In this episode of Speaking of AI, LXT’s Phil Hall chats with Prem Natarajan, Chief Scientist and Head of Enterprise Data and AI at Capital One, about his passion for technology, its potential to benefit society, product design principles, and much more.
Phil:
Hello and welcome. My guest today is Prem Natarajan, Chief Scientist and Head of Enterprise Data and AI at Capital One. Prior to joining Capital One, Prem spent 12 years at Raytheon BBN Technology and was VP of Alexa AI at Amazon. In parallel with this, he spent ten years at the University of Southern California. Now, what most people might know is that Prem is an inspiration and that he is an industry legend. But what those of you who haven’t met Prem yet might not know is that he also has a deeply incisive sense of humor. With these things in mind, I’m excited – nervously excited – to welcome Prem to today’s edition of Speaking of AI. Prem, hello and welcome!
Prem:
Thanks Phil! Excited to be here. You know, our association goes back a long way, and I couldn’t be more delighted to be talking with someone about AI.
Phil:
Fantastic. I’m going to lead off with a question about where things might be headed. What do you think might become the battleground capabilities for generative AI in the next year or two?
Prem:
My answer might feel a little general, but actually I can’t imagine an area of human endeavor that is not going to become an area where somebody or the other is trying to apply and employ and harness the potential of generative AI. So in fact, in my mind from coding, to code development, to writing documents, to communicating with others, to finding information, to organizing our lives, etc. I think generative AI is going to have – applied properly, thoughtfully, responsibly – it’s going to have a beneficial impact on pretty much every aspect of our social and professional life. I don’t quite see it as much as a battleground as an opportunity to deliver so much magic for so many people around the world that there is enough magic to be created for everybody to work on it.
Phil:
A very diplomatic response, I think. Well, a couple of observations I’ve made. One is that we did a survey of senior executives working in AI-related areas and we asked about areas of application. And when we got the results, it was one area where we couldn’t see any trend at all. And my first reaction when I looked at the data was, “Okay, there is no trend in this area”. My reaction when I thought about it a little bit more was, “Oh, no, no, no”. What this is telling me is that it’s everywhere. It’s not that there’s the absence of a trend. The trend is that it’s being applied everywhere. But my second comment on that, though, is with regard to the battleground status of this, and I suspect that Microsoft and Google might see that differently. It feels a bit like an arms race between those two organizations.
Prem:
I don’t know about specific organizations and what their perspectives might be on it. I like to think about it in terms of individuals and humans and where are we likely to be feeling most opportunity driven or like which leads to competition in some sense. There is an ecosystem developing around AI and around generative AI in particular, right? And so when we look at that ecosystem and we look at some of the fundamental building blocks of that ecosystem, obviously the large language models or foundation models are one place where there is going to be a lot of competition and there are some open hypotheses that are still being explored. There are some people who are betting on one model increasingly becoming better at everything, right? There are others that have the belief that custom tuned models for specific domains are the way to go. There are others who are somewhere in between on that spectrum of one thing that is everything and versus lots of siloed things, like some balance of customization, the fine tuning, supervised fine tuning balance with large language models trained on a broad variety of data, etc. So I think if we think about it as one of the battlegrounds for human endeavor, I think one of the hypotheses that’s going to be tested is what is the right balance of massive models that are capable of doing more and more and everything well, versus specialization. So that’s certainly one area that’s an intellectual battleground, if you will.
Then the next level of battleground is what are some of the capabilities that are needed to unlock the application to two different use cases? If you’re in health care, for example, you may want the models to be more explainable or be more able to trace their reasoning when they come to a conclusion, right? And so that’s going to be another area of competition. Who has the models that give you the maximum performance consistent with what we see today, but are also able to explain their working, if you will. Yet another area of work would be in terms of reasoning. All of these are having interactions, whether responding to requests or, you know, the human is playing a big role in directing the conversation in some way. But when you get into more complex tasks, how are they at reasoning, etc.? So that’s yet another battleground.
Then I’ll elevate it and say there’s also a bunch of supporting technologies that need to be built around these, because these things, as we know, can make up their mind that certain things exist that don’t exist. And can do their best to try to convince you that they exist.
So things like vector databases, other knowledge bases in combination with these things. There’s a whole battleground for that. Then there is a lot of competition around which model-building platform should be used. Is it one that one of the big providers of cloud services do, or is it some of the smaller people are saying, these are also specialized endeavors. So there’s just a number of these areas. I will just say, like you said in your survey, you know, where you could either see battlegrounds or fertile fields of opportunity. And right now when I stand and if some generative AI thing and I look all around and I’m seeing, you know, 360 degrees of fertile fields all around.
Phil:
And when you mention this continuum between the idea that one model will dominate everything and at the other end that specialized models will be required for each vertical. Where do you stand on that continuum? What advice are you giving to folks?
Prem:
My belief right now, because I’m building stuff or I’m in a place where I have to build stuff
that can be used, and because we put a premium on the reliability and the robustness and in a lot of other things, given the kind of enterprise we are, things have to really work properly. We have to appropriately govern things, etc. My view is kind of more conditioned by the state of affairs right now. And I think we’re somewhere in the middle in the sense that we need the pre-trained models, but we can’t actually employ them for use cases without customizing them on our data. And that customization, Phil as you know, takes different forms. One is supervised fine tuning where you might collect data that is representative of your use cases and do that. And that’s one of the battlegrounds is what LXT is, who is going to be the, you know, the most proficient developer of instruction tuning data set and who can actually partner closely with people who know AI to be a data partner for that. I mean, this is your history. You’ve led the way there on the data side in the previous generation of AI development. You know, our joined work on trans tech, some of the fun that we’ve had. But coming back to this, I think that balance of finding, but the finding, it takes multiple things. One is supervised fine tuning. The other is how do you construct the prompt? And once you start about thinking, constructing prompts, now you’re saying, how do I involve vector databases in anchoring on curated facts, on data that I know is on hard ground that I have complete confidence in, etc. To me it feels like in the near-term future, maybe near to mid-term, we’re very definitely living somewhere in the middle of that spectrum.
Phil:
I’m going to shift gears slightly here. So you’re at Capital One. Capital One is very evidently a tech forward organization, but it is before being a tech forward organization, it’s a banking organization, it’s a financial institution. So it’s a financial institution first, very tech forward second. I’m not sure if I’m asking you a question that’s sort of out of limits here, but if you were advising your competitors on what they what anybody would need to do, maybe let’s put aside your competitors, but organizations that are less tech forward than Capital One, what do they need to be thinking about in order to not be left behind?
Prem:
So I don’t know necessarily that I mean, anybody needs to be left behind. But my view is each of us has to chart a strategy that reflects an approach to serving our customers. I like to think about it as all of us exist to serve our end customers in some way. And we’ve each adopted a particular approach. And at Capital One, we certainly adopted an approach that is tech based. To leverage technology to deliver the best value that we can to our end users and do so responsibly and thoughtfully. And so the path for us is around the way I told you, I think we’re somewhere in the middle. All of this in the end is simply a means to providing the value and the service to the customer that you want to provide. So in the abstract, it’s kind of hard to say like, you know, because somebody might say they want to provide customer service in a particular way and that that’s how they are differentiated, etc. But I think simply the fact that there is such a variety of services and service providers available and ways to leverage this technology, that every enterprise, whether it’s in finance or in law or in healthcare or in education, right? Everybody will find a way to leverage this technology to become better.
In some sense, you know, Phil, this was my view when we were in the darker world as well. The thing that we’re competing with is the current version of ourselves. How can we leverage technology to become better every year, every week, every month? And so I look at what’s available today and say literally everybody will be able to be on this continual improvement path, which means society at large and their end users and their customers will benefit. So I kind of think it’s in that more exciting phase now than like just a relative advantage, though, I certainly think like we, you know, we’re kind of leading the way in thinking about how to leverage all of this technology, harness it responsibly, think about the implications, but still use it properly. You know, apply the governance that we know how to and use it to deliver the best value we can to our customers.
Phil:
Right. I read one of your recent interviews and in the course of the interview, you said there is a deep imperative to operate all of this in a responsible, thoughtful way, which I think, you know, I think most people would agree with you. But when I think about a point like that, I wonder what is responsible, what is thoughtful. And in thinking about this question, are there some baseline truths or are those truths rationalized by competing perspectives and interests? For example, if you’re thinking about operating responsibly, is there a prioritization of being responsible to customers, to employees, to shareholders, to society as a whole? And is there a conflict in thinking about that sense of responsibility and thoughtfulness?
Prem:
I think one of the great, one of the great aspects of modern technology and its advances
has been that we’re able to make advances on all of these, because these are fundamentally things that make us all more capable at doing something, more capable of finding the stuff that we want, more capable of delivering value to people that we want to deliver value to. So in that sense, I’ve long seen technology as one of the democratizing influences. If, for example, back when we were working on speech-to-speech translation, the thing that excited me about that was the fact that even if you and I don’t share a common language, you know, as humans, we want to share a conversation, right? And so the fact that speech-to-speech, you know, real-time spoken language translation could allow us to have that conversation, even when you don’t share the language, was super exciting to me. In that same way, I think all of this technology, all of this AI, so when I say responsibly and thoughtfully, I’m kind of connecting to a few different motivations that at least are within me. One is, I want to make sure that this technology is making things better for everyone. It’s not in some way a trade off in the sense things are better for someone vs. worse for someone. In fact, that’s one of the things I find most inspiring about advances in technology, not just AI, is that it allows us to lift things for everyone. So that’s the first part.
Second is, as these things become more and more effective, more capable, I think we want to bring in more perspectives right from the design stage. I strongly believe in the notion of inclusion at the design stage. How do we bring in many different perspectives. Ultimately, this thing is meant to serve humans. So the more diverse perspectives I can bring in terms of its use and how something might be used, etc., etc., help us, help inform us to build a better and more robust product for everyone, right?
A third thing is there are lots of things we care about when we’re in such a business as we are, you know, customer trust and safety and all of that. So thinking about those things as we’re designing it, not just as an optimization after the fact, but right from the design stage, how do we bring all of those perspectives and to me, that’s the essence of a responsible and thoughtful approach. Be inclusive, make sure diverse perspectives are seen, make sure you’re constantly testing all of your hypotheses, that you’re generating hypotheses to test in your testing, and then that you’re also testing them in the real world before you actually launch them into a product.
Phil:
It’s very reassuring to have someone like you in this role. I’m really pleased to see that. So I mentioned DARPA connection a couple of times during the course of the conversation here.
Prem:
And, well, that’s how we’ve met you know…
Phil:
It was an exciting time. It was. Yeah, I would come to those DARPA group meetings where for those who in the background here who don’t know, we worked on DARPA projects where multiple technology leaders would be competing with one another to make advances. And the competitive environment actually produced very rapid advancement in the technology. And it was exciting. It felt like we were doing something that was going to change the world. Or, well, let me back that off slightly. It felt like we were doing something that might change the world
if we could get it right. But it also, to me, felt far from certain that we were ever going to get there.
Now, I’ve got a quote here at the end of five years of hard work on the trans tech program, the director of the program, Mary Mader, described the accuracy that we achieved
as “enough to be interesting but not enough to be useful”. And at that time, I have to say I had doubts as to whether this would ever be possible. It was a struggle to make the technology work in even the cleanest and most controlled of environments. And the reality is that we were designing technology to work in the dirtiest and most difficult of circumstances with all the issues that come with that. How strong was your confidence that we would ultimately achieve
the breakthroughs that we’ve seen, particularly since, let’s say, 2012, 2013? There’s been massive breakthrough after massive breakthrough. And now the recognition, translation and synthesis pieces that went into trans tech are all at level of advancement that I found difficult to imagine back then.
Prem:
Indeed. So I’ll say this, I was super hopeful. I mean, partly when you’re you know, you’re in the business that I’m in, which is building technology that is going to be useful a few years from now, right? Deep inside you, deep inside you lives a technology optimist, right?
Phil:
Yeah, absolutely.
Prem:
So in some sense, it’s also a mindset in my mind that you think this problem is so worth solving. I think the way I’ve approached a lot of these things Phil, is think about the feasibility later. But first, ask yourself, is this problem worth solving? Is this something that is a hard problem, but one that will make so much value to the world or deliver so much value to the world when it is solved, right? Once you pass that filter, then the next thing you have to ask is, are the elements of this problem interesting to me and to me, anything that involved human language and machines learning them, etc., etc., that I’m and still does remain very interesting. Just fundamentally, they appeal to me, you know. It’s one of those things you don’t know why, but they appeal to you, right? And so I’d say my mindset at the time was, boy, this is such an important capability or set of technologies to develop. And boy, this was so hard. And to me all that meant was, oh my God, I’m going to have fun for so many years solving this problem, which is that, you know, so it was it was less and less the question of, oh, this is overwhelmingly hard, and more like, oh my God, this is going to be so awesome to work on because… and deep inside that, the basis is if you thought you would never succeed, then it would be hard to keep that optimism channeled, right? So I think deep inside I always believed this is coming. Could we have predicted the specific arc of things? Probably not, right? We didn’t predict the specific arc of things. But when you look back at where we even during the speech-to-speech, translation times, you look back, you can see that there was several inflection points in the technology even during the course of the program. Even during that five years, which she said it went from being interesting. I don’t remember the exact quote, but not entirely useful or something like that. I would say even the users of it were finding it useful in very specific instances, right? It was just not useful in every situation. That’s a big change. Like at the start you had no technology and people were simply stringing together different components to deliver a capability. By the end, we had an integrated thing that was talking to each other, etc., etc. and being able to deliver a thing which some people found useful in some cases. So my way of thinking about it was like, oh my God, look at this. At the start of this thing, we thought we could never get that in five years. Halfway through it, we said, oh, maybe we can get that in five years. And then in five years we look back and we say, oh my God, I can’t imagine what a distance we have traveled in five years. And if that’s how I feel about the first five years, do I want to continue on that journey for the next five years? Heck yeah.
Phil:
Yeah. Awesome, right? Absolutely.
Prem:
Yeah.
Phil:
Very cool. I have to say, that was some of the most enjoyable time in my life. I felt like I was on such a voyage of discovery and I loved attending the events associated with it. It was something else. I just have a couple of wrap up questions here. One is, what do you see as the biggest risk to successful AI deployment?
Prem:
I think there are a few performance curves that we have to bend, right? I mean, we have to make this more affordable across the spectrum. We have to bend the cost of being able to deliver good, useful AI across a broad spectrum of use cases. I think that’s the first in my mind. The second, I think, is I’ll go back to my responsible point, I really think across the board we need to make sure that we’re focused on understanding the behavior of the things we’re building and the potential outcomes and bring in a diversity of perspectives. If that means that we move a little bit slower than we could otherwise, in my mind, that’s okay, because at this point I see the potential for so much impact and for so much beneficial impact if it is done right, that I think one of the fears could be that, you know, so that there is a race to kind of be the first on something and that’s not ideal or something. But beyond that I think there are a bunch of other factors to consider: who owns the data, and how do we build these models? There are some business models to be worked out in different things, but in the kind of society we live in where we’ve managed to find our solution, I’m pretty sure those I’m optimistic about, like we’ll find a fair arrangement that works for everyone. But there are some of those things that need to be worked out. But overall, you know, Phil, I’ll go back to being my optimist self.
Phil:
Yeah.
Prem:
I think we’re going to figure it out, you know?
Phil:
Yeah, I was actually on a call about 12 hours ago late last night discussing a fairly difficult problem that we’re facing and our head of technology said to me, he said, look I’m pretty confident about this. I’m 90% confident that I’m right. And I said, my optimism is at the same level as your confidence. My confidence is not, but my optimism is at the same level as your confidence.
Prem:
But this is why we all love language. So, I mean, look. Absolutely. Wonderful semantic nuance that you’re teased out and it’s just such a joy to hear it, right?
Phil:
So, Prem, my final question for you. If you were running this interview, what is the question that you would really be hoping I would ask and what’s the answer?
Prem:
If I was running this and I know this is in the back of your mind, how is my family doing? How’s my wife? How’re my three girls? And you know, they’re doing wonderfully. You know, they’re the joy of my life. And, you know, part of what also drives me is to, you know, keep using AI to make this a better, more joyful place for all of them to be in. So I’ll say that was the question that I know is at the back of your mind and I asked it for you and I answered.
Phil:
Prem, it is always a delight to spend time with you. I’m pleased we could do this today. I’m sad that we’re not in the same room or perhaps over a dining table.
Prem:
Same. But I know we will be again soon.
Phil:
I can’t thank you enough for your time today. I think people are going to find this tremendously insightful and interesting, and I thank you from the bottom of my heart, Prem.
Prem:
Thank you, Phil. You know, the feeling is entirely mutual. And I’d also say, you know, I’ve seen you go through so many incarnations, you know, musician, you know, data maven, you know, growth officer and now podcast host! It’s just amazing. So, you know, I have to figure out
how to follow that kind of trajectory.
Phil:
But thanks again, Prem. I really look forward to the next time we speak.
Prem:
Yeah, take care.
