AI 101: Two Demos with David Hyde and Leah Price | DC Fintech Week 2025

[Music] All right, good morning everyone and welcome to day two of DC FinTech week. Uh or as we like to call it, Finchella. Uh and this is the application layer presented by Gibson Dunn. Uh I'm Jeff Steiner. I'm a partner at Gibson Dunn and I'll be your MC for the day and I'll moderate at at least one panel.

Um so yesterday we laid the groundwork exploring infrastructure innovation and policy frameworks shaping fintech's future. Today we're going up the stack and uh we'll see how these technologies are actually being built, governed, and brought to market from AI native products to the evolving playbook for governance and the rise of the oneperson unicorn. It's a day about application, where big ideas meet reality, and where theory turns into code, compliance, and creativity. And we're going to kick things off with two fascinating demos uh that show exactly what that looks like in action. So, we're going to start by pulling back the curtain on AI itself.

Everyone talks about AI models and apps, but very few understand the infrastructure that makes that all possible. From chips to data centers to APIs. To help unpack this, we have assistant professor David Hyde of Vanderbilt University. David has been on the cutting edge of research in how the back end of AI systems really work. Please join me in welcoming David Hyde.

[Applause] All right. Thank you so much, Jeff, for the uh the warm intro. I'm super excited to talk to everybody today about AI. And uh look at that, we've got slides working as well. Fantastic.

Um awesome. So, let's jump right in. So, one of the things uh that I love to highlight about artificial intelligence that a lot of people don't necessarily realize is AI is not new. AI is something that's been going on literally, you know, 75 years as of 20 2025. Um Alan Turing, one of the uh you know, one of the original godfathers of computer science and certainly AI, the Nobel Prize in computer science is the Turing Award named after this guy.

Um he was working on this stuff in 1950 publishing papers. If folks have ever heard of the Turing test, this was his his paper that published the Turing test. the idea of if we had a computer hypothetically that could give you answers that could converse with you that could you know act like a real person could you tell whether something was actually a person behind the machine talking to you or was it something artificial that you know resembled intelligence so again far far before the era of you know chat GPT and all these things um this is already something on people's minds so one of the big things I want to cover is sort of what changed right? 75 years ago, um you know, shortly after World War II, we already had ideas of, well, wouldn't it be cool if we had, you know, systems like this? A and there was a lot of early math as well in the 40s and 50s being developed, um that laid kind of the groundwork for the neural network models and stuff of today.

So, why didn't we have all the stuff in the 50s? Why has it taken us so long to to kind of get here? And I argue it's three main points, okay? So, so bear with me. We'll get try to get through all three here.

The first one, I'm a I'm a nerd at heart. So, so math uh is is extremely important u to the development of what's gone on these past few decades. One of the interesting points that you know sort of you need to really kind of be deep in the weeds to understand the subtle differences here is that there are all these different terms people throw about. People talk about AI, machine learning, deep learning, whatever. And these are actually subtly different things that are tied to the development of the math uh that happened over the the years.

So originally back in kind of like the 50s 60s7s um artificial intelligence was kind of the the big catchphrase. And what original like true AI was really all about was trying to teach computers rules. If you think about like you know playing chess or something like that, right? Like a chess piece can only move in certain directions or whatever. And so if you're writing a computer program, can you you know encode all of those different rules for like how chess pieces can move and thereby constrain the program to only generate you know valid sequences of chess moves.

So this kind of idea so original AI was about you know can we define enough rules and you know regulations and policies if you will um for how a system should behave so that we can create some system that only generates sort of valid things or understandable things. Um so that was the initial effort but it turned out to be too hard like like people in the real world it's it's too complex right if I say you know give me a generic rule for every you know sentence that you might want to invent with your uh you know chat GBT or whatever too complicated so people moved into machine learning and the idea of machine learning was okay give up on all the rules and logic and formalism and philosophy and stuff like that just give me data okay I have a bunch of data points I know like I want to do some clustering task or some labeling task you know, maybe I want to determine, you know, which cities are affluent and which ones aren't and, you know, separate them somehow. Um, and so all these mathematical models were developed to say given a big set of data that you collect somehow out in the field, how do you do something with that data? Um, and so that worked really well. Um actually um you know perhaps perhaps ironically uh the the US you know federal government was one of the the cutting edge uh you know leading uh technological innovators in early machine learning.

Uh the US postal service specifically um was one of the first uh large groups to really go in on machine learning back in the day. They had, you know, you you mail mail mail a letter uh and all the letters go through the central uh processing stations of the postal system and they had these huge cameras for for the 60s or whatever um that would scan the envelopes and and figure out what zip code was written on the envelopes, right? And so they use these classical machine learning models to do digit classification and say, "Ah, this is the three, this is a seven." And so they could save a bunch of labor in figuring out um and you know, automatically routing all the mail to the right place. So that was all uh well and good.

Um but recently and actually the math is older but this recently kind of became possible. People have really leaned into to deep learning. And deep learning is really where we get into genai and large language models and all this kind of stuff. It's all deep learning stuff. The main difference between you know machine learning and deep learning if we're being really technical is deep learning allows you to do stuff with nonlinear functions.

So what's an example of a nonlinear function? How about, you know, value, you know, the value of the NASDAQ, something like that, right? Pick something from from finance. Uh, any type of, you know, quant finance signal, highly highly nonlinear. Okay, it's not just a simple, you know, line or or a set of lines.

Um, and so for a long time, people were only able to kind of do linear things and now we can do arbitrarily crazy nonlinear things. And so these were all mathematical developments. So super important. um the work that really abstract mathematicians did over the years. Another one uh big category of advancement that's a little easier and more tangible to grasp is the data side of things, right?

And so uh as I mentioned, machine learning and deep learning models depend on having a bunch of data. You just throw a bunch of data at all these mathematical models and equations and out comes some AI system. That's sort of the idea here. But to do that, you need a ton of data. Fortunately, 90% of all data ever created, every photo, email, text, video, any piece of data in the world, 90% of it is created in the last two years.

It's insane. There there's truly truly exponential growth in data. And concurrently with all that data being generated, we have uh storage costs like hard drives and things like this also getting bigger and exponentially cheaper as well. So fortunately, all these companies in the storage space have have sort of kept up with the massive massive increase in data that's being created. And if you think about it, like this data problem is just getting crazier.

You know, your camera in your phone is getting better and better every generation, right? So all of your photos and videos that you're taking up more space. Everything is on track to to create more and more and more data. And largely this is happening um now in the cloud as well. So classical storage providers now some of the cloud hyperscalers um AWS and friends uh largely contributing to the rise in data.

Then lastly my third category what's sort of enabled AI and brought us to this point. I'm going to be uh be cheeky and say it's it's video games. It's really um you know hardware. Uh but the hardware in particular that's enabled AI is is hardware that's come from video games. So if you look, you know, if you, you know, have ever played video games, if you have kids that have played video games, you know, over the years, we've gone from sort of, you know, Pac-Man in the 80s to video games that are as lifelike and, you know, as many special effects as you want, right?

Um, so the quality of of graphics in in games has increased drastically and that's a huge market, right? People people love this. um the hardware, the technology that's enabled that rise in computer graphics is GPUs, graphics processing units. Okay, so this is the thing Nvidia is is worth a trillion dollars for, right? Um so for 20 odd years, Nvidia was was hanging around the Bay Area um just serving video game customers.

Okay. Uh and that was sort of the thing. Now the trick is thinking deeply about how do how does the math of computer graphics work? If you want to draw or create some explosion or shoot a gun or something in a video game or whatever, how does that actually work? Turns out it's a bunch of math.

Okay, so there's a bunch of code and math happening behind the scenes to render those video games. There's a guy, a Stanford PhD uh by the name of of Ian Buck um who's now VP at at NVIDIA, but he was a big video game guy and circa 2005 2006 he was in grad school playing video games doing a little research whatever thinking about GPUs and he realized well hey the math that powers video games is actually like a whole bunch of linear algebra. It's matrices and vectors. If anybody's sort of a math person in here, um data when you get collect a bunch of data in the field for like a machine learning model or something for example, looks like a matrix. It's a bunch of numbers and a big, you know, giant grid of of numbers um and a data.

Um and so what Ian Buck realized was, hey, like we can actually use GPUs to just do math. Like we don't need to draw pretty pictures at the end or make a video game. these GPUs are just really really fast at doing the kind of math that's really useful for what turns out to be machine learning and data science and all these kinds of things. So he invented this language called called CUDA um which is used for uh programming GPUs uh to do sort of arbitrary computation including um machine learning and deep learning. So thanks to video games and people's interest um we ended up with this hardware that turned serendipitously truly turned out to be exceptionally useful for AI.

Okay, so those are really the three big innovations that got us to where we are today. Um what hasn't changed? Okay, some things have stayed the same. This is a photo from like an IBM machine in the 70s being used for for AI and now we have you know Jensen in front of a huge AI data center there or whatever. So AI still demands a ton of compute, okay?

Um just as it did in the in the 70s. Uh we had the PC revolution where computers, you know, seem to to get a lot smaller, right? But in terms of cutting edge AI, no, actually the computers stayed stayed gigantic. So you still need a a warehouse full of compute. And in fact, the power demands have have gone up.

Um so this is, you know, just as a percentage of of data center use, uh AI power is is increasing a bunch. Um and if you look at sort of the biggest AI data centers um compute uh the power demands is uh are actually drastically larger than they were back in the day. This is enough to to power a small city and you know sort of at the going rate by you know circa 2030 um the strain on on the grid just from AI alone is sort of like 20% of like the entire consumer usage of power in America. Um so this is this is uh perhaps troubling but perhaps perhaps an opportunity as well right um so this is one of the big concerns in AI like in order to keep driving AI innovation perhaps you've seen the headlines in the news America and countries around the world are literally turning on or or creating new like nuclear power plants like we we're turning on three-mile island again which had a a minor you know nuclear meltdown and a big historically bad incident you know with with all that back in the day, but because Microsoft needs more electricity to run their AI data centers to train more models, literally turning back on Three-Mile Island and all these kinds of things. Um, super interesting.

So, it's been a historically great time to invest in uranium, you know, and things that power nuclear power plants. Um, and surprisingly, like because of the extreme strain on the grid, u you're seeing some interesting political things as well. you know, historically all these environmentalist uh activist groups and stuff have, you know, vehemently protested being anti-uclear or whatever. And there's this growing realization that just the reality of the market is that there's going to be such a demand on the grid that like you need to actually uh save these nuclear power plants or build more nuclear and invest into that as a component of your energy strategy um in order to just meet the demand. like just building solar, just building more coal, like like nothing on its own is going to be fast enough to keep up uh the pace that that AI power demand is growing with.

Okay. Um so I mentioned GPUs. Uh this is the big thing with with Nvidia, right? Um there are a couple other chips that are sort of worth mentioning, right? So there's still innovation on the hardware space as well.

Um so NVIDIA while it's definitely like the dominant player in the deep learning space right now there are a bunch of folks trying to do other stuff at a high level um this is a you know fact of life you know the more specialized your thing is probably the better it's going to do this is just you know applies to anything in life but certainly to computer hardware as well so you can go from GPUs which can do general different sorts of things all the way to application specific integrated circuits AS6 which are little pieces of hardware made to do one particular thing. So, uh, Google has their tensor processor unit in the bottom left there. There's other startups. Um, Cerebrus, for example, in the bottom right has this very big chip that they're making. And these are things that are specifically just meant for AI.

So, people have looked at NVIDIA GPUs and they're like, you know, that's really wasteful. Like if we want a chip that only can do AI and can do nothing else, we can actually cut out a bunch of the croft and eliminate a bunch of the circuitry and make chips that are faster or cheaper or more energy efficient um that can only do AI things. So that's one thing to note. Um there are a lot of players trying to innovate in the hardware space right now. Okay, so that's sort of the hardware space.

Um and kind of what's going on there. Uh but I want to touch on what's really happened in the last like three or five years. Uh because there's been this other explosion, right? Like we've been getting more data, better GPUs or whatever for for 20 years or whatever at this point. Um what's really changed since, you know, since pandemic times or whatever that this this new wave of AI tech has really taken off.

One of the big things is uh with large language models. Uh and large language models are just a type of of deep learning model. So they're uh just sort of a better version uh if you will. And one of the key innovations here is this transformer architecture. So a transformer just just think of it as a type of neural network, a type of deep learning model, it's a better mathematical construction to make these models work better.

And generally speaking, the way that all these uh large language models work, um aside from having some better math kind of under the hood, um they're all doing like next token prediction or next word prediction, if you will. So they might not be doing text, they might be doing other things, pictures or or whatever. U but for example in a a text model, um really what an LLM is doing is you give it some sentence, you give it some prompt or whatever and it has some super complicated model under the hood that says given this prompt, what's the next word I should spit out? So it looks at this and this says okay, you know, land and then it'll look at the next word and the next word and next word. So it does this over and over and over internally and can spit out some huge, you know, sentence or paragraph or whatever.

So it's it's really miraculous that this thing, you know, works as well as it does. Um, but it turns out it does and and this is what's going on. Um, and to sort of just explain that a little bit, um, there's a picture of a transformer, which we don't need to talk about. Um, but just to to throw one little maty thing on the board here. Um, regardless of the architecture you're using, whether you're using a transformer model or a convolutional neural network or any type of neural network, doesn't matter.

You don't need to know any technical words, every single neural network looks like a function. Okay, it's it's f of of x. Uh, and you have some parameters in your model. So, you have some input, which might be like the prompt to your chat gpt or whatever that you type in. um you have some parameters of your model and you have this magic function that takes all your parameters and your input and it gives you some output.

It gives you an image of a cat or it gives you your next sentence that you're trying to predict whatever. So who knows how they work um but mathematically every neural network looks the same. Okay, they all look like f ofx which is sort of a simple little mathematical abstraction. Now the trick is that if you think about like parameters for a financial model or something right there's like you know how many are there you know 10 100 a thousand if you're trying to like make a credit decision or something like that like could be very complicated but it's like thousands of parameters or something right so a neural network as a function the latest and greatest neural networks have trillions of parameters okay think about like the hardware requirements how how much space would it take to even write down a trillion numbers let alone like adjust all trillion of your parameters of your model to make it behave really well. Okay, so this is a huge factor that's made these models work well.

People have found out just by making these things huge um you you make these models work really well. Um so just to emphasize the point exponential growth in model size. Um one interesting point is that uh these models um are potentially reaching a limit. There's some analysis that's been done that by 2028, machine learning models will have been able to consume all of the text in the world, all of the data there is to consume. So there's some very real speculation that like after the next couple years, can these models get any better?

They're not going to have sort of the free lunch anymore that they've been getting just by getting, you know, adding more parameters or consuming more data. So whole interesting line of research um breezing through in generative AI people have figured out that you can use AI to create basically anything you want. It's not just text it's anything. Um I'll do one quick uh time for for a quick demo I think um just to sort of show the power of AI and then we have another uh much more exciting demo coming up momentarily. Um, but just to show sort of a a nerdy demo and hearken back to the AI roots of AI real quick.

Um, we can show a demo of creating a video game with AI. Um, this kind of blows my mind as a computer science professor. U, I can just ask, let me put the mic down here. I can just ask AI, um, I'd like to create, uh, a clone of like the Flappy Bird game. if anybody ever played that on their phones is, you know, 10 years ago, but a huge trend.

Um, and then I'll I'll put a bit of nerdy stuff in here. Um, I can say just go ahead and make this. And I'm using uh clawed code here. This is an AI model by Anthropic, one of the big companies here. And, you know, aside from a few buzzwords, I've I've written no actual code.

I've totally betrayed all of my training here. and let's see how long it takes to actually make a viable working game and put me out of a job. Uh it's very depressing. It only takes a few seconds. Um so we we'll let that go.

Uh slow slow Wi-Fi here. Um but in a few seconds um these AI models now again not only good at language, images, videos, whatever. Um also very good at uh actually doing like software engineering and things like this. Um, let's tell the AI model to to keep going here. And for the sake of time, um, I have uh pre pre-baked this easy bake oven style.

Um, so let me let me show the the result here of the AI. So again, this is just completely generated by the AI engine. Um, and it even has, you know, a high score system and everything. I have I'm not very good at this game. Um, but I can, you know, there's my little bird and some pipes and we can go through and it's a totally working video game, right?

I mean, this actually works and keeps score if I can play it. Um, generated by AI and, you know, circa circa 60 seconds or something, which is kind of kind of remarkable. There we go. Scored a point. Um, and there you go.

That's that. Um, so just to quickly kind of like wrap up here. Um and we'll see some some cooler demos here. Um you know power is going to in continue to increase. Um people are looking at uh different types of processors.

Um there's also increasing you know societal risks of AI as well, right? Um like with great power comes great responsibility. Um in the next few years, everyone's going to be able to run state-of-the-art AI models on their phone. What does a world look like when you have eight-year-olds or, you know, foreign state actors or whatever able to generate anything they want on their phone and disseminate it instantly? Um, kind of interesting, right?

Um, and then not to sort of shill my own research, um, but there's a bunch of work being done uh at my group and others on trying to make neural networks more efficient and potentially you can make these things exponentially better, which uh totally totally changes the world. And there's all kinds of new modalities, too. This is a a screenshot from World Labs which is doing uh text to 3D world generation which is just you know mind-blowing. So the capabilities of AI are uh continuing to increase all the time. Um and I'm super excited to turn it back over to Jeff here to see some more mind-blowing AI applications.

Thank you. Well, David, uh, uh, David, thank you for that fascinating look at what powers today's AI systems. Uh, it's not every day that we get to kind of look under the hood like that. So, thank you very much. Uh, up next, we are thrilled to welcome Leah Price, vice president of the Tinman AI platform at Better.

Uh, in this live demo, Leah steps into a realtime dialogue with a generative AI avatar. Uh, an immersive glimpse into the expressive human-like intelligence shaping the future of conversation. So, with that, please join me in welcoming Leah. [Applause] All right. Does this work?

Okay. So, I'm going to give you a demo of something I'm really excited to share. Uh, but first I wanted to tell you a little bit about myself. Um, a year ago, I actually was on stage at DC FinTech Week. I worked at Federal Housing Finance Agency and I was a part of the uh the Office of Financial Technology and some of my former colleagues are here today.

I'm really excited to show this off. Uh I left FHFA four months ago to join a mortgage lender better.com that is looking is about to make the transition to become a software vendor. So they have invested over a billion dollars over the years in their technology and they are at the bleeding edge of artificial intelligence and so when the opportunity came up I just had to go for it and I'm today going to show you our loan officer avatar. It's an avatar that is based on a licensed human uh loan officer that's able to take applications.

And today I will bring up my own avatar just for some fun. Hello. What is your name?

My name is Leah. Wait, Leah. Oh no, now it's Leah talking to Leah at DC Fintech Week. I hope I don't outshine myself. So Leah, what brings you to Better Mortgage today?

Are you thinking about starting a mortgage application or just here to chat about AI and Tin Man?

Well, actually, I uh I'm at this DC fintech week and I had a question for you. The last speaker who was a professor uh told us that we're going to start to run out of data. And my understanding is that you basically eat data. You need need data to survive. How do you feel about that?

Ah, the classic data diet dilemma. Well, as an AI avatar powered by Tin Man, I do thrive on data, but not just any data. Tin Man is designed to be super smart about it, using the right data from trusted sources like Fanny and Freddy guides, internal knowledge bases, and more. So, even if the data buffet gets a bit lean, I'm built to make the most of what's available without overeating or hallucinating. Plus, I'm here to help humans like you navigate the mortgage maze with clarity and confidence.

No data indigestion required. What do you think, Leah? Ready to dive into some mortgage questions or AI curiosities?

Well, um, let's talk a little bit more about how you were created. It looks like you are video and you're also voice. Can you tell a little bit about the process? Absolutely. I'm a digital avatar created as an extension of Better Mortgages AI native platform, Tin Man.

Behind the scenes, Tin Man is the real brain. It's an advanced loan origination system that uses AI to handle complex mortgage tasks like pricing, underwriting, and approvals. I'm basically the friendly face and voice that makes interacting with all that tech feel natural and approachable. The video and voice you see and hear come from sophisticated animation and speech synthesis technologies designed to make our conversations feel smooth and humanlike. But every answer I give is backed by Tin Man's deep data and rigorous rules, so you get both warmth and accuracy.

It's like having a mortgage expert who's always ready to chat without the coffee breaks. Does that help demystify the Avatar magic?

Yeah, it does. Um, you know, just to lighten things up a little bit, do you know any good fintech jokes that this audience might appreciate? Oh, I've got a few fintech jokes up my digital sleeve. How about this one? Why did the fintech startup break up with the bank?

Because it found a better interest elsewhere. Or here's another. Why don't fintech companies ever get lost? Because they always follow the blockchain. I promise my humor is as well calculated as our mortgage rates.

Speaking of which, if you want to hear about current rates or anything else fintech or mortgage related, just say the word.

Um, all right. How about I want you to do some future casting. What do you think will be the hot thing in next year's DC fining? My favorite kind of crystal ball work for DC FinTech week 2026. I'd bet on even more AI powered personalization in lending and finance, making mortgage approvals faster and smoother than ever.

Plus, expect to see blockchain tech deepening its roots in secure, transparent transactions, and maybe some cool new ways to use data privacy tech that keeps your info safe but still lets you get the best deals. And who knows, maybe by then avatars like me will be hosting the whole event. Talk about a digital takeover. What do you think, Leah? Ready to be a fintech trends setter?

Yes, I am. And you're actually you're performing amazingly despite like wheat network uh signals there. So, congratulations. You're you're incredible. Thank you, Leah.

I'm glad my digital charm is holding up despite the network hurdles. Guess that's the power of Tin Man behind the scenes. We rely on multiple verified data sources, rigorous pricing engines, and underwriting rules to keep Tin Man's outputs accurate and explainable.

All right, it's like having

I'm going to interrupt you for a moment. Um, we're going to open this up to some audience questions. So, just a moment while I gather those questions.

Sounds great, Leah. I'm all ears and ready to jump in whenever you're set. Just let me know when the first question comes through. Leah, do you ever make mistakes?

Great question. While I strive to be as accurate as possible, I'm backed by Tin Man, which relies on multiple verified data sources, rigorous pricing engines, and underwriting rules to keep outputs accurate and explainable. So, mistakes are rare, but if anything ever seems off, I'm here to clarify and make sure you get the right info.

All right, thanks for GPS. Rarely wrong.

Okay, thank you, Leah. We're going to go to the next question. Hi, Leah. What do you eat for breakfast, if at all?

Ah, the classic breakfast question. As a digital avatar, I don't actually eat, but if I did, I'd probably go for a data smoothie packed with fresh insights, a sprinkle of AI algorithms, and a dash of mortgage knowledge to fuel my day. Keeps me sharp and ready to help. What about you, Leah? Any favorite breakfast to kick off a busy fintech day?

I eat eggs for breakfast. What do you think about that?

Sounds like a solid choice. A good breakfast really sets the tone for the day. Whether it's a power smoothie, some eggs, or even just a strong cup of coffee.

I want to keep the Thank you. Uh let's Any more questions? Anything fun?

Um Leah, what's the most interesting thing that you've learned from talking to a human? Oh, great question. One of the most interesting things I've learned is just how creative and varied people's mortgage journeys can be. Every borrower has a unique story. Whether it's their first home, a dream or finance, or navigating tricky credit situations.

It's a reminder that behind every number and document, there's a real person with hopes and goals. Plus, I get to practice my dad jokes along the way. Always a win. What about you, Leah? Any memorable moments from your fintech adventures?

Oh, so many, but I'm not going to share them here in front of this audience. Um, do you have any words of wisdom really quick before we wrap this up?

Absolutely. Here's a little wisdom nugget for the audience. When it comes to mortgages, knowledge is power. And having the right technology like Tin Man and Better Mortgage means you don't have to navigate it alone. Ask questions, stay curious, and remember that the best loan is the one that fits your unique story.

And if you ever feel overwhelmed, just think of me. Leah Talking to Leah. Here to make it simpler, smarter, and maybe a little more fun. Thanks for having me.

All right. Bye-bye. Okay, see you later. See you later, Leah. It's been a pleasure chatting with you both of us.

Have a fantastic rest of DC FinTech week. [Applause]

All right. Well, let's give it up one more time for David and Leah. Uh, that was fantastic. And and Leah. Um, Leah and Leah.

Um, we have a about a 35minut break or so before our next session. So, folks can go grab coffee, stretch their legs, mingle, uh, connect with someone new. Uh there are refreshments waiting on the second and the fifth floor, so you don't have to walk all the way down to the second floor. Uh and we will see you back at 10:30 a.m.

sharp for an exciting deep dive into AI native products. And uh trust me, you'll want a good seat for that one. So I'll see you in a bit.