms are still not made public in some cases. In others, they are being made public, so this is why I think it became popular.
Speaker 2:Machine learning in AI has always had value chains. It has always been used in industries, including ourselves and including in city. We've been using machine learning, but it just wasn't as cool. Quote unquote, if you can see me. It wasn't as cool and now it has become cool.
Speaker 1:Looking at where you are now in terms of the work that you're doing. What are some of the biggest roadblocks that your team is currently tackling in terms of AI and specific to finance at this moment?
Speaker 2:It is a super hot topic. Everyone from our CEO down is interested. Senior management is very interested. We are very interested, which is a bonus, which is a plus. This is amazing. We do want that kind of interest because that unblocks a lot of things that would have traditionally been blockers for large organizations like ourselves funding budget resources. But in the case of Generative AI, we see so much value in this that some of these things are melting away because of the potential of this technology. So that's the good part.
Speaker 2:But your question was around what are the blockers at this space? The blockers are this technology is emerging quite fast and we want to make sure that when we utilize this, we utilize it with the right risks and controls in place, and that's not so clear-cut at the moment. The reason for that is these are huge, huge models, massive models. So think about them, as I don't know what the equivalent would be If the same as if you were to picture a model like this in my space, it would be those computers from back in the 50s 60s that filled out rooms. That's how big these models are behind the scene, thousands of GPUs working in tandem.
Speaker 2:So to explain the inner workings is a little difficult and if we can't explain them to a level that we're comfortable with, especially for high-risk applications? If you're making loans to individuals, anything that affects an individual's capacity to operate freely in this world has to be taken carefully, and we need to be able to explain our models. So some of the issues are around explaining how these things work, the biases that are inherent in these models and, especially when you talk about Generative AI, trying to explain away the hallucinations or at least detect them so that we can act accordingly. The reason these are difficult is not for us, but this is an evolving field.
Speaker 2:In the research area itself. There is a lot of research going on to ensure that we reduce these biases that are inherent in the data, the models or otherwise. But what I can tell you is we are trending in the right direction. Lots of areas of focus now on how to use these models ethically, soundly, reliably, consistently across the various use cases that occur in financial services. So I'm very confident that the blockers that we have today that is, ensuring that we ethically use this, that we monitor these we will overcome these. We will start to have robust risk and controls in place and then off we go, utilizing this to improve the lives of our clients, customers, internally, ourselves, with robust risk and controls in place.
Speaker 1:In terms of managing things like bias, and many financial institutions have had particular challenges in those regards. Will this ever be autonomous in terms of things like loans or something of that nature? Will there always be an element of human intervention in terms of harnessing this type of technology and decision making?
Speaker 2:Certainly for us, in the medium to short term, there are definitely going to be humans in the loop. The terminology is humans in the loop, humans on the loop, humans outside the loop. So certainly in the short term, and especially for high risk applications, there are likely to be humans in the loop. But I think it's a fool's errand trying to predict the future. James. You already yourself were saying that this stuff is moving fast and it's moving fast and improving quickly.
Speaker 2:I think the way to look at it is for high risk applications, we certainly need robust oversight, with humans in the loop or any other mechanism that gives us a level of comfort that we need as an organization and this industry needs and, more generally, even outside financial services we need to ensure. But for low risk applications, applications such as finding a sentiment in an email you received, categorizing incoming emails into buckets, helping me write an email giving me a suggestion for a paragraph or a next word that I'm reviewing these are low risk, depending on what you do. Obviously, if your whole business is writing really elegant emails, then that's a core component of your business, but in most cases they're low risk and for those we may or may not need a human in the loop and I think a lot of the regulation is also looking at it like this Risk based approaches, where you treat models high risk models differently for low to medium risk models. I think that's the future, combined with better tools to explain these things.
Speaker 1:You know that's fascinating the way you're thinking about high risk versus low risk, and I was at a lecture a few years ago and the gentleman was talking about autonomous cars and his thesis was we could have autonomous cars today, but we as humans weren't ready for it. So things like lane departure warnings and lights in terms of sensors of what's around you. We're training us to get comfortable with that technology and I think you know that's just a very interesting parallel. In terms of low risk, I wouldn't think of you know finishing my email for me or if we're throwing in a paragraph, but that's a really interesting concept for I think, for people to think about this, especially around AI. So thank you for sharing that. You know, do you think it's important that all companies have a team like yourself you know to explore AI's and you know what benefits you know is city deriving just by having you think about this all day.
Speaker 2:Listen, the self-serving answer here is of course you need people like me and viewer. You should be paying them really well because they're really hard to get. That would be the self-serving answer I agree Absolutely.
Speaker 2:Please make sure this stays in the final cut. But, joking apart, I think it depends you know what you want to use machine learning for. If it is a core component of your business that is likely to give you a competitive edge, then you definitely should have data scientist in your team, because that is the differentiator. But if you're simply consuming a technology it is not your core product then maybe you don't need as many data scientists. You still need data scientists. You still need people who understand what's happening under the hood, ultimately because you need to know what technology you're using. As a parallel, you didn't ask me whether you need database administrators today. Do you need cybersecurity experts in your organization? Because we understand for large organizations, yes, no organization, even small ones, don't necessarily function nowadays without a cyber expert, without a database administrator, without a project manager. So I think a data scientist is very much a core component of any business today A data scientist or a data science team to help you increasingly understand what's happening under the hood and navigate the world of machine learning and data science that you will increasingly be using as an example. I bet you that this recording session probably uses some sort of AI in the background to maybe look at the video, enhance some other aspects of it. We use it every day. Your phone has so many things that are AI related. I think understanding that would be key, and data scientists or experts in the field will help you do that as an organization.
Speaker 2:So for Citi, we are very much interested in enhancing our internal processes from a risk and control perspective. Operational enablement large organizations have many manual touch points. How can we eliminate those and or use machine learning, AI or other technologies in fact, to streamline those processes? How can we create more value for our clients using the latest technologies, including machine learning and AI? And the way Citi benefits from having a center of excellence that cuts across the organization is that we centrally set standards to ensure that we are getting the most out of the technology within our risk and controls environment. We want to ensure that we understand how to use these technologies across the organization. We build out a community lots of thought leadership.
Speaker 2:I talked to our clients directly. I was just at an amazing conference yesterday where we got Really good feedback on people just trying to understand how to apply AI to their organization or to their particular business. I talked to third party vendors and startups that are offering state of the art solutions. You need a data scientist to understand what's being offered under the hood. Is it simply everyone labeling AI and not doing enough of it, or where is the value add? So me and my team look into that as well. Many, many aspects of this, and we need to. Everybody needs to get stuck in.
Speaker 1:We talked to, obviously in the beginning of the podcast and the popularity around chat GPT. But there has been, especially within financial services where there is so much proprietary data. Right, chat GPT it's public, it's harnessing the power of the internet, but how are financial institutions looking at AI specific as released to the proprietary data versus these, an open AI type solution?
Speaker 2:That's a great question. I'm surprising the word data has been mentioned so late in this conversation. Usually pops right at the start and you're right to mention it. Without discussing data, we would have only partially addressed the world of AI. There's a famous quote in this. I was recently also presenting on this quote. In God we trust all else must bring data, and that's a brilliant quote. And then there was another equivalent quote is if you tease, if you torture the data enough, it will tell you anything.
Speaker 2:So, basically, there's two aspects of this. One is that thank you for your opinion, but we need data and in fact, with Generative AI, your proprietary data has become more important, not less, even though that these algorithms have been trained on nearly the whole of the internet or a curated version of the internet. We can all get access to these via APIs, but the value add, going beyond the standard services that everybody will utilize, the value add will come from your know-how as an individual, which we can discuss about. Whether your jobs will stay or not, that's another quick question. To have your expertise as an individual combined with the data that your organization has captured over time, this is going to become the differentiator in your quest to develop better products and services that don't just rely on general data but rely on the data your organization has captured over time. So your key motivation here, our motivation, is how can we now exploit that data with the latest technologies to offer better services for ourselves and for our clients? So data still happens to be the key.
Speaker 2:That the point that I was making tortured the data enough, it would confess to anything would tell you anything. The idea is we need to be careful. Huge amounts of different types of data means correlations, not causations, and it's hard to understand the correlations and causations when you're using such large amounts of data. So you need even more expertise in data analytics and the business side of things to really get the most out of it. Otherwise, you're running down a rabbit hole. Number of Nobel laureates in Switzerland and the amount of chocolate eaten is highly correlated but has nothing to do with it. So there's no such example like that, where you don't want to get caught up in causations and correlation that have nothing to do with each other. So we need more expertise. We need more data.
Speaker 1:You did mention it briefly and I think it's the one question that is on everybody's mind, and there was an amazing study in terms of the London School of Economics recently looking at jobs that would be changed or no longer necessary because of robotics, ai, things of that nature. I myself have prompt engineer as a skill on my resume. What's your position on the future of the workforce and the use of AI?
Speaker 2:Firstly, we need to all have prompt engineering in our job description. The sooner the better. So anybody listening in, please do a course online and get that prompt engineering going. Joking part, the reason actually you need prompt engineering is like I was explaining. You have this room-sized model, quote, unquote, and you're interacting with it. The clearer you are in your interaction, the better results you're going to get. And prompt engineering is simply that. How clearly can you articulate your ask so that you get the right response? In fact, we need that in our daily lives, james. If you can't articulate what you're looking for to another person, you're going to mess up and have misunderstanding. How many misunderstandings are down to language? Nearly everything. So prompt engineering in general in life, but also specifically for these models, definitely needed. Now to the second part of your question jobs.
Speaker 2:Jobs have always been changing right. There was a famous study that I don't know, the, in the 1980s. 50% of the jobs in the US Don't exist anymore, or there are newer jobs now with newer titles. If we were doing this back in the 80s, you wouldn't be talking to the global head of AI as an example, but that job did not exist. Or if it existed, nobody cared. So the point is jobs always change and we will continue to change. The issue isn't that your job might get obsolete because of this. The point we should focus on is how quickly will this change happen? Because that's really where the fear is right. The fear is, I'll come in tomorrow and, oh my god, chat, gpt, we have started doing my job. I do not think so. That is not how it works. It has never worked like that in the past and it's unlikely to do that. Jobs will change over time. It is the responsibility for our governments, our regulators, large organizations like ourselves in society in general to prepare us for what's coming down the line. As an example, how many of us go to university now as compared to maybe 50 years ago? Many more people do, because there is a benefit to go into university. Similarly, we need to see, we need to look at universities and look at courses, and computer science could have been, could be done earlier, in primary school or in secondary school. We need to tweak ourselves a little bit and get ready for how machines will help us more in the future. So I'm and so I certainly don't think this is the case.
Speaker 2:My final point in this is this is very much an assistive technology. Very much so, and I think we mentioned this right through our conversation so far. Your know-how is your know-how the machine has captured. That machine relies on it, in fact, to give you the right responses. Without you it is a statistical machine churning out random stuff that may or may not be useful to anybody, and so it's your know-how and your experience, your common sense. We consume so much more data than these machines. The internet is nothing. Our experience is stemmed from way more than the fit and the internet the physical world, our relationships with others.
Speaker 2:Not everything can be captured in language, james. If it could be delighted. Actually, many things can be captured in language, as GPT for and GPT 3.5 turbo shows us, but not everything. So my, my take away from this is learn how to use these models. Your know-how is the differentiating factor here. It is what you know, have learned over time that will get the most out of these models, not the other way around. You will be working with these models to get the most out of them. That Inside, that you provide and ask the right questions to these models, is what will give your organization yourself the edge up when you compete in this Hope that makes sense. Oh, if you agree with absolutely.
Speaker 1:I mean, you know it was. It was funny. It made me think back to one of my college friends, and he was, he was Greek and he, his household, they spoke English and and Greek and and at. Sometimes, when we were sitting dinner with his family whatever you always pick a you'd say something in Greek, you know. And I said, well, why are you switching back and forth between English and Greek? And he said, well, because there's a better word for it. And I said, well, that's fascinating, right, so it was in his own mind.
Speaker 1:He was, he was, he was translating between the two language for precision in his thought, and I think you know when I think of the prompt engineer, the precision and and Understanding that interaction, I think, as an assistant, it's an amazing, an amazing tool, an amazing opportunity, you know, for the market across the board. But but sadly, we are at the final question of this podcast and we we call it the trend drop. It's like a desert island question. So if you know, if there's only one technology, ai technology Within the capital markets that that you could track, what would that be?
Speaker 2:That's an easy one. Today I mean six months ago I would be telling you something else. A year ago, I'd be thinking for something else. Right now, generative AI transformer based large language models Is what I'd like to track. They are evolving at a phenomenal speed. They are and they're doing amazing things. They can turn out paragraphs, they can consume thousands of pages of document in a prompt and respond to that. I would be delighted to come back in Two, four, five, ten years time and see what this stuff is doing. That and robotics if we can improve our physical robotics, combine that with large language models, we are certainly going to find I'll find ourselves in an amazing world. I'd like to track those two things and if you forced me to pick one, I'd love to track our large language models and where they're heading well, I'll say this those dogs from Boston Scientific scare me if you stick a lot of language model on to them, they'll start talking to you.
Speaker 2:Now that's the scary part, and that's coming gosh.
Speaker 1:Well, I'll wait for that YouTube video. Well, the kind of Prague I want to thank you so much. This is was enlightening. I feel like I could probably talk to you for another hour easily and I appreciate your time today and and I'm sure our audience did as well. Thank you, delighted to be here. Thank you Coming up next week We'll dive into more technologies and our processes. You need to be tracking if you were going to capital markets. Take note of this important episode and join us next week. But first, if you enjoyed the podcast, make sure you hit the subscribe button, leave a comment, a like and check out our other episodes.