Key Wealth's National Call - AI: Everything You Are Afraid to Ask but Need to
George Mateyo [00:00:00] My name is George Mateo, I'm the Chief Investment Officer here for Key Wealth, and I'm really thrilled to have you join us today for a really great, interesting, topical conversation around artificial intelligence, or AI as it's known as, more colloquially. I'm just really thrilled to have this conversation. Frankly, we've had a number of calls in the past year or so focusing on the economy, focusing on the capital markets, interest rates, the Federal Reserve, the administration. And this is, frankly, a nice departure and a breath of fresh air to talk about something even probably more broadly than all of that, in the sense that artificial intelligence has certainly dominated our conversation, and it's seemingly kind of creeping into every element of society, in the past several months and years. And I think, perhaps more importantly, going forward, that profound change that we've seen thus far already take place is poised to accelerate and continue to extend from here. So, I'm very pleased to have a subject matter of experts, a panel of experts for us today. to talk a little bit about AI. What does it mean, frankly? How do we kind of understand what it means? What is it? How did it come to be? How did it grow so quickly? How has it become so prominent in our lives? what's the outlook going forward? And then let's talk about some use cases, and I've invited somebody from my team to also talk a little bit about things we're doing here inside Key that you might be curious to know about as well. And then again, we'll expand the conversation at the end, talk a little bit about AI more broadly, and some important implications for us as a society, and also as an economy. So, again, I'm really thrilled to have a great number of people joining me on this conversation today. First of all, a quick housekeeping note. If you would like to ask a question, thank you, first of all, for your questions in advance. We received a number of questions as you registered, and we'll do our best to try and answer many of those in the next 45 or 50 minutes or so. But if you didn't ask a question, or if something occurs to you during the conversation here this afternoon, please do so. There's a chat box at the bottom of your screen that you can submit a question to, and I'll try and do our best to try and answer those over the next 45 minutes or so. Join me today, though, again, as I said earlier, are three domain experts as it relates to artificial intelligence and technology more broadly. I'll have each of them spend a few minutes talking about their specific subject matter of expertise, but quickly on the screen, you can see my colleague Mark Greeby has joined us from our key, technology team. Mark is an 8-year veteran with our firm and a 20… I think a 25 or so year veteran in technology overall. Mark, glad to have you along here. Mark is a terrific partner of ours inside our wealth business to really harmonize some of our strategic priorities and objectives from the wealth management business, and also bringing forth the technology infrastructure that make our… makes our firm work. Paige Bailey is joining us from Google DeepMind. She is at the intersection between research and applications to take AI and actually make it applicable, spending time with her research folks to actually think about new technologies and new innovations, and then bringing those to life. And then Jill Italiano is also with Google DeepMind. who's leveraging AI in the business application sense of bringing applications around artificial intelligence to businesses as well. So, Mark, as I said, I'll just have you spend just a few minutes talking about yourself, your background, and your, your focus on AI more broadly. Mark?
Mark Grebey [00:03:04] Sure, thanks, George. Hey, everyone. It's actually a pleasure and honor to be here, being a key technology employee. Typically, our business partners would hide us nerds from our client base, so… You know, it's rare that we get to actually interact directly with our client base, so I'm pleased with this opportunity, and I'm super excited to be here. So, as George mentioned, I'm an 8-year veteran of the firm, but a 26-year veteran of technology. I started out way, way back in the day working for a bank in Cleveland called National City Bank as a programmer. So I have about 15 years of programming slash engineering experience, and I've been in leadership ranks for a better part of 11 years now. My role at Key is I'm the head of wealth technology, and so I work directly with folks like George and some of the other senior leaders here to kind of ensure that we're aligning Technology strategy along to the business strategy, and… and are forward-looking. And, as a result of that, I am helping drive our AI strategy as it relates specifically to wealth here at KeyBank.
George Mateyo [00:04:12] Excellent. Thank you, Mark. Happy to have a fellow nerd along as well, so I'm glad you could join us, and thanks for that introduction. Paige, I'll pull you into the conversation next. Paige Bailey, if you don't mind spending just a few minutes introducing yourself as well.
Paige Bailey [00:04:23] Yeah, absolutely, and I feel like this session is one that all of us nerds are going to be very, very excited to have this morning. My name is Paige, I'm the DevAx Engineering Lead at Google DeepMind. My career started not really in computer science, but more in kind of the geophysics-applied math space. And I did computer science and carbonate geology for grad school. First started doing machine learning, I guess, around 2009, 2010, so before it was… before it was cool, before it was so ubiquitous around the world. And it's been just really wonderful to see how it's evolved and changed over the span of that time. Started at Google, first in Google Brain around 2018 timeframe. I was working on our machine learning frameworks, then eventually, eventually kind of… joined into the… the model-building world, for the first iterations of Palm 2, Gemini, Gemma. And now I'm kind of bringing our models to market through the Gemini APIs, AI Studio, and working with great customers like yourself to kind of really understand the use cases and the potential benefits of AI. So it's been, like I said, really great to see how everything has changed and evolved, and I know that there are a lot of questions, for folks who maybe haven't seen, kind of, the end-to-end timeframe of AI, so it's great that we're going to be discussing that today.
George Mateyo [00:05:58] Excellent. Thanks so much, Paige. And then on the other part of the country, we have Jill Celliano, also from Google DeepMind. Jill, please say hello.
Jill Taliano [00:06:05] Absolutely, thank you, George. So I'm actually a Gen AI Specialist within Google Cloud, so part of the broader Alphabet family. I've spent the last 25-plus years working, kind of, for and with, banks, so ranging from Central banks, all Federal Reserve Bank of New York, the megabanks, kind of regional, super regionals, and community banks. started dabbling in Gen AI back in 2017, and just super excited, been with Google for 4 years, and, you know, really in the last 2 years, I feel like it's come to a point within industry, that it's actually a primary focus, right? We've gone from experimenting and kicking the tires to actually, you know, starting to put stuff in production. So you'll see, you know, I think people understand that they're using it today as we talk about that, but they'll actually see it emerging, both within the banking industry and, frankly, every, every industry, around the globe. So from a, focused perspective. I'm really working with the large banks, all banks, around leveraging AI to optimize business processes, things like personalizing customer experiences and accelerating innovation. For me personally, I think AI is and will be one of the most exciting seasons, I think both personally and professionally, and we're super excited to be here today to share a little bit with you.
George Mateyo [00:07:29] Awesome. Thanks so much, Paige and Jill and Mark. All right, so I'm going to start maybe just at, really, like, 50,000 feet or so, and actually, I did have ChatGPT actually answer this question for me, so this is somewhat pre-contrived, and maybe I'll ask you, Jill, or Paige to see if you agree or disagree with the statement. But I thought just for our listeners and for our audience members, it might be helpful just to really frame the discussion, first of all, of what What do we mean by artificial intelligence? What is AI? And according to ChatGPT, it said it's a field of computer science focusing on creating machines. that perform tasks, Typically requiring human intelligence, such as learning. Reasoning, problem solving, and decision making. Now my wife would say, I'm not good at any of those things, so maybe to some extent, AI's already surpassed me, but I think the key… the three key things here that kind of bubbled up from that was, again, simulating human intelligence, so again, kind of able to think like humans do, basically, or mimic human cognitive functions, as it says here, learning, reasoning, perception. The second thing is analyzing vast amounts of data to really kind of understand if there's patterns in the data, what relationships exist between the data, and then using the data to make decisions, and then taking the data and then actually performing not just simplistic tasks, but complex ones as well. So, again, this is what, what came out at me when I asked ChatGPT, what is artificial intelligence? Paige or Jill, any thoughts? Is this fair, or would you add anything to this, or take anything away from that?
Paige Bailey [00:08:59] Yeah, so… so one thing that I, that I might want to, to, to think about a little bit is this, this feels like it's very, very much focused on perhaps embodied intelligence. So, like, like, when I hear a machine, I think physical machine. As opposed to maybe some of the more, digital kind of automations that we have. That we have running today. So, so I, I feel like, when, when I imagine AI systems. I imagine, you know, how are we kind of incorporating data from everything that we've seen previously, and using that to inform decision making? for potential future events. And that, that applies to predictive machine learning. For generative AI, it's a similar sort of idea, just how do we take in all of the vast amount of data, the vast amount of knowledge. That, that humans have been capturing over time, and then use that, to kind of help, create content. So maybe it's text, maybe it's code, maybe it's images, or editing images, maybe it's video content. And, and sort of using, using that, that really, kind of firm backbone of, of, kind of accumulated knowledge to, to, sort of help us. help us create, down the future. I'm curious what the other panelists, are, are imagining as well.
George Mateyo [00:10:32] Thanks, Paige. How about you, Jill?
Jill Taliano [00:10:35] Yeah, no, I mean, I think she hit the nail on the head. For me, as I think about within, banking, I think about our ability to be able to analyze tremendous amounts of data, market data, economic reports, and actually being able to take that, personalize that information, and probably most importantly, action it. So to me, it's a very exciting, foundation, I think, to what's to come.
Mark Grebey [00:11:00] Yeah, and I'd have to agree with that, Jill, just to, you know, kind of put the KeyBank's perspective on it. You know, what makes the idea of AI exciting to us? Well, first things first, you know, AI is not new, right? we here at KeyBank are running somewhere around Probably around 100 machine learning models for various, various facets of the bank, whether it be compliance checking, consumer credit validation, cybersecurity, fraud. operations, you name it. So we've had models in place since probably the late 90s, right? It's where it gets really interesting now. You know, I think the latest innovation is the generative AI, and that kind of was introduced in probably 2021-ish, and you know, to Jill and Paige's point, large swaths of data, documents, you can very quickly, analyze this data, but do something, do something interesting with it, right? Like, I think you know, analyzing large swaths of data without having some sort of outcome or output, really, you know, there's not really a lot of value there. But, you know, what I get excited about here at KeyBank and the things that I'm thinking about is how can we leverage generative AI to From a client service perspective and a client experience perspective to make sure that we're servicing our clients, more quickly. And easier, right? And so, you know, I look at it as a way to collect large swaths of data and make it available to our employees so that they can actually service all of you better, and more, and quicker.
George Mateyo [00:12:37] On the data piece, Mark, I'll go off script here a little bit and just take a second to ask you a question about data itself. How do you… how do you ensure the integrity of that data? In other words, there's a lot of data, and maybe Paige or Jill have something to add to this as well, but it seems to me that given, as you mentioned, this huge amount of data that exists. How do you feel… do you feel… how do you get comfortable knowing that the data is good data, for lack of a better term?
Mark Grebey [00:13:04] So, who wants to take that first? It's me?
George Mateyo [00:13:06] I started with you, Mark, just because as you kind of mentioned that, you know.
Mark Grebey [00:13:10] Yeah, like, any computer system, garbage in, garbage out, right? To use the old adage, I think. You know, developing a… developing confidence in your data is kind of core and fundamental to being able to leverage it for Gen AI purposes or for AI purposes, right? And so, here at the bank, we have a data governance model That has identified and documented the lineage of, what we call critical data elements here at the bank. And, we're doing this for regulatory purposes, but, basically… We have an entire program here of, and an entire organization that's dedicated to data quality and ensuring that we know lineage. So we feel confident. At least at the moment, and that we have high-quality data that we can operate on. But in areas where, you know, we may not have such high-quality data. that's where I think the, you know, still the human interaction, as we're training models that are using data, we depend on humans to… You know, help train those models up and help identify those areas where, you know, if it's either a data problem or a large language model learning problem, so that we can resolve those before we actually operationalize it for general use.
George Mateyo [00:14:25] You also mentioned, Mark, I think this one's really important, and kind of moving beyond just the definitions or so, I think it was kind of interesting, as I was preparing for this conversation, to really kind of go back and think about how we kind of came to be where we are today. And you referenced the fact that AI has been with us probably longer than many of us realize, in the sense, I think, that there was a scientist back in the mid-1950s or so that thought that computers could learn how to think. And again, that seems, you know, something more forward-thinking, but again, it was, you know, several decades ago. After that moment first happened, there was a lot of experimentation that took hold. Things that kind of blossomed in the 60s and 70s really catalyzed the movement in a big way. there is kind of interesting. It seemed like, from what I read and what I've been, learning about as it relates to artificial intelligence, that there's a period called the AI Winter, where things were kind of stalled in terms of development. Not a lot of progress was made, and things kind of… slowed quite a bit, as far as I could tell. But then somewhere around the 1990s and 2000s, when IBM's Big Blue took stage, and I think, took down a significant chess master and really dominated the fact that… it showed the fact that I think computers could learn and could be used in certain applications, that we had a bit of a revival in that period of time, and then things took off even further in the 2010s. As we say here, AI has started to rise with the advent of Watson and Siri and Alexa and other things that many of our audience members probably know firsthand, but now we seem to really kind of take this into overdrive in the last few years, literally, with the advent of ChatGPT, DeepMind, things that you're working on, Paige, for example. I'm sure there's many on this list, many other things not on this list that we could talk about. But it was really quite interesting just to see this evolution, again, knowing that this started several decades ago, and has really catalyzed itself in the past couple years. And indeed, if you think about just the level of adoption, I'm sure many people have seen certain versions of this chart in different forms, just showing the number of months and years, rather, in which critical mass was attained. Critical mass may be being defined as 100 million users, and if you look very closely and maybe squint a little bit at this page. you'll see on the top, top part of the screen showing that JAT GPT relays over 100 million users in less than 3 months, which is really quite staggering, and you compare that to other technologies. things like Facebook and Spotify, Uber even, you know, Netflix, other things we're more familiar with. This, this advent, this adoption, this rapid adoption was really unprecedented. As you can see quite clearly, and that also shows up here on this next slide. I was kind of intrigued by this as well, just to show the number of people using AI, the number of searches, you know, within 2 years, we were at the point where Google is, with all due respect to my colleagues at Google, that within, you know, two years, we saw the same level of activity in 11 years in terms of the sheer amount of searches that were conducted on AI's platform. So, maybe if, just for a second, in your mind, Paige, what do you think has kind of catalyzed or kind of led this adoption to be so rapid, and what's kind of behind the more recent explosion in AI adoption overall?
Paige Bailey [00:17:21] I really do feel, like, if we go back to that timeline slide that you had a little bit earlier, I think one of the important things to note for some of these, for some of these events, right, like in the early days when, John McCarthy and Marvin Minsky were thinking about kind of symbolic AI and kind of this more heuristic, like, rules-based system. it was still really challenging for everyday humans to know how to build AI models or build these systems, or even to interact with them. You know, you probably had to program, you know, perhaps in the early days using something like punch cards. And then, you know, evolving to assembly, and to basic, and then to, around when I first started, things like You know, you might have been using C++ or Python, but it still required you to have a lot of background knowledge, it required you to often kind of accumulate the data and train the model yourself. To know how to program. And honestly, those are all kind of roadblocks, that, that kind of get in the way of doing the important thing, which is helping people solve problems, or helping people, you know, do something useful, as we were talking about just a second ago. And I think that part of the wonderful thing about these AI-enabled chat applications, like Gemini or Claude or ChatGBT, is that they really unlock the door and kind of empower and democratize AI in a way that we've never really seen before. You know, instead of having to learn something as archaic as a programming language, you can just speak naturally into a phone, or speak naturally into a chat application, and ask for it to create a web app, or to automate a workflow, or extract out insights from a PDF or a slide deck. And I think that's the magic cold piece, is that…
Jill Taliano [00:19:22] You know, these models, or at least Google's models, I don't know if it's the same with some of our competitors, but they understand hundreds of languages.
Paige Bailey [00:19:31] So you can speak, just with audio, with, you know, typing in text, and suddenly be able to, to do things that previously you might have had to have a PhD and, like, 5 years of programming expertise in order to do. So that's why I think that what we're seeing now is really, really unique and really magical, is because, you know, all of us are suddenly, able to use these AI tools. They're not locked in some ivory tower, with only a few having access.
George Mateyo [00:20:03] Great perspective.
Jill Taliano [00:20:04] Yeah, and I might add on top of that, I think the fact that it's made it so accessible to, nontechnical people, I think the other part of it is we've really gone from the shift from kind of just analyzing data, to being able to actually create… and images and everything else, and sound, right? To creating it and actually reasoning and working with it. So that's one of the significant changes I think we've seen, really, over the last, I'm gonna say. 18 to 24 months. Folks probably see that if they go online and do a Google search. You know, today you have the ability to interact, with Gemini or with an AI engine. You see the same thing with GPT and some of the others. But it's really allowed you to interact with information very differently than we could just a few years ago.
George Mateyo [00:20:51] Fantastic point. Do you think, in Google's mind, is there a certain, Use case or use cases that are really interesting to you right now, as you think about where this is going. Maybe for individuals and businesses alike, what would be some of the bigger use cases that you're imagining or envisioning as you look out years from now?
Paige Bailey [00:21:11] So, so I am… I'm happy to start, and then I'd be really curious to hear Jill's perspective on this, too. So, so working at DMind, it feels a little bit like working at Disneyland, honestly, in the sense that you, you go into the office, and it's like, oh, there's a robot, like, or, or, oh, like, there's, you know, selfdriving cars just kind of, like, roaming around.
George Mateyo [00:21:32] And it's, it's really, it really does feel like you're, you're living in the future just a little bit.
Paige Bailey [00:21:39] The… I think the, the pieces that I'm most excited about, are… are both kind of, you know, kind of touching on this idea of embodied intelligence, so… You know, our Gemini team has a model called, you know, Gemini 2.5 Pro, and then an implementation called Gemini Live, which allows you to just have a conversation with the model, and Like, the same as you might have a conversation with a human. And Gemini Live is actually used by our robotics team to kind of build plans for the robots to accomplish, and then kind of assign different, actions, so you could tell the robot, like, hey, go make me a salad. Or hey, go clean up that spill in the corner. Or, hey, go wash the dishes, which, like, cannot come soon enough, as far as I'm concerned. Like, make the bed, wash the dishes, like, sweep the floor, like, sign me up. But that is… that is something that I'm really, really excited about, both for our own robots, as well as, you know, our partners at Boston Dynamics, and Everyday Robots, and Enchanted Tools. And then the second thing is the media generation piece. Like, I've been really loving playing with Nano Banana, which is one of our latest models, which allow you to do image generation and image editing, and then, of course, BO3 and Genie3, which allow you to do, like, really interesting video generation and kind of playing around in explorable worlds. So those are… those are the two things that I'm most excited about.
Jill Taliano [00:23:10] So, for me, I think, first of all, I'm super excited that Paige gets to live in kind of the art of the possible world. I end up coming back to what I'd call the art of the feasible, so what things we can do today. So I'm very excited about what's coming, but I think in terms of on a personal basis. I look at, kind of, the personal assistance and productivity pieces. She talked a little bit about, kind of, the creativity and expression side of it, but also, I'd say, accessibility and communication. So, I mean, today, people probably use it for search and discovery, or folks are familiar with virtual assistants, asking, you know, Siri or Alexa, Google Assistant, you know, a question for something, or they see it filling in a, you know, completing a sentence in email. But what about things like, you know, I use every day, transcribing, you know, and summarizing meetings, writing up next best actions? For me translating. I mean, I used to have to use translators for, you know, for contracts and things. to be able to actually use an app today to do that translation. You know, while I'm not using an automated car yet, I'm not using Waymo, I do use Maps and Waze every day. But I'll even give you, you know, an example on a personal basis. You know, I've got an aging parent. They're looking at, you know, retirement communities, and that's a super overwhelming activity for them to do, and to be able to take that and say, okay, well, let's come up with some criteria, let's figure out what your priorities are, and to be able to put something like that out to Gemini and say, you know, help me find a retirement community with a certain set of criteria. I'll say in-house memory care within 10 miles of where I live, assess all the things that are available in my area, build me a Matrix. And then tell me what I'm missing. What didn't I ask for? That took something that could have been extremely overwhelming, for my parents, and my dad happens to be an engineer, so he loved that process. He was able to take something, you know, natural language again. He didn't have to know how to program, and actually be able to build out, and I'm happy to say he selected a place, but those are real-life examples. I mean, there's lots of fun stuff, too. You know, creating storybooks for your kids, or, you know, creating your own video games, or planning trips, and things like that, but I think that, for me, the productivity enhancements in my everyday life have been tremendous. I see those same types of activities with clients as well.
Paige Bailey [00:25:39] Yeah, and when you mentioned that earlier, Jill, my mom… my mom is also, so she's, she's in her 80s right now, and is looking for… and she's healthier than I am, honestly, like, doing yoga, like, Pilates all the time. But she's, she loves living in, kind of these senior, kind of living… areas and is thinking about, kind of moving from where she currently lives in Texas to one that might be a little bit closer, to where I live. And when you mentioned that you had had that with your dad, using deep research to kind of investigate new places. She did the same thing with, just in California for senior assisted living houses. So it's, it's a great, great idea.
Jill Taliano [00:26:23] Yeah, absolutely, even, you know, for anyone that's got school-age kids. I mean, I'm researching colleges now, and so it's just… it's a tremendous tool, that actually renders information very quickly, helps you think through things. Because you don't know all the questions to ask, right? And it'll prompt you for things that you didn't ask or that you didn't think of, and then just that continuous iteration and ability to interact and continue to refine and refine, it's amazing, I think, the productivity that can come out of things like that.
Mark Grebey [00:26:54] 100% agree, and, you know, just to kind of tag on to that, I like the art of the feasible Jill. I think I'm gonna take that and internalize and begin using it here at KeyBank, but, you know, if I look… you know, as a technology leader here, and a consumer of AI, you know, the things that I get excited about are, well, first things first is, like, the glasses I'm wearing now actually have AI built into them. And so they've become a…
Jill Taliano [00:27:20] ubiquitous part.
Mark Grebey [00:27:21] my life. I can interact with Meta at any given time. And ask it questions, and the first thing I do every morning is when I put on my glasses, I ask. what's the weather gonna be like in Cleveland today? And then that helps me figure out what I should be wearing. I also ask, what are the top 3 national headlines? What's happening in pre-market? So I… I have a way to do that way more effectively than, you know, than back in the day, where you would, you know, read a newspaper or, you know, wait for the weather report.
George Mateyo [00:27:48] Wait, what's a newspaper? What's a newspaper, Mark?
Mark Grebey [00:27:51] Oh, well, yeah. So, you know, I have it, it's become basically a ubiquitous part of my life, just in a personal level, and it's helped me become a more efficient person. Another personal, advancement that I'm really excited about is, you know, it would be nice to have the hour and a half that it takes me every week to cut my grass. back, and I know that there are plenty of products on the market that you can train up that will do that for you, I just haven't… Splurged on making those purchases yet. Now, as far as professionally, and how we can, you know… from an Art of the feasible perspective, make it productive here for, you know, for KeyBank, there are many ideas that we have in the hopper, that I've been working on advancing. I'll try to do something contextual for the people on this phone call. A lot of you are trust… are clients of our trust business, and so, you know, we hold, we hold your trust documents, and we might have, some of, you know, your wills and some more legal documents. Well. We have around 2.5 million, you know, documents associated with trusts in our repository, and back to the, you know, kind of the idea of, you know, being able to, look at large swaths of data, I believe that we can, we can actually, train a model that sits on top of that, those documents, and completely and simplify… completely simplifies the ability to find the document, to make sure you've got the right one. To look at the discretionary language in it, and to answer the question that seems to be on every trust holder's mind now. Can I… can my… can my kid buy a Corvette with, funds from this trust? And it should be able to be able to produce an answer. I've… I've kind of jokingly named that agent Varo, which is the first, the first ever commissioned librarian, commissioned by Julius Caesar during the Roman Empire. So we are working on the idea of making Varro real. As, as we speak. And again, to bring that back to the client service perspective, we think if we can somehow make that large language model work against that 2.5 million document repository, we can bring a And a level of efficiency to client interactions as they're trying to get information about their trusts. That, you know, think about how much easier it would be if Varro could go get the document for you. Varro can confirm it's the latest one. Varro can get the discretionary language. You can ask it questions about discretionary language. You can ask it questions about distributions. If we can operationalize that, I think our, you know, our trust office becomes way more efficient, and in turn, you know, makes it easier for all of you to interact. So that's one example of something that I'm working on now here at the bank.
George Mateyo [00:30:38] That's great. I'm going to stop sharing the slides so we can have a nice, intimate conversation on these next couple topics, because we have a, you know, all this stuff sounds really exciting and innovative and, as you said, Paige, kind of, like, in the toy room with Disneyland and metaphors and things like that. I think it is pretty… cool and exciting stuff, and some really practical things too, but we did get a lot of questions in advance, and also, since the call started, just around the notion of, again, data integrity, as we rely on more AI for more things, we're more dependent upon algorithms. Algorithms have some biases, they have some shortcomings. How do you think about that, Paige? Or Jill, if you want to weigh in first on the Google perspective, and then Mark, I'll ask you to kind of reflect on the same question, but the key takeaway is how do we kind of get comfortable with the risks, and maybe what are some of the concerns as we become more dependent upon AI?
Jill Taliano [00:31:30] Sure, so if you want me to go first. Yeah, I mean, look, I think that there's, you know, AI is not perfect, right? Certainly, there are risks inherent in anything that we do here. We can actually use it to turn the table a little bit, though, too, right? So, in terms of using it for fraud, using it for, you know, malicious intent, being able to detect cyber attacks, so for as much as there are risks that are involved, there's actually opportunities to leverage it to counter, in particular, security risks. So I think that there's lots of ways to be able to leverage that, today, but in terms of bias, biases, right, I mean, I think that humans are flawed. Data can be flawed, so anytime that the data contains, any historical human biases, it is absolutely possible that AI will make a mistake. It could even amplify those, right? And I think that that's why, as we look at things like, we call them guardrails, but being able to put guardrails in place that will protect us as much as possible. We talked about training, the models to do that. I mean, today we still have a lot of human in the loop. you know, our human-in-the-middle type activities, that help us look at those things and try to detect those and put measures in place to, to prevent and reduce them. But I think there's absolutely possibility for that. AI is not infallible.
George Mateyo [00:33:04] For sure. How about you, Paige? Any perspectives on that?
Paige Bailey [00:33:06] Yeah, I just want to plus one what Jill said. I do feel like, I'm very proud in that at Google, when we train our models, we start with responsible AI from, like, t equals zero. Like, we're thinking about it as people are acquiring the datasets, as they're… Kind of looking through, making sure that they're the right kind of information to put into the model. And then, we also have adopted best practices around as these models are getting served, so that they can be used by us and all of our downstream, applications, like Docs or Sheets or YouTube, but also through all of our customers that, that they have guardrails in place to help make sure that folks are using the models appropriately. So these could be, anything from, kind of. you know, having configurable safety filters around specific kinds of information to helping people know how to configure their system prompts so that, like, certain kinds of outputs aren't acceptable from the model. And even things, like, for multimodal image generation. we have a project called SynthID, which is kind of an invisible-to-the-eye watermark that, that allows you to spot AI-generated visual content. So I'm sure that if folks have been on social media anytime recently, you've probably seen a whole bunch of generated videos or images. being able to have a way to kind of identify that that's been AI-generated behind the scenes, even if it's not visible to the human eye, is something that, again, I'm super proud that Google is pioneering and kind of… Creating as a standard, or helping create as a standard that other folks can adopt.
George Mateyo [00:34:58] We did get a couple questions on that last point, Paige, you mentioned, which is how do people… how would a layperson without a whole lot of knowledge, like myself. how would he be… how would I be able to discern what is AI-generated versus what is not? Are there certain, kind of, takeaways or certain things you'd look for or have us identify if we wanted to discern what is generated by AI and what is not?
Paige Bailey [00:35:19] Yeah, so, so some, some hints are, you know, just kind of looking at something visually. Whenever we use Imagine or, the Nano Banana model, the Gemini 2.5 flash image preview, to create an image, there's a human visible watermark in the lower right-hand corner. But then there's also that, that digital watermark, and so if you, if you take that image and kind of check. To see if it's been, to see if it includes Synth ID, then that's a good way if you have access to the image. Another way is that I think that probably all of us have seen enough text that's been generated by these models to kind of pick up on some things that are pervasive through the model vernaculars. Like, that's a great idea. Or, not only did they do this, but also, you know, like, these sorts of common patterns are good indicators that models developed the content and not a human. I see the same thing for, kind of bullet points. You know, if you see, like, a series of bullet points, a text paragraph, and then there's bold text for the bullet points, then that's a good indication, that it might have been AI-generated content. We also have been starting to do text watermarking.
George Mateyo [00:36:43] For, for AI-generated content, there are ways to, to kind of.
Paige Bailey [00:36:48] Put into the outputs or configure the outputs to help people understand if the patterns are indicative of a model generating the text-based content. So there, there are ways, and we are improving in terms of our ability to understand, but it's, I encourage everyone to be vigilant, to ask questions, to double-check, and to verify information. We've also, in our Gemini app, incorporated something called citations, so if you ask for… if you ask a question and the model gives an output, it will give you citations, where you can go and double-check and verify the sources that Gemini has shared. So, but those are, those are all just some techniques, and it's not just, you know, for all of the folks on the call, it's not just y'all who are trying to wrangle, with understanding, it's, it's everybody. And so knowing… knowing that we need to all be watchful is the first step of the battle, and so we've just gotta, like, know what tools are at our disposal and kind of pay attention.
George Mateyo [00:37:49] That's super helpful, thanks for that. Let's maybe shift gears a little bit now, and again, we had a few questions in the chat that just popped up in the last 30 minutes or so, and also ahead of time, and it relates to just the sheer infrastructure needed to really drive and power AI. And we have a few questions around, you know, what does that mean for the environment, what does it mean for sustainability, and that's a fair question as well. Also, just more broadly, people are kind of curious to know your thoughts or our thoughts on does the U.S. have enough power and electricity, and can we actually make this work? I mean, we've got… there's a huge amount of infrastructure that needs to be built, and is already being built. I'd be curious to get your thoughts, Jill, maybe I'll start with you, to see if you have a view on, can we meet the need, can we meet the demand based on the infrastructure we have in place today from the energy perspective?
Jill Taliano [00:38:35] So, yeah, so I think from my perspective, it's certainly one of the biggest, I'd say, physical constraints. You know, in terms of space and capacity for the data centers, in terms of the power. Obviously, we're looking at, you know, renewables and natural gas and nuclear and other ways to build, you know, build out and fuel our… the kind of the huge grid that's being used today, so I'd say infrastructure. and energy are both there. You know, one thing I might mention, so Google actually, I think we were the first to release a tech report, because a lot of people say, well, how much, how much power is being used? It's being consumed. So, and I'm gonna… I hope I get this stat right, but on, an average query, right, which is an average question that goes in, it's about a quarter of a gigawatt of, of, electricity, so what does that mean? That's like powering a microwave oven for a second. Right? So if you think about that, but then think about the volumes of, whether it's businesses, I mean, people, I mean, ChatGPT is huge for students, and things, right? So it's not just companies and large businesses that are doing this, it's the general consumer, you know, market as well. So I'd say, like. I think… I think we might be challenged right now, and the outlook and the growth that we're seeing, is massive, right? I mean, they're talking about 25-plus percent increases, you know, over the next 5 years. I mean, that's massive if you think about, the data and the data center growth over the last 10 to 20.
Paige Bailey [00:40:10] And I will say, Google is… I believe we have a commitment, to reaching net zero emissions on our infrastructure by 2030. So, so Google does have a firm commitment to operating sustainably, and we also have kind of one of the, the lowest power… kind of consumption… it's called PUE, so our, like, PUE, our power usage effectiveness is kind of leading in the industry, and has been dropping significantly from when we first started measuring. There's also, you know, kind of techniques that you can do during the model training and model serving stages, so things like quantization, that, that we've been investing in, so you can have smaller models, which means they need less energy to run, but without degrading the model quality. So, you get the same or near the same outputs, but instead of needing, you know, like, 8 TPUs or GPUs to run, you just need one. And so Google has been investing in those kinds of techniques for, you know, many, many years, and I think that We're going to keep investing in them, even as we kind of diversify and kind of explore different energy options. And full disclosure, like, I started my career in the oil and gas industry, like, oil and gas.
George Mateyo [00:41:42] Ronnie.
Paige Bailey [00:41:43] Yep. Oil and gas is definitely, like, top of mind, I think, for a lot of the data centers at the moment, to figure out, you know, how to get the energy to power all of the world's… all of the world's models and inference calls.
George Mateyo [00:42:01] Mark, anything you'd like to add?
Mark Grebey [00:42:04] Well, I'm gonna depend on the brilliance of folks like Paige and others at Google to figure out how to do this sustainably. But I did want to touch on one thing. about biases that I just really didn't get a chance to bring the KeyBank perspective. You know, one of the things that, you know, the reason that, you know, one of the reasons why we've been, you know. pretty, conservative with, with our, you know, rollout of, of Gen AI here at the bank. Well, first and foremost, I think the regulatory guidance is still evolving. And so we, you know, the only thing that we know for sure right now is we need a counsel, and we need a process that That, approves, whatever models that we want to work on and introduce. So, we have that in place, but… You know, back to the biases idea, one of the reasons… one of the things that we really wanted to make sure was true before we started, you know, kind of, unleashing the creativity within our organization to leverage Gen AI was… There were two things that were concerning to us. Number one, auditability. And I think, Jill mentioned this, or maybe Paige with, you know, with the citations, but, You know, any sort of responses that come out of an agent or a model, they need to be auditable. And so we spent a lot of time and energy making sure, you know, that we were satisfied with the auditability and consistency of results, and when we… when we, when we put models out there. The second thing that we were really concerned about that we've since moved past is, A term that they use called hallucinating. You know, back to the original point of large swaths of data, I mean, You know, you want to make sure, you know, given that we're a bank, we are very security conscious. We are… we have a very conservative security profile here. Obviously, we're in the business of, you know, technology, and when you're dealing with, you know, with real money. you have to have that, and you have to be conservative in that way. And so, you know, one of the things that we really wanted to work through, and our Google friends have helped us, is with the idea of, you know, trying to reduce and or eliminate hallucinations. And what hallucinating is, is when an AI model or an agent somehow brings a data or a data set into its thinking that it wasn't supposed to or didn't have access to. And so, I think we put a lot of time and energy into that as we were building out our Gen AI platform to ensure that we're, kind of controlling, the guardrails and controlling the lock and key to all of our various data sets so that we don't introduce hallucination into the models that we use and put our clients' data at risk.
George Mateyo [00:44:49] No, those are really helpful perspectives, and I think that addresses a number of questions that we received ahead of time, Mark, and also some things in the chat as well. You know, I think we also get some questions on the more global perspective about what does this mean for the economy, right? What does AI mean for the workforce? And I think Jill and Paige, you both addressed just the sheer number and the sheer enhancements to productivity that will likely come from AI. We gave a lot of great examples around those type of things, but there's probably also a cost, too, and I think people are right to suggest, or right to ask about at least, that maybe some significant job displacement might be… might occur as well. You know, there's certain It seems to me that there might be certain industries or certain occupations that might be rendered less relevant, or maybe even obsolete because of AI, and I would argue that we've seen this happen time and time again throughout our history, frankly, that we see new innovations take hold, and frankly, some jobs are displaced and done away with, but other new ones emerge. So, with that in mind, I'm curious to get your thoughts, Paige, you know, maybe I'll start with you on your thoughts on, or Google's thoughts on. Job displacement, any economic implications from AI, and maybe where we see that going over time as well.
Paige Bailey [00:45:59] Well, of course, I don't speak on behalf of Google, but I can at least talk through some of the things that I've seen in terms of accelerating the software engineers that we have. At Google today, and we have seen, like, pretty, comprehensive adoption of AI, assistive AI tools across our entire software development lifecycle. So, everything from. Designing new features and applications to actually doing the work to build them, to deploying them, maintaining them, and, kind of also monitoring for, for things like, for things like security. I… if anything, I… I've seen us, kind of, ship features so much more quickly, so on average, we're releasing new model updates or new feature updates for the Gemini APIs around once every 5 days. So it's, it's really, really rapid, the, the amount of work that we're able to get accomplished now. And we're still hiring engineers. Like, we just find that a single engineer can do so much more just by themselves, assisted with a fleet of agents. Like, as an example. one of the engineering managers, I used to work with, in the machine learning framework space. Had a team of around, I want to say, like, 70 or 80 engineers that were building machine learning frameworks. They, they then, decided to step back into an IC role, and since then, they've been managing, effectively, a team of 10 AI-enabled agents. To, to have just about as much impact as that size of a team, but, but on a completely new, like, greenfield project. And, and so I, I think that we'll be seeing even more of that going forward. Like, there's a really high percentage of startups today that are solo founder-led, and we see that solo founder being able to do product marketing, being able to do engineering, being able to do customer support. assisted with these AI tools, but being able to, but being able to operate, as just part of a small team or as a singleton, as opposed to, as opposed to needing a large number of humans in order to accomplish the same amount of work. And it's not so much that you know, those jobs are getting displaced as much as, like, we're seeing so much more innovation, so many more projects getting deployed, so many more people who are able to spot ways that AI could help assist them or help kind of enable their business. And now they finally have the space and the time to tackle all of those P2s and P3s on their roadmap that, you know, otherwise would have never gotten solved.
Mark Grebey [00:48:52] Yeah, Paige, I couldn't have said that better myself. I mean, I think in, you know, from an organizational perspective here at KeyBank. You know, we realize that financial services is fertile ground for AI advancement. Just to give you kind of a little bit of a background, we… you know, I… we typically go through a capital planning process every year, and the reality is we only invest… we're only able to invest about… into about one-third of the, you know, the things that we want to do, right? So that implies there's about 66% of things that we want to do that we just don't have the time to do, and I think… you know, just given the efficiency that AI brings to everything. You know, as we adopt that more and more, then we'll still be able to, you know, we'll be able to do some of those more greenfield projects that we ought to do that we just simply don't have the capacity to do now. So I think, you know… in the long term, efficiency is the play, and it enables the people we have to work on, you know, work on more ideas by virtue of gaining efficiency in their own everyday lives. You know, that's what I think the promise is. Now, we just have to, you know, we just have to find a way to deliver on those ideas, but yeah. That's my perspective, I agree with Paige here.
George Mateyo [00:50:08] I don't see it replacing people, I see it, enhancing what those people can do. Excellent. Thank you all for that. We also have a few questions that come in… have come in around the topic of legislation, you know, kind of public policy issues. I wonder… I don't know, maybe, Jill, I'll throw this one to you for starters. Any thoughts on just where legislation around AI is headed? Anything that we should be mindful of as regulators get involved, potentially, or more involved? So I'd say it's absolutely an evolving, space. You know, my personal experience, I do a lot with fair lending, for example, and I would say it's, it's…
Jill Taliano [00:50:46] challenging. They're… they're absolutely, I don't want to say over-conservative, because I'm not sure you can be over-conservative when it comes to, you know, to banking and our money and our investments. But it is absolutely challenging, right? I think the rigor, that is there today, I think we'll see more. We're seeing, you know, a lot of legislation, and regulatory coming out of the EU. You know, sometimes before it comes out of the U.S. So I absolutely think we'll continue to see more. What I would generally say is, is all the banks that I work with. When it comes to, you know, risk, compliance, audit, I think go generally above and beyond what's actually required of them. Their reputational, you know, risk cannot be underestimated. So, while there's lots of regulatory, you know, I think, things that are coming down the road, I would expect that the banks, KeyBank, you know, for sure being one of them, that's going to go above and beyond to make sure that we're, you know, really in lockstep with your risk appetite and your client's risk appetite.
George Mateyo [00:51:53] Super. Thanks very much. Yeah, I'm sure it is going to evolve a lot in the next several years as well, so TBD on that. We did get a couple questions also just around what we're doing in the investment office. with respect to AI, and the short answer is a lot. We're doing a lot of experimentation. We've used AI in certain applications to help us clean through a lot of different data sets to look for patterns, and that's, I think, where AI can really be helpful to us. It doesn't make decisions for us, so it's actually a supplementary tool to what we do with human brains, if you will, but it's an important part of our work as well, just given the sheer amount of data that exists, so I'm sure we can cover that in another session, but I do want to address that, because there are a few questions around how we've integrated AI into wealth management. Mark, you spoke about that directly with the applications in reviewing trust documents and so forth, but there is a good bit of effort being put forth into our investment decision-making as well, utilizing AI tools. I know we're getting close to the top of the hour, so be mindful of everybody's time. Maybe I'll just close and ask you each to think about one piece of advice you might give to our audience in terms of how they actually would use or consider using AI, one thing they might consider incorporating into their lives, or any kind of useful hints that you might offer our audience. Around the use case of AI in their everyday world. So, maybe, Mark, you want to start with that one? You ever talked about the glasses, anything else you, you're really fond of?
Mark Grebey [00:53:19] Yeah, I mean, I definitely would encourage folks, if you haven't, just to go pick up a pair of AI glasses. They're really something. You know, if anything, you know, I don't have to depend on, calling friends anymore on my way home. I can just have a conversation with Meta and talk about things. But it's a way to… feel AI on your person, and interact with it, and get familiar with it, and figure out how it can help you. I think it's been… you know, my original use case, I bought them I was taking a trip to Japan earlier this year, going to Tokyo with my family, and I'd heard these things actually did, translation. Now, unfortunately, this particular version doesn't actually translate Japanese, but it does do, German and Italian and Spanish. But as the technology gets better and better, I think having something like this in a… You know, if you're trying to speak with someone speaking a different language, it'll… it'll be, it'll be extremely valuable, but I would encourage everyone to, you know, consider getting some AI glasses and learn how to interact with them.
George Mateyo [00:54:27] Excellent. How about you, Jill? You've talked about that a little bit with your parents and your father in particular. Anything else you would recommend?
Jill Taliano [00:54:32] Yeah, I mean, I would say be curious. Be curious, try it, don't be afraid of it. Find one thing in your life that… that you feel like has an opportunity to… to either be more efficient or you need help with. It could be managing your five kids, you know, sports schedules. Whatever it is, but… but try it, because I think, you know, the, it's not as scary as you might think it is, and I think you'll become addicted to trying to find new things in your life that you can make more efficient. I'm constantly doing research on banks and clients and industry trends and things that are going on. I don't have time to read, you know, the 10K for every client. I download it, throw it into Notebook LM, have it write me a summary, and I listen to it as a podcast when I walk my dog. Like, every second of the day gets utilized, and so for me, that's an efficiency opportunity. But I'd say try it, test it, you're gonna love it.
George Mateyo [00:55:26] Alright. Paige, final words?
Paige Bailey [00:55:29] Yeah, I feel exactly the same way. I also encourage folks, if you haven't already experimented with it, the live feature in the Gemini app. It supports over 100 different languages. You can have conversations with anything that you can, you can kind of show, show Gemini via the app. You just pull out your camera and kind of have a conversation with the model. It's hooked up with Search, so it can, accomplish tasks for you. and do kind of these deep research activities that, that we were just discussing. Strongly, strongly recommend folks, folks play around with that, and, and I also feel like. You know, similar to what Jill was just saying, I feel like every single minute of my day has now become hyper-optimized. You know, I am using Gemini to help answer emails, to help do research, to help write code. To help, kind of figure out what are the top priorities for people that, in terms of. Bugs or new feature requests that need to be fixed. It really does feel like, like I'm getting time back, and time is the most precious resource that any of us has.
Mark Grebey [00:56:43] Hey, Paige, when do I get my Gemini glasses?
Paige Bailey [00:56:46] Well, we announced them at I.O. It's just a matter of when they're available for folks to start using. We were super quick to the market with Google Glass, and so now we've gotten to a place where the models are good enough to power them.
Mark Grebey [00:57:00] I'm excited.
Paige Bailey [00:57:02] Yeah.
George Mateyo [00:57:03] All right, that's a great, great note to end on. So, hyper-optimization sounds like a really great goal, and a noble pursuit. I do want to thank you all for your insights, Paige, Jill, and Mark, this was a fantastic conversation. We could probably go on for another few more hours, given the fact we just probably skimmed the surface, I think, and I think there's probably a lot more to talk about, so maybe I'll have to invite you back sometime later this year or next. But, thank you so much for all these great insights. They were fascinating, and I thank our audience for listening. The call is being recorded, so we'll make sure the reporting goes out to those who missed it or weren't able to listen to the full program. Thank you for being a client of ours as well. We appreciate your business, and we look forward to staying in touch and keeping you updated on many other topical items, such as artificial intelligence and other things as well. Thanks, everybody. Enjoy the rest of the day, and stay well.
Paige Bailey [00:57:54] Thank you.
Mark Grebey [00:57:55] Thanks, George.
Artificial intelligence (AI) is no longer a futuristic concept—it has become a present‑day force reshaping industries, economies, and everyday life. View the Key Wealth webinar replay for a dynamic panel discussion featuring experts from Key Wealth and Google, as they unpacked the evolution, mechanics, and societal impact of AI.
The discussion focused on:
- AI – What is it? How has it come to be? What lead to its rise, and why has it become such a big topic in the last 2-3 years?
- How does it work? What are its benefits, and what are some of its limitations?
- What are the implications for the economy and society as a whole?
Speakers:
- George Mateyo, Chief Investment Officer, Key Wealth
- Mark Grebey, Wealth Management Technology Executive, Key Wealth
- Paige Bailey, AI Developer Relations Lead, Google DeepMind
- Jill Taliano, Generative AI Specialist, Google DeepMind