AI Innovation at Fidelity

Maureen Olejarz Welcome to Tech on Deck podcast, brought to you by Fidelity Investments. I'm your host, Maureen Olejarz, CIO of Enterprise Software Engineering. Adam Ely And I’m Adam Ely, the Head of Digital Products and Engineering. Adam Ely Each episode takes listeners inside the walls of a fintech industry. Maureen Olejarz Anything from cybersecurity, artificial intelligence, cloud and crypto to the intersection of product and technology. Adam Ely Tech on Deck breaks down the topics top of mind for technologists today. Maureen Olejarz Plus, we'll give you insight into the exciting and challenging careers in fintech. Adam Ely Welcome back to another episode of Tech on Deck. I'm Adam Ely, Head of Digital Products and Engineering at Fidelity. Maureen Olejarz And I'm Maureen Olejarz, Head of Software Engineering at Fidelity. We are super excited for this episode as we have a special guest joining us today, Fidelity's head of artificial intelligence innovation, Vipin Meyer. Vipin, welcome to the show. Vipin Mayar Thank you. Thank you for having me. Maureen Olejarz All right. So Vipin, if we get started here today, we have a lot of things that we could talk to you about. You know, AI is such an exciting topic and as a hot topic, it's an understatement. And based on your career and we're going to learn a lot more about all of that and what you're doing. So, what we'd like to do is start off with tell us a little bit about your views on artificial intelligence and what led you to the position, how you think about it in the industry and career opportunities that exist out there. Vipin Mayar All right. So, let's just begin by saying artificial intelligence is at the top of the hype curve, and it's been there for many, many years, which means that the expectations from AI far exceed the expectations from any other technology. We at Fidelity have been focused on that for quite a while. I've been leading this for the last seven, eight years, and over the course of the seven or eight years we've done hundreds of use cases. Vipin Mayar We have filed for over 120 patents, published over 50-60 papers. So, there's a lot that's been done. Maureen In the space of AI. Perhaps I'll begin by just highlighting the work for you and then you can ask me some more questions. Maureen Olejarz I think that's great. Yeah. Vipin Mayar Regarding the work, there's the thing about AI is that it is very broad term. It can be applied to almost every aspect of a company's business. And I would bucket all the work in two buckets. Bucket number one is things we do for our customers. And bucket number two is things we do for the productivity of our employees. Vipin Mayar Now, on the customer side, maybe I'll give you a few examples. Just to make it come live and bring up the strength of what we've done. So, one of the things we did, we were very early on this, we decided we needed to understand the intent of a customer when they were interacting with us. So, Adam, if you were calling us or Maureen, if you were on the website, we have AI that's looking at your behavior and your profile and then determining the purpose of your interaction. Vipin Mayar And if it’s a call, we then send it to the right rep. If you're online, we try and give you the content you're seeking. It's been a real big boom in improving customer satisfaction and driving to successful outcomes. Another big piece on the customer side is customers are looking to plan their financial matters with us and they're sitting inertia and they’re setting up a plan. Vipin Mayar We are using AI to create what I call default plans for you so that the barrier for starting a plan goes away. You can look at what they've done for you and you can lift off from there. And then there's a lot of work we do around paper and processing it in an automated way so we can take the friction and delays for anything that's associated with paper. Vipin Mayar Let's look at things we've done for associates and drive productivity into many, many here. But I'll begin with advisors – advisors’ that on the front line have a ton of appointments. Each appointment is a very important moment with the customer. There's a lot of preparation that goes into these appointments. We've created an appointment prep tool that synthesizes the whole relationship to give some recommendations, Vipin Mayar give some suggestions, big boom for them improves the interaction and makes them more productive. Then we've also embraced co-pilots, especially with Gen AI. Co-pilot is an assistant to the human being, and we've got co-pilots running in many different places, including in the development area. Plus with our knowledge workers that are looking to synthesize a lot of vast amount of data and the AI helps them summarize that. Vipin Mayar Those are just a few examples, gain a lot of breadth of use cases, and we are doing some really, really innovative work in all these use cases. Maureen Olejarz I think that's great. Thanks. Adam Ely There’s so much in there. And to your point. There's so much in there to understand about AI. I am going to give you a softball question. I'm pretty sure that you're going to expect this for your background, but I think it's a good grounding point and it builds on sort of your in the you mentioned Gen AI right there. Adam Ely What's your view on Gen AI right now? Vipin Mayar Gosh. So, let's just for a moment, Adam, just take a step back because I want to ground my view on what we are seeing. And you might remember the moment you first heard about Chat GPT. November of 2022 - the day it came out. The next day I was on GPT 3.5 looking at it and I was blown away by its capability and that happened to millions and millions of people. Vipin Mayar And fast forward now, Gen AI is now integrated into everyone's psyche and we've got to give credit to Chat GPT for creating this word and this momentum for us. Now what is Gen AI? Before I give you a view, let me just tell you what Gen AI is. Very simply these algorithms are creating new content, be it videos, be attached, be it images, be it all kinds of strategy documents. Vipin Mayar And the other thing to know about Gen AI is it's done by large language models that are fed tons and tons of word, and it uses pattern detection to create content. And that's an important concept in that it really doesn't try and understand the meaning of the question. It really creates a pattern around the outcome. So, it's been very successful, and these models are now everywhere. Vipin Mayar So I feel, Gen AI, it's transformation. A good analogy which people have used, it’s like early days of electricity. It's here. It's going to transform a lot of things. It does have a lot of issues which we'll get into, but it is here and the amount of money that's going into it, I added the number of zeros in that it's it is innovating at a fabulous, fabulous pace. Vipin Mayar And we're expecting a lot of really exciting things to come out really soon. Adam Ely So let me pull on that a little bit. If we say we're in the early days, just like when electricity first came out, so maybe in one or two cities, but everybody understands it or everybody knows what it is because everybody understands it. A few of us do. Or I’ll put it another way, many of us me do not. Adam Ely What do you expect in maybe 2024 for the next year? So, it's second year kind of out there in the mainstream. What are you expecting to see? Is there some great big new things? Do you think it's just more adoption? I'm curious, what do you think the next year? Vipin Mayar Yeah, yeah. So, we are expecting a lot and I expect it’s going to - the application of Gen AI in different verticals is going to come out in 2024. I'll give you five or seven quick things on what we expect. So, the first thing is we expect this will become multi-modal. So today GPT is largely text with some images, but you're going to see a lot of these models become multimodal. Vipin Mayar Second thing, the term “large language models” has been used and it's because these models are large, a trillion different learning functions in them. They consume a lot of electricity. In 2024, there's a big movement to make them much smaller, much, much, much smaller. So a trillion come down to a few billion. And the reason is they are very expensive to run. Vipin Mayar And frankly, we don't have enough GPUs to run them. So you're going to see much, much more smaller size. And the good news on that is they can then work on smaller devices. On the phone, on your watch, you're going to see these generative A.I. applications become smaller. Now, the third thing, which maybe feel a little bit astounding to people, is that the data being used to train these models will soon be synthetic data, artificially generated data. Vipin Mayar Now, that could feel bothersome, but the fact is a lot of the data that is being fed into the models has questionable quality in them, which leads to a phenomena called hallucinations. We won't go there yet. There's a greater need for high quality data, and there's a feeling that we can manufacture high quality data from looking at high quality data that's real, but create synthetic versions of it. Vipin Mayar You also going to see number four, you're going to see open source really, really start to catch up to the GPTs and start to what would take them. And I think a lot of people are looking forward to that. Next, you’re going to see that the prompt window, which is the interface that most people have into these models is going to become very, very powerful. Vipin Mayar It's going to have reasoning built into it. It's may become the excel of the future skills in prompt engineering. And then the last thing I would say is you may remember this not so far back when the app started coming on the iPhone, we would say “There's an app for it!” and we would just be amazed like “there was an app for this” and “an app for that”. Vipin Mayar Well, that's what's going to happen with these LLMs, large language models that will be relevant for this and for that for every single task. And a lot of that I think will happen in 2024. Adam Ely So it's interesting. I want to pull on something you said, but I'm going to come back to it in a second because there's something really interesting here. If I go back to say maybe just mid-last year and it's amazing, we're talking a timeline of like 15 months or something, I was just doing math in my head. If I go back to mid-last year, the big conversation I heard everyone talk about was the limitation in electricity, GPU availability and how this was going to stall the industry, stall the innovation. Adam Ely But what you're saying is they're sort of fixing that because they're moving these smaller models trillions to billions. Billions still feels pretty large to me, but these smaller models, which aren't going to need as many GPUs and go onto the form factor, I can go on to my phone, go on to my laptop, and I also remember the conversations right after a lot of people talking about the quality of data like they couldn't get good data to put in. Adam Ely So now we're looking as an industry to solve that problem by creating good data. So, this industry sort of fixing its own problems to keep its innovation going, is that right? Vipin Mayar Yeah, I mean, I mean, that's that's a very astute observation that the constraints that we saw last year through innovation being addressed. I mean, you mentioned two of them. I mean, there's another very fundamental one that's coming out in research papers, which is the underlying transformer-based methodology is the compute intensive, very, very compute intensive. And people are now looking at other methods that are not as compute intensive. Vipin Mayar And this is not a signal for saying that this is the end of the transformers. They will be supplemented or in certain cases substituted with less compute intensive methods. Adam Ely There's a lot to pull on there about like tech careers and just like how an industry I think that we thought, Maureen that's like transformational A.I. just in general. It's actually creating a bunch of now different opportunities for people as well. Maureen Olejarz I think so in a transformative industry that's trying to mature and trying to scale. Adam Ely Vipin, before you fully move off, there is something you mentioned A.I. hallucination. This is a term that I hear a lot, a lot of people sort of throw around, but I think taking a minute to talk about that would be really good. Can you expand on that a bit and maybe give an example for people that aren't really familiar with the term or how the tech works so they can get their head around it? Vipin Mayar Yeah, hallucination. Fascinating word from psychology roots go back to the 17th century. Now it's being brought up in context of AI. A little bit disturbing, but it is a disturbing phenomena in that it refers to the instance where a large language model presents things as fact, but indeed it's not. Now, the first time it came, it became really public Vipin Mayar And you may remember this, there was an incident happened last year where a lawyer referenced three citations of legal instances to present a case, and all those three, in fact, were non-factual. They were made up. Yeah. And so, the question that leads to in this is why is this model making things up? And that's what’s called hallucinations. Vipin Mayar Now, the couple of things to remember. Firstly, it is a pleasing model. Its aim is to please you. If you ask it for something, it will answer it for you. As yet it doesn't have the sophistication to say, I don't know. It goes back and gives you the best pattern of sequence. It thinks it matches the words you're putting in. Vipin Mayar So, we need it to be less friendly and we need it to say, “No I don't know enough of this”. The second thing about hallucinations, since it's probably one of the biggest issues, is that a methods to detect it. And now you're seeing models that are created within a large language model that check for hallucinations. For instance, simple thing like you also question these checking models would create ten versions or fifty versions of the same question and ask it to respond to different versions of the same question and see if it gives you the same answer. Vipin Mayar That's one instance. We need it to be fixed. We need these hallucinations to go down, and it's largely a factor of its seen data that is not factual. And that's why back to your point of the industry fixing the issue, I think as we create more real high-quality data, it will get addressed as well. Adam Ely I feel like with AI hallucinations, students can now use that as an excuse in place of the dog ate my homework. Like if they're if they're if they're term papers just a little off, they can blame their copilot. Maureen Olejarz No, for sure. How about bias? Vipin Mayar Yeah. You're asking such great questions around, you know, how do you know the response to a question is not biased? And I mean that issue -- Maureen Olejarz Because that's different than hallucination. Vipin Mayar And that's because the data, it's it's not truly representative, it's factual, but it's not a true representation of the population, for example, that issues existed in AI for a long, long time. And prior to GenAI we had methods to check on the quality of data coming in by looking at distributions that quality of the output by looking at distributions. Vipin Mayar But GenAI doesn't give us that same ability because it's being fed trillions and trillions of words. And on the output side also, it's very hard to know the bias detection piece. So, I think it's another one of those issues that the industry is trying to get their arms around, and it's being done by a set of models that then check for bias. Vipin Mayar And those models have been trained on bias data. So, they've been trained by saying this is biased, this is biased, and this is not a representation of bias. So, you're seeing those models getting developed to check it. Fascinating. Maureen Olejarz It’s a fascinating industry and evolution, right, of technology and capability for the industry. Adam Ely So, Vipin, as we've been talking about this technology and we kind of touched on a couple times that this is kind of an emerging industry and there's a lot of people working on lots of things, even solving the industry's own problems to continue innovation. So, it's a really interesting career space. I'm curious, can you tell us a little bit about how did you get down this path and maybe what kind of people do you look for to join your team to work on these kinds of challenges and apply this technology? Vipin Mayar Yes, And I'll first speak about myself. And like you, I've been on the tech side, and I've taken on non-tech product roles. Let me just go back in time. So, I spent five years studying engineering and then went on to get a business degree in marketing and finance. My earlier career was in I.T. consulting and I also happened to, in these consulting assignments come across vast amounts of modeling data. Vipin Mayar And in the late eighties and early nineties when I was doing this, I saw that the quality of the modeling that was going on wasn't of high quality and in some ways a little bit shocking. When I saw some of the largest companies in America basing their campaigns on predictive models that were not all that good. Anecdotally, you know, since you're from the Boston area, you might find interesting in the year 1992, which is what roughly 30 years back, I was in the throes of looking at models that were predicting cable behavior and adoption. Vipin Mayar So, there was a fan amount of consumer modeling going on, predictive analytics, not necessarily high quality. So, I had that experience and then I joined Fidelity for the first time in 1995, had a job on the institutional side running data analytics, and then in after six years, I left Fidelity and had some very interesting experiences in New York City working for McCann. Vipin Mayar McCann's one of the largest global marketing companies, works with some of the big brands and handles their entire marketing investment. And through that, just completely saw many different verticals how optimization was taking place. And those were the early days of digital, Adam. And so, we were also looking at data that we had never seen before exhaust and search and mobile. Vipin Mayar You know, it was fascinating time I happened to write a book on the topic and was actually surprisingly, I was trying to write a business book and then realized that no business person wanted to read it, but in fact was a really good textbook. So, I had some offers to teach in New York and it became a reference book on some courses. Vipin Mayar Fast forward from there. I rejoined here and seven or eight years ago when AI in the days of big data, I started making a difference in image recognition and started understanding language, not generating it, which is now, but understanding it. We set up a center of excellence and I started leading that center of excellence. So, my career took me from engineering to the business side to AI, which requires both those skills ideally. Adam Ely So, one of the one of the things I love about this is I've known some of your background, but I did not know the marketing work that you did. I didn't know you went down that path and it just it constantly reminds me that you never know what path somebody is going to take and where they're going to end up. Adam Ely All these experiences come together for this value. That's an incredible journey. So, when you think about where your team is today and where you're trying to go, are you looking for people to join that team with similar backgrounds? Is yours are different? Like what's that makeup? Yeah, yeah. Vipin Mayar And it's changing as we speak because the capabilities are getting more democratized. So, I would say we look for four different skills. Now, if you're a unicorn, you have all four skills, which I know the two of you might be. But now most people, you know, you've got some combination of these, so I wouldn't be concerned if you don't check the box on all four. Vipin Mayar But if you do, please pick up the phone and call me. But otherwise, these are the four skills. So, we look we do look for math skills. Now, the math skills could be sufficient if you've done engineering or econ or math. Some people have gone all the way to do PhDs in math. That is still appealing to us because we'll put you on higher end problems. Vipin Mayar But you will check the box on math skills if you've done some quant-oriented education. The second thing we look for is computer science, and it could be programing skills. It could be just understanding hardware, operating system software and being able to program. So that's the second skill. The third skill we look for are data skills and actual experience with using large amounts of data and a no sequel instances. Vipin Mayar Hadoop – if you're familiar with some of the methods? MongoDB is just being able to work with large amounts of data to be the third skill, and the fourth one is business skills, because AI is solving business issues and it's not understood that well. So, we feel like people who can understand the domain and think about applications for the business. Vipin Mayar So, we have needs of all these four skills. And you know, if you have these some combination of these, come join us. Adam Ely Well, you just disqualified me as a unicorn since, as you said, math skills. So, I was out there. But I mean, that's an amazing makeup and I didn't expect I was trying to predict where you were going. I didn't expect business skills. Makes a ton of sense as you explain it, but I did not expect that. That's really interesting. Maureen Olejarz Well, I think for Vipin, he's been on the business side and in tech. I think it's a great combination. So, this has been fantastic. We could talk for hours with you, so let's bring it back to AI and financial services industry and then kind of weave in what's exciting to you in the financial services industry. This is the industry we're in, in Fidelity. Maureen Olejarz And then just talk about, you know, tech careers and how you might think about that. Vipin Mayar I think you know, this Maureen that AI there's certain sectors that lead when it comes to A.I. Financial services is one of them. Maureen Olejarz As you said, advisors and everything else is. Vipin Mayar A great place to be at. Maureen Olejarz Yes. Vipin Mayar Now, now, you might say, why is that? So let's just unpeel that for a second and say financial services has a lot of data interaction. Transaction data, financial services tends to be quantitatively focused and has a great culture of measurement, experimentation. All necessary ingredients for A.I. financial services by itself has a very complex product. You're not selling a widget That is a quick impulse buy. Vipin Mayar You're selling something that takes over time and AI is good at handling complex problems because it requires human plus machine to come up with the right solve. Those would be some of the reasons. Now, I'm very bullish about people joining financial services and especially Fidelity to place to make a career. And the reason is coming into financial services - Vipin Mayar You make a difference in people's lives. At the end of the day, you know, if you've sold an algorithm, it's not to sell something. It's really making a difference. Secondly, we have very high ethical standards. We are highly regulated. So, the work that goes on is of the highest quality and highest standards, especially of Fidelity. We we take that very seriously. Vipin Mayar We have strong governance around it. And then I would say financial services has been quick to adopt agile principles and the way we operate with, especially with the A.I., is a very agile, nimble way so that you get the benefit of a big company and the investments and the use cases. But yet you operate in small teams and all to make a big difference. Vipin Mayar So you can see big fan. Vipin Mayar Because of AI and financial services and you know, there's a lot more for us to do to help people live live richer lives. Maureen Olejarz I think that's great. So we couldn't say it any better. Adam Ely It's yeah, it's amazing. Super inspiring. Yeah. So, if we want to help people find that career path for themselves, or I'm going to ask selfishly, because I want to know how to learn more in this space too, and get better. Where can they go to learn or where can they go to experiment to build a little bit of those skills, whether it's just learning concepts or maybe they actually are trying to figure out what the tech stack is and how to get started. Adam Ely Do you have any recommendations for people? Vipin Mayar Yes, yes, yes. So the good thing is learning and education has been transformed by A.I. Maureen Olejarz So there you go. Vipin Mayar So let's take the easy ways to learn. So, the first an open source takes a battery, right? So let's first think if you're looking to learn some of the newer concepts around neural nets and what's driving, you know, transformers, a lot of courses online, most of them are free. Andrew INC’s course on deep learning AI is really, really good. Vipin Mayar The good thing about those courses is they give you environments to also practice project work. So, I would begin there and make sure that you understand some of the terminologies. The next place I would go is go to open source. Now you might say like, how do you go to open source? You can type in open-source dot com, but it looks like you may not get there. Vipin Mayar “Hugging Face” is a supermarket for open-source algorithms so “HuggingFace.co” and once you go there it's very easy to look at the algorithms that are available. You just need to navigate to the type of algorithm and select one the first time and then take you a couple of minutes to do it. Just to read up on which one you want to play with. Vipin Mayar But then it gives you the ability to take it and work it and what's called spaces, an environment where you can actually learn the algorithms. And the only missing piece is data. And the beauty of this is it's a data supermarket with over tens of thousands of data sets. So, you can move the data to the algorithm and you can be in business literally in minutes. Vipin Mayar So you can get pretty much hands on. And then to keep up, I tend to use “Medium”. I like medium as a reading source. I also tend to follow a lot of people on LinkedIn and people like to talk about what they're seeing so if you follow people, there'll be no shortage of content that just siphoning the one that you really want to read. Vipin Mayar Maybe next time I see you, you'll be proficient in a couple of these algorithms. We'll check back with you. Adam Ely A little busy. A little busy. I don't have a lot of time. Proficient may be too strong a word, but I'll take the challenge of at least trying to learn some more and being able to demonstrate. I'll take that challenge for sure. Just keep a low bar on expectations there. Maureen Olejarz Vipin, I think we're going to have to check in with you in a while, too, as the as we continue, whether in Fidelity Financial Services, and then the work that we're doing, whether it's for customers or for our associates and productivity and things of that nature with the moving marketplace and capability, we definitely want an episode in the future where we can talk about some additional work that we're just beginning. Maureen Olejarz Okay, so Vipin, we typically go and ask our guests, you know, just to make you more real to people, listening is a fun fact about yourself. Vipin Mayar Yeah. You know, so far I Maureen, I've liked all your questions except this one. I don't know, there's anything fun about me, but, you know, and I'm like, “What should I say?” I mean, I could say a couple of things. We've had one or two other guests that have said that, and they they came up with. Vipin Mayar Ultimately, I could say I'm a pickleball addict, but I don't know. That's fun. I just it's like, okay, you place pickleball. Or I could say, you know, lately I have gotten into a card game called Bridge, and I like to solve tough Bridge hands, but I figured that's still not enough. What I think maybe really interesting is that today, or any given point in time, I have a choice of four citizenships for now. Maureen Olejarz Wow! Okay. Alright another surprise. Vipin Mayar Yeah. You might say like. Like, how can you have can you choose four? Okay, so I am a U.S. citizen. I've been naturalized here. I was born in the U.K., so I have a birthright to British citizenship. My parents are Indian, and I grew up in India so I can get Indian citizenship. I'm married to a Canadian and I was a spouse of a Canadian. Vipin Mayar I can get Canadian citizenship. Maureen Olejarz Did you study all that? Adam Ely I don't think I know anybody that has that much optionality. And that's a first for me here in that that that is a fun fact. Beats the pickleball for sure. So, Vipin, it's it's been fun. I mean, we've dug into everything from GenAI and hallucination to tech careers to your optionality and citizenships around the world depending where you want to live. Adam Ely And it's amazing. You know, I know we're coming to the end, so we wanted to thank you for taking time out of your busy day, helping us understand some things our listeners understand financial services and AI, and really kind of what they could do with their career based on what you've seen. So, thank you so much. Maureen Olejarz Thanks so much, Vipin. Vipin Mayar Thank you for having me. It's been fun. Adam Ely Thanks for joining us for Tech on Deck. We hope you enjoyed the episode. If you haven't yet, please give us a five-star rating and subscribe to the podcast on Apple Podcasts, Google podcasts, or wherever you get your podcasts from. Maureen Olejarz Thank you to our listeners and recording studio and editors who make our episodes possible. To learn more about tech opportunities, head over to Tech Dot Fidelity Careers.com. Adam Ely See you next time.