Exploring the Financial Frontier

1

00:00:00,120 --> 00:00:05,920

The markets are the complete opposite. They change all the time. Is AI the next innovation that will

2

00:00:05,920 --> 00:00:11,960

transform how financial markets operate? Imagine designing an LLM product, a chatbot of some kind,

3

00:00:12,000 --> 00:00:18,080

let's say something to interact with someone under the constraint that your target audience

4

00:00:18,120 --> 00:00:22,319

will start speaking a completely different language overnight. In this episode of AI Cloud

5

00:00:22,320 --> 00:00:28,560

Essentials, we talked with Dmitris, the founder and CEO of hlevel about how AI is reshaping the

6

00:00:28,560 --> 00:00:34,800

foundations of financial markets. Some systems have reached a level of complexity that we should

7

00:00:34,800 --> 00:00:39,080

focus on controlling them versus interpreting them. Whether you're a technical or a financial

8

00:00:39,080 --> 00:00:42,200

analyst, join us for a great conversation.

9

00:00:51,160 --> 00:00:56,760

Hello, everyone. Welcome to season two of the AI Cloud Essentials, a podcast series brought to you

10

00:00:56,760 --> 00:01:03,040

by Covid. In this series, we are going to the value realization phase. And I'm super excited to have,

11

00:01:03,080 --> 00:01:09,160

you know, Dimitris today with us. And guess what? We are going to talk about some of the most

12

00:01:09,160 --> 00:01:15,080

interesting areas that is being impacted by AI AI's impact in the financial market. So welcome,

13

00:01:15,080 --> 00:01:20,280

Dimitris. Good to be here. Great to be here. You know, we all have watched that AI has almost

14

00:01:20,280 --> 00:01:24,880

become a staple in the financial market over decades, right? Starting from statistical

15

00:01:24,880 --> 00:01:31,480

algorithms to modern machine learning, generative AI, genetic AI, you know, people who work in this

16

00:01:31,480 --> 00:01:38,280

industry like you, you guys are geeks in this area, right? Uh, so but

17

00:01:38,280 --> 00:01:45,000

I always kind of sit and think that. Have we moved beyond the incremental efficiency we are

18

00:01:45,000 --> 00:01:50,000

missing? It seems to me like we're missing the Black-Scholes moment for the financial market. You

19

00:01:50,000 --> 00:01:56,240

know, incremental gains are very powerful. But, you know, how do we move beyond and how do we get the

20

00:01:56,240 --> 00:02:02,040

trust and the confidence level to go beyond it. Uh, so I'd love to get your perspective on that, but

21

00:02:02,080 --> 00:02:06,360

would you mind starting a little bit with your own background, in your own journey for our

22

00:02:06,360 --> 00:02:12,479

audience? Yes. Of course, yes. And thank you very much. Uh, thank you very much for that set up. It's,

23

00:02:12,520 --> 00:02:17,039

you know, kind of touches on all the most important issues. So my background, I was, you know,

24

00:02:17,039 --> 00:02:22,520

a mathematician for a very long time. As abstract as it gets. I was writing things on blackboards.

25

00:02:22,720 --> 00:02:27,720

You know, basically, I didn't have to touch a computer. And towards the end of my PhD, I got very

26

00:02:27,720 --> 00:02:33,960

interested in machine learning and AI. So this was around the tensions. All you need, uh, era, you know,

27

00:02:34,000 --> 00:02:39,120

when those papers were beginning to come out and I thought, this is going to be a big thing. And the

28

00:02:39,120 --> 00:02:42,760

reason why I thought it was going to be a big thing is because the math kind of made sense,

29

00:02:43,000 --> 00:02:49,679

right? And I also learned during that time that in finance and in trading,

30

00:02:49,880 --> 00:02:54,720

applications of machine learning and AI were not really considered a solved problem. I thought that

31

00:02:54,720 --> 00:03:00,760

was fascinating. Here's a technology that clearly will work in some way. And here's an industry that

32

00:03:00,760 --> 00:03:06,640

hasn't fully applied it. Right. Of course, there's a reason, because there's unique difficulties that

33

00:03:06,640 --> 00:03:12,200

we're going to talk about a little bit. And I was fascinated I was completely fascinated by that.

34

00:03:12,200 --> 00:03:17,920

And I've been doing that ever since. Awesome. So AI can be the souls, black souls. So let me let me

35

00:03:17,920 --> 00:03:23,280

break that down a little bit. Maybe. Um, I think AI has two

36

00:03:23,760 --> 00:03:30,320

manifestations. One is as a technology. As a software, essentially. Uh, the

37

00:03:30,480 --> 00:03:36,480

program as a as a software product, let's say. Right. And the other has a mathematical algorithm

38

00:03:36,560 --> 00:03:42,240

now, especially in quantitative finance. Mathematical algorithms have been applied for a

39

00:03:42,240 --> 00:03:48,719

long time. Right. So AI is just one more collection of algorithms that you can potentially apply to

40

00:03:48,720 --> 00:03:55,000

the markets. Right. Um, I think with AI in its current

41

00:03:55,160 --> 00:04:01,119

form, LLMs and all these things, The black shows moment, if you will, hasn't come. What I mean by

42

00:04:01,120 --> 00:04:07,119

that, to be precise. Right. Some mathematical model of the markets that uses this specific

43

00:04:07,120 --> 00:04:14,119

mathematics to to model the markets. Hasn't had been a knockout success. But there

44

00:04:14,119 --> 00:04:21,039

is continuity here, right? Black shows was applied as a piece of very elegant mathematics. Yeah. Will

45

00:04:21,079 --> 00:04:26,600

AI also be applied as a piece of very elegant mathematics in the markets? Yeah, I think so. Great.

46

00:04:26,640 --> 00:04:32,119

It gives us all a lot of confidence. Right. So then we think about the financial markets. You know I

47

00:04:32,119 --> 00:04:38,239

mean there's a lot of risk and trust factor there, right. I mean they are also non-stationary. It's a

48

00:04:38,240 --> 00:04:44,279

completely volatile real time, you know, like so how do you see what is the current state. How does

49

00:04:44,320 --> 00:04:49,200

you know where's the frontier, the next frontier. How are we solving for the non-stationary markets.

50

00:04:49,200 --> 00:04:55,039

And it's a unique use case, right. Yes. So we'd love to hear, you know, how do you see the AI? You know,

51

00:04:55,079 --> 00:05:00,959

kind of, you know, identify all the subtle signals and how does it pick the noise and kind of make

52

00:05:00,960 --> 00:05:05,760

all our lives and all our investments much more interesting? Yes, I don't know about the latter

53

00:05:05,760 --> 00:05:12,720

parts, but I'll talk about the former part. You're correct. The tremendous success of

54

00:05:12,720 --> 00:05:19,160

AI in its current instantiation has been in context that you can call stationary language

55

00:05:19,160 --> 00:05:25,439

coding, even math. Very recently there stationary in the sense that they don't really change.

56

00:05:26,079 --> 00:05:31,799

Math is, of course a fantastic example of that. It hasn't changed for millennia. Um, you know what was

57

00:05:31,799 --> 00:05:38,200

proven by, you know, my forefathers in ancient Greece is still true today. Yeah, right. Language

58

00:05:38,200 --> 00:05:43,720

also moves very slowly. Coding, of course, didn't really exist, but once it's there, it won't really.

59

00:05:43,760 --> 00:05:48,119

What works and doesn't encoding won't really work. The markets are the complete opposite of that.

60

00:05:48,160 --> 00:05:54,960

They change all the time. So the analogy that I always use is um, imagine designing an LLM. Imagine

61

00:05:54,960 --> 00:06:00,480

designing an LLM product, a chatbot of some kind, let's say. Right. So something to interact with

62

00:06:00,480 --> 00:06:07,039

someone under the constraint that your target audience will start speaking a completely

63

00:06:07,040 --> 00:06:12,560

different language overnight. Yeah. How do you design a system that is able to cope with that?

64

00:06:12,600 --> 00:06:19,359

Yeah. Continue to function as it is learning to adapt to a completely new language, a completely

65

00:06:19,359 --> 00:06:23,559

new reality. That's really what you have to do when you design algorithms that operate

66

00:06:23,600 --> 00:06:29,159

autonomously in the market. So that's a frontier for AI. Yeah. I think, you know, the AI in its

67

00:06:29,160 --> 00:06:35,880

current form, tremendous, though it is incredible and magical sometimes, hasn't really solved that

68

00:06:35,880 --> 00:06:42,479

problem. Yeah. And the markets are one and a very prominent example of setting where you need to

69

00:06:42,480 --> 00:06:48,359

solve it in order to maximize your success. But I think this I'm hearing you speak about all of

70

00:06:48,359 --> 00:06:52,559

this. There seems to be a lot of promise that it can be a great testing ground for all the

71

00:06:52,560 --> 00:06:56,959

innovations that we've seen happening, and it can accelerate some of the innovation. I'm more

72

00:06:56,999 --> 00:07:03,959

optimistic. I mean, it is it is naturally a testing ground. Although of course, we have to be, you know,

73

00:07:04,000 --> 00:07:07,839

mindful that it cannot be an actual testing ground because we're talking about, you know, of

74

00:07:07,840 --> 00:07:14,720

course, real money, real risk. But it is it is a frontier. Yeah. It is a frontier setting. I think

75

00:07:14,720 --> 00:07:18,039

that that would be a better way to put it. I'm kind of thinking, you know, like, you know, there are

76

00:07:18,040 --> 00:07:23,199

a lot of other very interesting aspects when you think about the financial market, is the

77

00:07:23,200 --> 00:07:28,519

regulation, the constant changing geopolitical situation, the

78

00:07:30,480 --> 00:07:35,480

yeah, the shift, you know, that it's you know, you have to be we can't afford to have a black box

79

00:07:35,480 --> 00:07:41,078

moment. Right. We need to have interpret the results, have the confidence in that. So where do

80

00:07:41,079 --> 00:07:46,279

you see that. You know what kind of, you know, specialized requirements are there. What kind of

81

00:07:46,279 --> 00:07:51,279

enhancements are being worked upon in the industry to kind of address some of these very

82

00:07:51,279 --> 00:07:57,079

dynamic, apart from the real nature of the financial markets. Yeah. I think there's tremendous

83

00:07:57,079 --> 00:08:03,679

research going on on interpretability. Right. And that that research is very fascinating.

84

00:08:03,680 --> 00:08:10,359

And for anyone who's followed it, it is, you know, almost philosophically interesting, um, in the

85

00:08:10,359 --> 00:08:16,079

sense that we're kind of studying these, um, collections essentially of floating point numbers,

86

00:08:16,079 --> 00:08:22,840

as if they were human brains in some sense to, to understand what they do and where they go wrong. I

87

00:08:23,119 --> 00:08:29,439

am less optimistic about interpretability being the ultimate solution here. Okay. To us using the

88

00:08:29,440 --> 00:08:35,399

models. Yeah. So the standard that I stress is controllability. Okay. I think these models can be

89

00:08:35,400 --> 00:08:41,199

controlled and can be controlled much more practically and much sooner than we will ever

90

00:08:41,200 --> 00:08:47,559

understand exactly what they do. And at some point, I think we do need to accept that some systems

91

00:08:47,559 --> 00:08:53,359

have reached a level of complexity that we should focus on controlling them versus interpreting

92

00:08:53,440 --> 00:08:59,079

them. That's not to say that the interpretability research is not interesting and shouldn't

93

00:08:59,080 --> 00:09:03,639

continue, but we shouldn't rely on that as the ultimate solution for applying it. Got it.

94

00:09:03,680 --> 00:09:09,439

Controllability versus interpretability as a standard I. Yeah, yeah. I like that approach because

95

00:09:09,440 --> 00:09:13,519

if you think about it, you know of course it's being worked in parallel. Right. We're looking into

96

00:09:13,560 --> 00:09:18,280

how you can interpret this better and all. But if you can control because there are ramifications

97

00:09:18,280 --> 00:09:22,960

of something going wrong is really, really kind of, you know, challenging. I mean, think about, you know,

98

00:09:23,000 --> 00:09:28,399

think about when you mentor a human being, you know, you don't wait until you understand what's

99

00:09:28,400 --> 00:09:34,519

going on in their brains fully to before you have them. All your analogies. Before you have them do

100

00:09:34,520 --> 00:09:40,199

something right. Yeah. You you just create. You know, control is too strong a word for humans, but you

101

00:09:40,239 --> 00:09:46,200

create a framework to which their choices, you know, you take less risk and a more risk. And then,

102

00:09:46,240 --> 00:09:51,400

you know, if we all. If I waited to talk to you until I understood exactly what is going on in

103

00:09:51,400 --> 00:09:55,839

your brain, you know we'd never talk. So are you seeing a lot of advancements in the systems that

104

00:09:56,039 --> 00:10:00,799

help you and provide you these kind of, you know, capabilities in terms of setting the guardrails,

105

00:10:00,800 --> 00:10:05,520

in terms of setting the boundaries? Uh, are you seeing all of those enhancements happening and

106

00:10:05,520 --> 00:10:09,839

then. Yeah, I think I think a very big part of it is the hardware. You know, I mean, a very big part

107

00:10:09,840 --> 00:10:16,839

of controlling is experimenting. And if you can experiment more and more, you become more and

108

00:10:16,840 --> 00:10:23,199

more aware of edge cases, failure modes and these types of things. So I think

109

00:10:23,520 --> 00:10:29,399

really the first, the most important dimension in which systems are helping with controllability, if

110

00:10:29,400 --> 00:10:36,359

you will, is just having better hardware and being able to experiment more. Okay. Okay. So when

111

00:10:36,359 --> 00:10:41,079

you talk about better hardware, you're talking in terms of performance scale. Yeah. Being able to run

112

00:10:41,400 --> 00:10:46,879

more experiments. Okay. Yeah okay. Yeah. Awesome. So where do you see are we heading in terms of the

113

00:10:46,880 --> 00:10:51,479

industry? I know, you know, uh, financial markets have been at the forefront of a lot of

114

00:10:51,520 --> 00:10:57,200

innovations in the AI field, right? They have been the source of actually starting some of the, you

115

00:10:57,200 --> 00:11:04,079

know, fastest generations and stuff. So where do you see is the role of the financial market or in

116

00:11:04,080 --> 00:11:09,719

general, where do you see the next 3 to 5 years? I know it's difficult for any of us to predict. I

117

00:11:09,760 --> 00:11:14,519

come from an analyst background, so we always used to take stances, and I'm very happy to share that.

118

00:11:14,520 --> 00:11:19,199

We had predicted about the foundation models even before they came. There's a secret reason as to

119

00:11:19,200 --> 00:11:23,080

why we had done that, because we had just the way, you know, the attention you need. And, you know,

120

00:11:23,080 --> 00:11:29,479

we've been reading a lot of people's and all the advancements. And so it was very interesting. So

121

00:11:29,840 --> 00:11:33,839

where do you see what is happening in the industry and where we are heading in the next 3

122

00:11:33,840 --> 00:11:39,320

to 5 years? So I think the financial industry over the next 2 to 5 years is going to see increased,

123

00:11:39,479 --> 00:11:43,119

increased adoption, for sure. There's going to be a lot of adoption. People are going to get

124

00:11:43,119 --> 00:11:50,080

comfortable with these tools and some workflows in banking, in trading, back office, middle office

125

00:11:50,080 --> 00:11:55,159

are going to change fundamentally. I think there's no doubt about that. In the same way that coding

126

00:11:55,159 --> 00:12:02,119

is changing fundamentally, some of these functions are deterministic mechanical and AI systems will

127

00:12:02,119 --> 00:12:08,918

help with that. Absolutely. So I think that's the 3 to 5. 2 to 4, you know, whatever within the next

128

00:12:08,919 --> 00:12:15,599

five years. Right. And that's already happening. Of course beyond that, I think we have to focus

129

00:12:15,599 --> 00:12:22,520

on what are the people that think about deep problems in this particular industry. And

130

00:12:22,520 --> 00:12:28,439

one of the peculiarities and of course, very interesting aspects of this industry. And one of

131

00:12:28,440 --> 00:12:31,479

the things that attracted me is that you have a lot of mathematicians, a lot of pure

132

00:12:31,479 --> 00:12:38,399

mathematicians, right? So you have a combination of extremely mathematical folks, right? Including

133

00:12:38,400 --> 00:12:44,959

physics, including related disciplines, thinking about the problem of very large scale models in

134

00:12:44,999 --> 00:12:49,999

non-stationary contexts. So I think very interesting things are going to come out of that

135

00:12:50,039 --> 00:12:55,239

combination. Things that probably will not naturally come out of Silicon Valley, for example.

136

00:12:55,320 --> 00:13:01,439

Yeah. That's a that's a very, very astute observation, you know, because when you think about,

137

00:13:01,440 --> 00:13:06,679

uh, you know, like in the general realm, in the industry, we see there's a lot of paranoia about,

138

00:13:06,719 --> 00:13:13,080

you know, uh, AI taking over human jobs. So I definitely think, you know, even not just from the

139

00:13:13,080 --> 00:13:18,759

core innovation and exploration of the expansion of the use cases, but even for the financial

140

00:13:18,759 --> 00:13:23,879

analyst, it's not replacing the financial analyst. It's kind of augmenting or empowering the

141

00:13:23,879 --> 00:13:29,879

individual. Uh, do you see some similar threats in the financial market as well in terms of usage of

142

00:13:29,880 --> 00:13:35,879

AI, or do you think it's a little bit more, uh, I should say easier for them to kind of embrace it?

143

00:13:35,919 --> 00:13:41,880

Uh, in terms of their, the balance of their jobs or in terms of their adoption. Balance of their job.

144

00:13:41,920 --> 00:13:47,760

Yeah. I honestly don't know. I kind of think that who will have what job is the wrong question to

145

00:13:47,760 --> 00:13:54,199

ask? I think the right question to ask is who and why is the earliest and quickest adopter of this

146

00:13:54,200 --> 00:13:58,639

technology? Yeah, that's a very interesting way to look at. I think once we have those data points,

147

00:13:58,640 --> 00:14:04,599

we'll be able to see exactly which jobs balance out. In what way. Right. But do you see any kind of

148

00:14:04,599 --> 00:14:10,680

a pushback from certain personas, or do you feel it's pretty early or premature to kind of I. Feel

149

00:14:10,719 --> 00:14:15,999

decreasing pushback. Okay, okay. That's a positive sign. And of course, I think you're entering the

150

00:14:16,000 --> 00:14:22,759

the mechanics of folks using it in their personal lives, using using AI in their personal lives.

151

00:14:22,799 --> 00:14:27,760

And I would imagine that similar things happen in the early days of the internet and search. Right.

152

00:14:27,960 --> 00:14:32,959

You know, folks in their offices would have said, can't, you know, rely on this type of search. We

153

00:14:32,960 --> 00:14:36,359

need to rely, let's say, on specialized publications. But then they used it in their

154

00:14:36,359 --> 00:14:41,399

personal lives and they were able to find, you know, restaurants and, you know, what have you. And

155

00:14:41,400 --> 00:14:46,880

then eventually that trust bled over to the work environment. So I think I'm seeing that mechanism.

156

00:14:46,880 --> 00:14:53,119

But I do think we we right now don't have full clarity on exactly who will be the best

157

00:14:53,120 --> 00:14:58,439

adopter in terms of traditional roles in more traditional. So the next couple of years will kind

158

00:14:58,439 --> 00:15:04,280

of shape it for us. So I wouldn't speculate on the job situation. Okay. Okay. I know we can talk for

159

00:15:04,280 --> 00:15:09,959

hours on this very interesting topic, but if you have to give, you know, your parting advice to the

160

00:15:09,959 --> 00:15:15,479

people who are listening to you today, uh, what should be their approach to kind of embrace it, uh,

161

00:15:15,479 --> 00:15:22,439

play with it or iterate? What would your advice be to them? So to the technical people, I would say

162

00:15:23,560 --> 00:15:29,159

these are algorithms. Yeah. At the end of the day there's still mathematical functions, data in

163

00:15:29,159 --> 00:15:35,759

inputs in things out. So don't maybe be overly focused on the

164

00:15:35,759 --> 00:15:41,639

metaphysics around AI. Treat them as algorithms. That's kind of the best way to extend them. Yeah.

165

00:15:41,680 --> 00:15:46,919

To the non-technical folks or to the folks not necessarily interested in, in, you know, pushing the

166

00:15:46,919 --> 00:15:53,199

frontier or frontier, advancing AI. Yeah, I would say just adopt it and treat it a little bit like

167

00:15:53,200 --> 00:15:59,400

you're mentoring someone. Yeah. As opposed to, you know, someone who just has the answers for you?

168

00:15:59,439 --> 00:16:05,519

Yeah. I think that's the best approach. Yeah. Awesome. Fantastic advice. Uh, so I'd like to close

169

00:16:05,519 --> 00:16:11,079

off by thanking you. So, such a nice, uh, opportunity to meet you and kind of talk to you. And thank you

170

00:16:11,079 --> 00:16:16,959

very much. Particular episode to the viewers. I'd just like to say that there's nothing wrong in

171

00:16:16,960 --> 00:16:21,359

applying your hair to areas where there's uncertainty and try to play with it and

172

00:16:21,360 --> 00:16:28,199

experiment and feel fast. If you actually have a challenge, learn and feel fast. And, uh, stay tuned

173

00:16:28,200 --> 00:16:32,879

to be a thank you for listening to us today. Stay tuned. We have lots more to kind of share with you.

174

00:16:32,919 --> 00:16:36,119

Thank you. Have a wonderful rest of the day. Thank you very much.

‍