Advanced Fraud Prevention in the Age of Artificial Intelligence
In this episode, Andy and Frank sit down with Pavel Goldman-Kalaydin, head of Artificial Intelligence and Machine Learning at Sumsub, a global company specializing in KYC, AML, and anti-fraud technologies.
They explore the challenges in verifying identities remotely, the rise of deep fakes for fraud, and the use of AI and machine learning to combat these threats. From discussing the impact of technology on security measures to Pavel’s journey in the field of computer science and AI, this episode offers insights into the evolving landscape of fraud detection and the intersection of technology, AI, and security.
Join us as we delve into the complexities of anti-fraud measures and the fascinating world of AI and machine learning.
Show Notes
00:00 Securing customer journey from onboarding to verification.
04:44 2 years ago, typical attack to open account.
06:58 German video identification process prolongs account opening.
12:16 Analyze data patterns to make informed decisions.
13:34 Questioning deep fake implications for customer data.
17:42 Advancing technology makes image manipulation easier.
22:32 Financial fraud: creating defects for unexpected reasons.
25:53 Fascinating progress in beta software development.
29:23 Samsung creates its own products, understands customers’ needs.
29:58 Problem with defects, educate and ensure understanding.
34:01 Interest in drug development and AI technology.
38:57 Audible sponsors Data Driven with free audiobook.
41:05 Please rate and review our podcast.
Transcript
In this 349th episode of data driven, we are pleased
Speaker:to interview Pavel Goldman Khaledin, where he's the head of artificial
Speaker:intelligence and machine learning at Sumsub.
Speaker:Sumsub isn't your average AI startup. They're
Speaker:globally recognized for their work in k y c, AML,
Speaker:and anti fraud technologies. Our guest is the
Speaker:wizard behind the curtain, crafting tech to outsmart financial
Speaker:fraud does and deep fake artists. Quite the
Speaker:digital Sherlock Holmes, if you will. Now here are
Speaker:Frank, Andy, and Pavel.
Speaker:Hello, and welcome to Data Driven, the podcast where we explore the emergent
Speaker:Fields of data science, artificial intelligence, and,
Speaker:of course, data engineering, which is basically the underpinning of it
Speaker:all. And with me on this, journey is my favorite data
Speaker:engineer of them all, Andy Leonard. How's it going, Andy? Good, Frank.
Speaker:How are you? I'm doing alright. We we were recording this, the day
Speaker:after we did a 2 hour show,
Speaker:Kinda by accident, don't I see our guest in, it look kinda had this
Speaker:look of, uh-oh. No. It's not gonna turn. I can't do that today.
Speaker:But we are very excited here to in spite of our issues with Microsoft
Speaker:Bookings, in spite of our crazy hectic schedules, And in
Speaker:spite of your allergies and, really tasty jelly jam and
Speaker:and and biscuits Really sorry about that. No.
Speaker:I I don't know what it is on the East Coast this week, man. It's
Speaker:it's well below freezing, and I'm sneezing. Oh, that rhymed.
Speaker:Allergy station should be over for me. I don't know what's going on. For real.
Speaker:But our guest is actually, from Berlin,
Speaker:and one of my favorite cities in the world. In fact, they were singing the
Speaker:virtual green room. Had I lived in Berlin instead of Frankfurt, I probably
Speaker:never would have come back to New York,
Speaker:or the US, but he is
Speaker:our guest today is Pavel Goldman Kaledin. Hopefully, I said that
Speaker:right. He is the head of AI and ML
Speaker:at Sumsub, a global know your customer anti
Speaker:money laundering, anti fraud company, and,
Speaker:we're we're welcome to we're happy to have him. Although, I don't think he's in
Speaker:Berlin today. I think he's somewhere a bit warmer. Welcome to the show, Pavel.
Speaker:Yeah. Hi, guys. Happy to be here. Good. Good. So I have
Speaker:a lot of questions. You know,
Speaker:first off,
Speaker:I think I can kinda see the map, but What's the
Speaker:connection between know your customer, KYC,
Speaker:anti money laundering, and anti fraud? I think I think
Speaker:I see it, but I wanna hear you you kinda walk me through it because
Speaker:I haven't had enough coffee either today. So so what's the, like, what's
Speaker:the common thread? Because, like, because I I've not seen those 3
Speaker:kinda put together in kinda 1,
Speaker:sentence, but I can kinda see why. But
Speaker:I I I I can try to explain. But the thing is and we actually
Speaker:this is what we focus on. So we try to secure as a company.
Speaker:We try to secure the whole customer journey from
Speaker:onboarding. So this is the first step of when, for instance, like, I'm in a
Speaker:bank. So So I want to onboard some of my customers, and I want to
Speaker:make sure that this has real persons, for instance, that are not fraudsters.
Speaker:So I want to onboard them, make sure they are,
Speaker:that person, they actually pretend to be. And then and
Speaker:here's the thing. If I can, for instance, like, I'm a Journey
Speaker:person. But a month later. There could
Speaker:be some, you know, strange patterns of, you know,
Speaker:financial transaction happening. So probably, there are some sort of a pattern of
Speaker:money laundering. So this is where transaction monitoring comes.
Speaker:So you can actually this is a person. So this is but knowing customers are
Speaker:very simple. You can actually I mean, you
Speaker:can So basic basic attack is to be just pretend to be,
Speaker:a person. You you are not, basically. But then even if I'm
Speaker:not, I'm just a real person, I can actually, yeah, come up with some sort
Speaker:of, you know, few things to
Speaker:do. And then where just we try to monitor it, and then from a permit,
Speaker:make sure that, Okay. We can actually flag the transaction and
Speaker:then make sure it's it's it's getting looped. And then, I mean, there is a
Speaker:flag raised, and then, Probably, we can do
Speaker:something about that. This is just, like
Speaker:this. If we're talking about anti fraud, and here's the
Speaker:thing. Sometimes it's very easy to see that something fish is
Speaker:happening. So for instance, like, A very like, 2 years ago, it
Speaker:was a very typical attack. So I tried to, you know, open a bank
Speaker:account or, like, remotely, And I actually, I'll leave somewhere
Speaker:else, or I don't I I use a stolen document. What
Speaker:what I can do To do that, I can actually just print out the
Speaker:image of a person and just try to make sure that actually the
Speaker:KFC provider like us Tried to make us
Speaker:believe that I'm a real person. That was a very, you know, typical attack 2
Speaker:years ago. Now it's very easy to detect. Still peep some people use
Speaker:it. And that's it. And that's that for us. It is very easy to do
Speaker:that. But probably, I mean, this is not a real person. Some of you trying
Speaker:to use the printed out images. This is
Speaker:Fraud. We can actually or reject it or or ask a person. Can
Speaker:you well, I mean, we need your real real pay real real image.
Speaker:Or we can just tell our customers that, you have to take a look because
Speaker:there was something fishy going. And then it goes and goes and goes. And the
Speaker:whole customer journey, We try to make sure that the fraud is not happening. This
Speaker:is basically it. So
Speaker:fraud is kind of, I think, Cyber fraud or whatever the cool
Speaker:kids call it, I think is has has infected
Speaker:every industry. I mean, if I just I
Speaker:mean, I I get 2 factor authentication logging in the
Speaker:roadblocks, like, for my kids. Right. And I'm like,
Speaker:they'll they'll they'll they'll get in front of their device, and they'll be like, can
Speaker:you tell me what the passcode is that they texted you? Like, Sometimes
Speaker:some days it's the only way I see 1 of my kids. But,
Speaker:has the because I I wonder, like, has the pandemic kind of Accelerated
Speaker:kind of virtual fraud, or is that just independent?
Speaker:I think it I think it is. Because it, right now, it's but it's not
Speaker:Related to fraud. Exactly. But the thing is is that now
Speaker:people are used to actually work
Speaker:remotely, Or it's so it's not that common for you
Speaker:to go to bank in person. So you just call there. You just I
Speaker:mean, use over the internet, basically. It's like easier
Speaker:So and now you can actually, there is no way, you can actually verify
Speaker:that this is the only person. Right. Yep. And this is a final thing because
Speaker:for instance, in Germany, where I reside, most of the
Speaker:time. There is a regulation called it's called
Speaker:video ident. So for in Germany, in order For for
Speaker:me, if you are going to open an account, anyway, I really have to call
Speaker:a person, a live in person operator, And talk to him, and he makes
Speaker:sure that or she makes sure that, a a million person. But everybody
Speaker:do not like it, basically. Because, I mean, it it takes time. You have to
Speaker:talk, talk to a person. I I just want to open an account. So it's
Speaker:it's it's it's fast as I'm but but except Germany, all of the rest
Speaker:of European Union, I think across the world as well. It's, I mean, you
Speaker:just Send your image or video, some of your documents,
Speaker:and then the the account is up. So it's very easy. And people get you,
Speaker:getting used to it. And that's why it's easier to to to actually,
Speaker:do fraud because it's, I mean, it's it's a soldier to trade off,
Speaker:Make it easier, and then it's easier for fraudsters to actually do their business. So
Speaker:that's that's the thing. Gotcha. Do you see,
Speaker:you mentioned you see, Like, new scams, people
Speaker:are running as well. And you also mentioned a lot
Speaker:of what I I thought would be pretty effective ways to to
Speaker:combat those scams, without really
Speaker:giving anybody any ideas. Are there, like, brand new
Speaker:scams that have happened maybe in in the very recent past
Speaker:that, you're still working on ways to combat?
Speaker:I must say that, there is there will always
Speaker:be some sort of, you know, arms, right.
Speaker:Competition? Yeah. So you have to say or. There will always be,
Speaker:like, a new prod Of yours.
Speaker:And then we have to actually deal with that. But I can tell you a
Speaker:story. So for instance, like, so we asked him so not a big company. Yeah.
Speaker:The technology team is not that So big, we have to move fast. But
Speaker:in my team, the AI slash, ML, it's not
Speaker:anti money laundering, but artificial intelligence slash machine learning. We have
Speaker:a very small department aimed at creating defects.
Speaker:So we do not detect defects. We have to actually learn how to create
Speaker:them So you actually know how I mean, how people actually read Oh,
Speaker:that makes sense. So synthetic data. Interesting. Yeah. Yeah.
Speaker:And this is at and I can also tell you that I mean, and this
Speaker:is for me, it was, like, so sorry if, you know, a surprise because,
Speaker:Most of the like, let's talk about defects. So, yes, then what what is like
Speaker:recent type of fraud? Deepest, for sure. We had a report. I
Speaker:I think it, We published it 3 years 2 days ago or like
Speaker:yesterday on friends. So what's actually happening right now?
Speaker:And the thing is that deep fakes, They use usage of
Speaker:defects for fraud. It maybe it rest
Speaker:like 5 times. So like 2 years ago, like nobody actually
Speaker:knew so About defects. But now it's it's very easy to craft. It's
Speaker:very easy to craft. I mean, people like I mean, you are a fraudster. You
Speaker:have to actually, it's very rare
Speaker:prefer for you to just craft just 1 defect. It's usually something
Speaker:we call the serial fraud. You create like hundreds of defects. So now it's easy,
Speaker:very easy to create them. So now it's like a craft, like, hundreds
Speaker:of identities. And then I tried to bypass our security checks. So that's why this
Speaker:is like the recent trend. I mean, as so it's on the news,
Speaker:basically. And then we have to actually try to make sure that our solution,
Speaker:can detect it. And it's not sometimes, it's not that easy. Well,
Speaker:it sounds like, you know, there's there's stuff that people used
Speaker:years ago, and you've got that figured out. And it's probably not being used
Speaker:as much, at least alone. But now you've got,
Speaker:people coming up with, first, new ideas, and then second, they're
Speaker:doing combinations new plus older ideas. Is that
Speaker:accurate? But but, it is actually. And the thing is Okay. So,
Speaker:these are also like, Okay. Just imagine. We have a very
Speaker:sophisticated deep fake detector. So I I'm pretty sure that our, like,
Speaker:models are more or less, good. So, like,
Speaker:I mean, it's not 100% for sure. Mhmm. But what
Speaker:happens next? So can I actually, I mean, combat defects, 5
Speaker:years later? Maybe it's I'm so advanced. I so make like,
Speaker:our customers, like, ask us about it, like, once in a
Speaker:month. So what do you actually what is your plan, to talk about defects
Speaker:in 2 years. Right. Because now, you know, AI is like, it's very hard problem
Speaker:to solve. But here's also problem. There is a thing
Speaker:called mules. Have you heard about mules or money
Speaker:mules? This is, the the thing is
Speaker:that you actually go, hire a person.
Speaker:Usually, buy, pay some €50. And then
Speaker:actually this person passes a KVST check for you.
Speaker:And then Oh, wow. The person just sells sells here his or her
Speaker:account to you. And then this is a real person. I mean, it's not a
Speaker:defect. I found it that I could defect. Wow. It's not obvious and not defect.
Speaker:Yeah. But that well, this is that looks suspicious. But
Speaker:but I if I'm in a bank, I'm in a I'm a bank, for me,
Speaker:it's like a real person just trying to open up in a bank account. Yeah.
Speaker:And now we actually have to look around. So that's why so I
Speaker:like working with Deepgrams. I mean, it's very, you know, cool technology. You have to,
Speaker:like Yeah. It's technology. But Now you actually have
Speaker:to look around. You have to make sure what is, I mean, the
Speaker:pattern. What are the devices do you use? It's like lots
Speaker:of small Features or, signals, you have to actually
Speaker:combine or merge them altogether and then make a decision. Is it, like,
Speaker:specia or suspicious sorta? And this is like, but this is fun. This is
Speaker:like, you have to really look around, look collect lots of data, and then try
Speaker:to find, you know, your way into making a decision.
Speaker:Interesting. It's it's it's a fascinating the simple things are no
Speaker:longer simple. Right? Just signing up for an account, You know,
Speaker:it's just now it's become like this massive multinational worldwide
Speaker:cyber Security kind of exercise. It's a
Speaker:fascinating, Yes. For a
Speaker:customer, it is it must remain easy. Yes. I don't know like
Speaker:I mean, since, like even, you know, the really, really
Speaker:typical KBC check is includes recording your
Speaker:video. You usually have to do something like, you know, turn your head
Speaker:or something. I mean, if you have this experience. People do not like it. For
Speaker:them, it's like, why do you have to do this? That's it's it looks strange.
Speaker:I mean, just can I just open an account? And then it's like so it's
Speaker:also trade off unless you have to be simultaneously
Speaker:secure and busy. And this is Yeah. Those those are
Speaker:those are very much contradictory, forces. Yeah.
Speaker:Well, the other thing too, like, if I'm if I'm If I'm an average
Speaker:customer or paranoid me. Right? Like, if I go to a
Speaker:thing and they want me to look this way, look that way, Am I training
Speaker:their deep fake model of me? Do you know what I mean? Like, I mean,
Speaker:I'm kinda like, you know, obviously, I've done a lot of live streams and stuff
Speaker:like that, so I shudder Better to think what you know, where that could lead.
Speaker:But, what are your thoughts on that? Like, I mean, are do do you have
Speaker:people who are Do savvy customers
Speaker:do they get a little suspicious? Like,
Speaker:what are your thoughts on I'm not. I I
Speaker:must said that I mean, the defects that we see, they they
Speaker:can be crafted just for 1 1 image. Right. So like,
Speaker:here's the problem. So so like, there are, none of that, I mean,
Speaker:you can see them, but Usually, people send, you know,
Speaker:low quality images. So it's even harder for us to see it. Even harder for
Speaker:for human person for human to see that this is a problem.
Speaker:But there is also, I think, if I find a story that I
Speaker:know, that some of our models
Speaker:actually detect defects better than humans. So
Speaker:it's actually easier for a fraudsters to treat a leading
Speaker:person than a model. This model, like, can look back from certain artifacts with
Speaker:eyes or just, like, some sort of, you know, glitches.
Speaker:It's easy. But for person, especially the quality of the image is It's bad.
Speaker:It's like there is no way anybody can actually spot this is the
Speaker:problem. And this is great. It it is a problem. I I I must I
Speaker:must admit this is, I think, this is what we
Speaker:actually have to be have to hear
Speaker:about about creating deep fakes. I know that that
Speaker:is a very interesting thing. So, you know, about I mean, there are lots of
Speaker:things happening, around AR regulations, Especially in the
Speaker:European Union. Sure. And then so we actually tried to follow and then to
Speaker:make sure that everything is compliant. And actually, I wanted to say that we touched
Speaker:upon k y c KYT, which is know your
Speaker:transaction. There was also KYB and all your business, which is basically, you
Speaker:know, how we make sure that the company you work with is is
Speaker:I know fraudsters. And there is also a thing called k y
Speaker:a I, know your AI. And it says about transparency.
Speaker:So many people out there want to be to know actually how AI is used.
Speaker:So the k l it's it's a very new trend, I think. You have never
Speaker:heard about it because, I mean, it was going to be a week ago. Since
Speaker:I like, I want to actually know what's happening with all of this model of
Speaker:error, not just about touch prod, ground everywhere. But back
Speaker:to the problem with defects. The thing is,
Speaker:what to to say that,
Speaker:Oh, sorry. I lost the my my train of thought. But this is the all
Speaker:the time. Yeah. We I was just about to say that. But what you know,
Speaker:one solution to this, I I think, Pavel, would be
Speaker:if people did something, you know, like, I don't know, colored their
Speaker:hair Or grew a cool beard. I'm just
Speaker:throwing that out and with apologies to people listening and not
Speaker:watching. No. You know? I'm just
Speaker:saying. But but if you did but if you did
Speaker:grow a beard, would would or or or change your hair color or
Speaker:altered their face? Like, I know that, like, facial most facial recognitions
Speaker:use landmarks on, like, the eye sockets. Right. The a lot harder to change I
Speaker:was joking. Didn't mind. But, like, would it would it would that
Speaker:I don't know. Like, does that have any impact on these kind of systems or
Speaker:are they more like facial recognition systems? They are,
Speaker:it's, so we operate on the if you're talking about defect detectors or
Speaker:defect, models for defect detection. Yeah. There are
Speaker:some, I can't say that I face recognition. The
Speaker:models, they mostly focus on artifacts. So so for
Speaker:instance, like, a defect of a year ago, usually,
Speaker:had problems with eyes. Your eyes of a defect, they usually are
Speaker:very, you know, not really human.
Speaker:So it will be changed. It will be like as as as the technology,
Speaker:is getting more advanced. But like a few years ago, you can actually just crop
Speaker:Eyes of an image of a person, pretending to be a human person, then they'd
Speaker:make sure that this is actually a defect. Also I must say that
Speaker:Yeah. So a video is is is easier to detect because you can actually
Speaker:so, there is a thing called, I don't like the term in blindness because
Speaker:No, but nobody actually know what Linus is, but Linus is a detection.
Speaker:Linus detection is detection. If this is a
Speaker:leading person or not. And before, like, 5 years ago, it was
Speaker:mostly a distinction between, a video of a person or
Speaker:a printed out image. Now it's a detection of an image,
Speaker:defect, and the linear person. And at that time,
Speaker:you actually there are 2 types of fly misses. One tool that's passive,
Speaker:and we actually use also sometimes our customers actually ask us for
Speaker:pacifying. Let's adjust 1 image. But it's easier for
Speaker:us and for everybody else to ask a person to actually do something.
Speaker:And for defects, for instance, like, if I ask them to rotate, Sometimes some
Speaker:artifacts can appear. Some artifact. And then you can actually see that probably. I
Speaker:mean, this is not the only person. There are some sort of problems with visual
Speaker:artifacts. So it is it is like this.
Speaker:Also, I must say that there was also a challenge for us because there
Speaker:are, certain cameras. They have some sort of a
Speaker:beautifiers. So I'm pretty sure as I'm calling from my,
Speaker:my computer, and then my camera actually
Speaker:Advances my image. So my image is a little bit, better
Speaker:than I'm in the real life. So my my skin is is is a little
Speaker:bit better. So it's it is actually, Embedded into
Speaker:hardware. And for us, it looks like, some sort of, you know so there is
Speaker:a signal for us. It does some sort of, you know it's Oh, I see.
Speaker:So It's hard. You know? And you have to make sure that make sure that,
Speaker:okay, it's not defect. It's just the person using that, camera off my,
Speaker:computer. It's like, you know, you have you have to be really, a
Speaker:yellow error. Apple, I mean, installs
Speaker:another camera, and then you have to be actually tune your models to make
Speaker:sure that you actually do not penalize people from with
Speaker:I think about that. Yeah. The cameras are gonna behave differently if you use different
Speaker:cameras. So I'm here using my 4 k,
Speaker:camera. Kind of an outdated one, but it's still it does the job. But what
Speaker:if I pick up my droid Or, you know, my wife
Speaker:my wife, you know, she's the the device. She's got an
Speaker:iPhone. And if I'm trying to log in through her device, That would be different
Speaker:images, and it may change. You know, it may tell me, nope. That's not
Speaker:you. Those are gonna be different artifacts. That's fascinating. And I also
Speaker:think it's funny that you have an old four k camera, which
Speaker:is a pretty funny thing to say. Like For for podcasting, I won't
Speaker:No. I know. I don't wanna throw back to, theme from yesterday's
Speaker:2 hour show, but I'll just make this note. We we
Speaker:learned that we're in the top 2 a half percent of podcasts.
Speaker:So now I feel like I should have, I don't know, 16 k studio
Speaker:and Yeah. I should have a lot of time like Joe Rogan has in a
Speaker:brick wall. Exactly. Right. I don't I need something better than this
Speaker:old four k camera. But
Speaker:if all of a sudden You just want to open a bank account right
Speaker:now. Yeah. It looks strange because, I mean, a typical person is like you
Speaker:use your iPhone or you're like a regular computer. Like, with 4 k or 16
Speaker:k camera, it's like very strange. It's some something, you know. It's it's a signal
Speaker:for for every model and make sure that It's an outlier. Right? And
Speaker:it sounds like a big this is still obviously, there's way
Speaker:more complicated things than what you do, But outliers
Speaker:detecting outliers is probably 1 1 big tool in your tool belt.
Speaker:It is. Yeah. That's very hard if you have a Genuine person,
Speaker:and you are an outlier somehow. I mean, everybody can be an
Speaker:outlier in some sense. It's very hard because, yeah,
Speaker:So this is hard. So, like, at some point, yeah, colored hairs
Speaker:can be also an outlier. I don't No. It's just interesting. So I imagine, like,
Speaker:Instagram filters and things like that probably also cause
Speaker:chaos and things like that. Yeah. Of course. But, yeah, I
Speaker:mean So usually use, yeah, filters,
Speaker:a strong signal for us. I mean Right. And also I must I must have
Speaker:this defects. So going back, thing with defects is that
Speaker:it's not, like, specifically use the fraudsters. Here's the
Speaker:problem. You know, there are lots of cool things for defects. You can press
Speaker:advertising. Right. I don't know what what else. But, usually,
Speaker:you can actually adopt a person to, like, Replaced an
Speaker:actor in the movie. This is also a defect. It's a very cool defect, very
Speaker:sophisticated defect, very high quality defect. Still a defect. So those
Speaker:are our usage is actually for for that, I mean, not just for fraud.
Speaker:And then going back to our problems, it's like, I mean, And the
Speaker:even even that and even that from that, I like this example,
Speaker:but, the guys from the,
Speaker:I mean so we focus on financial fraud. Yeah. So it's more or less like
Speaker:people trying to actually sue money on, like, take over your account, something like
Speaker:that. But the thing is the defects, they are mostly created
Speaker:not for that. And this is a very interesting thing, I think. They are created.
Speaker:And, actually, I didn't know about that, but we actually knew that When they started
Speaker:to try and to create our Deepak's. So we went, you know, to the Internet,
Speaker:some strange forms to make sure what what people actually use
Speaker:What they create deep eggs for. And they create
Speaker:deep eggs for porn. It's like 98%, 89%
Speaker:Deepex, I slide 4. And this is also a problem because in in there is
Speaker:a thing called nonconsensual port. Deepex are used for that, And this
Speaker:is also a problem. So it's not our business, but the thing is that the
Speaker:same technologies is there. And you actually I mean, if you,
Speaker:I mean, work in the area, you can actually so the same model can actually
Speaker:be applied to detect, this type of defects. Right. So it's
Speaker:different, but, I mean yeah. Yes. It's, That was expressed to
Speaker:me maybe a year ago. It's fascinating how
Speaker:quickly this space is just Evolving or
Speaker:devolving, I guess, depending on your point of view. Yeah.
Speaker:But, no, you're right. Like, most of it is
Speaker:Those a lot of the deep fake kind of work is done
Speaker:for adult content. And, you know, and it's there
Speaker:the The legislation around this is gonna vary
Speaker:widely from place to place. But, like, you know,
Speaker:revenge porn laws don't apply. And there. I I think that was a big thing
Speaker:in, and there was a controversy somewhere. I think it
Speaker:was New Jersey, Where somebody had
Speaker:created deep fake images of either high
Speaker:school or middle school girls, which adds an extra level of
Speaker:legal Concern I have a whole lots of extra
Speaker:levels of concern. Let's be honest. But, like, you know and and and and
Speaker:there was this, you know, the big debate. And my first reaction was, I'm
Speaker:actually kinda surprised it took this long for that to happen,
Speaker:which is a very cynical take, I'll admit. But I can tell I I can
Speaker:tell you the reason. The thing is that Technology moves so fast. Yes. And
Speaker:legislation actually is always, like
Speaker:so even with with EAU, AI act,
Speaker:those I mentioned defects just a little because they started working on
Speaker:the regulations 2 years ago. And 2 years ago, it was not a problem.
Speaker:And now it's, like, all over, you know, the Internet, and then you have to
Speaker:actually tweak the, wording,
Speaker:but it takes time. Well, even still, like, you know, like, there's,
Speaker:a few months ago, they had these fake commercials that were created by with
Speaker:combination of 11 Labs and A few other companies to name them, so I
Speaker:forget. But, you know, they had a picture of Elon Musk, you
Speaker:know, eating spaghetti, and it looked weird. But you can easily see,
Speaker:like, You know, I was messing around with v q early versions of v
Speaker:q grant d q GANs in early
Speaker:2022, And that stuff looked
Speaker:weird, and it it really evolved. And this morning, I saw
Speaker:Pika AI, I guess, just went Yeah. Yeah. Yeah. Went to a wider beta.
Speaker:And, yeah, released and and and, like, I'm seeing what's created with that,
Speaker:and, you know, it still looks weird, it still looks cartoonish,
Speaker:but it's not The fact that we've gone that far in the span
Speaker:of, you know, less than 2 years, like, I think says something, like and to
Speaker:your point, legislation Usually takes years, to
Speaker:make. So, like, by the time these laws are written, they may not be valid.
Speaker:In the case of New Jersey, I think there's some debate over,
Speaker:does what sorts of laws that applies to? Because
Speaker:the the original, The faces
Speaker:were mapped on to something else, but that the
Speaker:something else I'm trying to keep our clean rating here. The something else were
Speaker:people over 18, but the bases were mapped onto it. So there's
Speaker:some debate over, do existing laws cover that?
Speaker:I'm not a lawyer. Don't look at me, and I'm not. But,
Speaker:it's just fascinating to your point. Like, this is moving quickly.
Speaker:Yep. It's definitely complicated. So we've
Speaker:reached the point in our show, Pavel, where we, like to
Speaker:ask a set of questions. They're in the chat. And
Speaker:I'll start out, with the, the very first question.
Speaker:How did you find your way into this field? Did this field find you,
Speaker:or did you find it? Yeah. I must say I have a
Speaker:story to tell. I just studied yeah. Studied computer
Speaker:science at, university And I actually worked as a software engineer
Speaker:at Motorola. You may remember this company, with
Speaker:HQ in Chicago back then, for 5 years.
Speaker:And then it was, 2011, which is, like, long time
Speaker:ago, the very first, massive
Speaker:online courses appeared. There was a one called AI class,
Speaker:and it later turned out to be a Udacity. And there
Speaker:was also a m l called ML class. It's a ML class. And
Speaker:this now this Coursera. It's like 10 years ago. And I was like, okay.
Speaker:Cool. I enrolled and actually, I pushed because it is like it was it was
Speaker:hard. It was like, you have to really, be involved. And
Speaker:then I felt like, okay, this is a cool thing. This is like a next
Speaker:big thing for me and, like, for everybody else. It was like
Speaker:12 years ago. So I quit my job, and I actually, so
Speaker:at the same time, I started to try to run a small startup with my
Speaker:friend, failed miserably. But I take, took my time, studied,
Speaker:for maybe half a year, and then joined a small data
Speaker:startup as a data scientist. And then it just
Speaker:started there. So it's I think I I find, my way into
Speaker:data. But Yeah. I don't know. So You want to
Speaker:I'm sorry. Go ahead. I just I just say it sounds like you were very
Speaker:intentional about finding your way into it. So that's cool. Yeah.
Speaker:That's cool. And I see you were You were at VK for a while too,
Speaker:which I've never seen VK, but I hear it's like a like
Speaker:a Russian language version of Twitter slash Facebook. It used
Speaker:to be. Yes. Yeah. Yeah. I don't I yeah. Obviously, now things are different, but
Speaker:yeah. Yeah. Yeah. Yeah. I worked there for 5 years, a long time ago. Oh,
Speaker:interesting. And, you know, if you're talking about the data, I mean,
Speaker:the, where it's like the the place where you can
Speaker:actually play with data. You can actually cool do many cool things.
Speaker:Oh, yeah. Nice. Nice. And he's being modest. According to LinkedIn, he was director
Speaker:of AI research, so he's super smart.
Speaker:But, what's your favorite part of
Speaker:your current job? Oh, I can't say it
Speaker:could create some defects, but, it's not
Speaker:it. I think
Speaker:no. I mean, I would say that what I like is, they,
Speaker:the the Samsung, Samsung is is now it's it's a product or any company. So
Speaker:have our own own products, whether, like, a technology company, yet we have our
Speaker:own product. And having that,
Speaker:actually, our own product, Actually helps us, you know, I know what our
Speaker:customer wants. Wonderful. I know the
Speaker:data. So it's like, you know, I mean, you have to actually so you have
Speaker:to look around. Okay. There is a problem with defects. I have to,
Speaker:like, make sure that I mean, I had, I actually have to understand this. This
Speaker:is a problem. And for many of our customers, I mean, I
Speaker:don't I would not like to say that we have to educate them or actually
Speaker:make make sure that they understand this is a problem with defects. And now we
Speaker:have when they understand, we can actually help them with their their,
Speaker:safety and security. One thing that this is, like, a little bit, I
Speaker:mean, Clumsy answer, but I'm sorry if you know.
Speaker:Yeah. Being closer to the product is is is is fun.
Speaker:Oh, sorry. Cool. So we have 3 complete
Speaker:sentence. And the first one is when I'm not
Speaker:working, I enjoy blank.
Speaker:Okay. Okay. Let me think for a while. There are many things I can
Speaker:say. No. I can say no. This is I think of this as I can,
Speaker:I can share? No. I I I I run or I can see job.
Speaker:Mhmm. Oh, cool. Cool. I run-in the the ring marathon.
Speaker:This is my Nice. There are Major Martins, like, 5,
Speaker:6 Martins across the world. So that's New York, Paris,
Speaker:London, Tokyo, Berlin, and,
Speaker:London. Nice. Like, 6 so that Very So Berlin was my 1st major
Speaker:marathon. So I ran it, this this September, and it was great. No. That's
Speaker:awesome. That's awesome.
Speaker:When you said Berlin, the first thing that popped in my mind was, Berliner
Speaker:Kendall wrote, which is like this local kinda drink.
Speaker:Yeah. Yeah. Yeah. Yeah. I know. That's like Yeah.
Speaker:Yeah. But I prefer there is a it's a vehicle. It's
Speaker:like a craft. Oh, yeah. From Berlin.
Speaker:Right. But I talking about Berlin, so I run. It was
Speaker:super fun, but, on my finishing picture, so
Speaker:it's my me, Ryan. So close to Bernsberg. It's a
Speaker:very central grid. Mhmm. And there is also a guy in the
Speaker:bottle question. And and I
Speaker:wasn't it was not slow. I wasn't slow. Yeah. There was a guy in a
Speaker:huge ball, like, I still running, like, finishing with me. Like, so it was, Oh,
Speaker:that's funny. That's fun. It's that's fun. That's funny. Very cool.
Speaker:Next, complete the sentence. I think the coolest thing in
Speaker:technology today is
Speaker:blank. Oh, it's it's it's hard to say. Let me I'll just
Speaker:think for a while. But, I mean,
Speaker:I think that so my my area
Speaker:seems like I expert a personally specified natural
Speaker:language processing. So I know about language models. And,
Speaker:actually, we had papers on language models, like, before they they
Speaker:were super big. So, like, on tuning language models. Yes. I
Speaker:found it really, really exciting that it in a
Speaker:year, it went from, you know, research
Speaker:Prototypes to, like, everyday product. This is Yeah. This was
Speaker:a compelling. So, like, my parents used Chargebee PCs. Like, I mean, this
Speaker:is like this is like a mobile phone. This is I mean, this is what,
Speaker:like, some sort of a milestone, last year.
Speaker:I think this is this is it. And he is that the actual unit
Speaker:for main things. You can build products on on language models. And
Speaker:this is also like. It's wild, isn't it? Like, you know,
Speaker:and and it's captured everybody's imagination in in good and bad ways.
Speaker:But, like, my father-in-law, you know, So he used to
Speaker:say Frank works with computers. Now he says Frank works in AI.
Speaker:Okay. You know? That's good.
Speaker:But I also like we used to say machine learning. So now you have to
Speaker:say AI. That's right. That's right. You have to say that data mining
Speaker:core something. So it's like, you know That's right. It definitely would.
Speaker:I wonder what it'll be next year. Who knows? Gen AI probably.
Speaker:Probably. So our next one, complete this Regulate, I think. Oh, that's
Speaker:right. Regulation. That's right. Regular. Our our last completes the
Speaker:sentence is I look forward to the day when I can use technology
Speaker:to blank. Uh-huh. I
Speaker:can't it's hard to answer because, I mean, like, I
Speaker:can't say it would be cool If I can, you know,
Speaker:develop drugs. And then there are very cool startups for drug design
Speaker:with AI. Yet, I mean, Just imagine we have
Speaker:a a a cure for cancer, but Right. We have so
Speaker:many diseases to care to cure. So let's say, I think I
Speaker:hope Once we fix anything, then there is gonna
Speaker:be a next, you know, next milestone for us to look forward. So I'm sorry
Speaker:if, you know, there's never I hope there will be
Speaker:no such date, I can say. Right. Right. That's
Speaker:a good one. I'm pretty sure you will agree with me. Like Yeah.
Speaker:Especially work with the technology. I mean So true. For sure.
Speaker:The next question, share something different about yourself, but remember, It's a family
Speaker:oriented well, not family oriented, but we like we we like it so that
Speaker:you can list it with your kids in the in the car. Right? Like, That's
Speaker:kind of a Yeah. Yeah. Yeah. And, yeah, and I live in Berlin across, very
Speaker:close. There's a very, how to say, kinky club, which is Berlin.
Speaker:Was that the the the tier garden? It's it's
Speaker:it's a it's a it is family friendly. It's it's like the most family
Speaker:friendly place in in in Berlin. You got some. Yeah. No. It's it's
Speaker:called KitKat. Yes. What I can say.
Speaker:I have, purple hair. Since last month.
Speaker:I don't know. So I can say that I speak
Speaker:a few languages, all of that. But, no, I'm I'm
Speaker:joking. So I speak Japanese. I don't I don't Japanese, for a
Speaker:long time. So I I can speak Japanese. I speak
Speaker:English, obviously, Russian. My parents are from
Speaker:Russia. And I also speak German. So I actually Studied
Speaker:German for 2 years. So I actually studied right now. So I had, like, my
Speaker:German classes 3 or 4 times per week, which is
Speaker:let me just go. Sorry. So I hope in a year, I will be able
Speaker:to do a podcast in German as well. Oh, Wendeschon. That is
Speaker:not
Speaker:Yeah. Yeah. And we just lost, like, We we just
Speaker:looked at our analytics, and, like, most of our listeners are from English language countries.
Speaker:So I think we just lost them. Maybe we
Speaker:can attract new listeners. Oh, I like it. I like the way you think. We
Speaker:wanna we wanna get to the top 2.4% now.
Speaker:Our new goal. So,
Speaker:Audible is a sponsor of the show, and I'm not sure if
Speaker:Audible is big in Europe. I think it is because I've seen a lot of
Speaker:German language audiobooks. It is a no. Okay.
Speaker:So do you do you listen to audiobooks? And if so, you have a good
Speaker:recommendation. Otherwise, we'll take a recommendation on the regular good Fashion
Speaker:paper dead tree book. No. I have a couple. I think I can
Speaker:give you a couple of examples. This is like,
Speaker:I like this was the most, you know so so I'm so my
Speaker:background is from many, places,
Speaker:since Israel, Russia, and Germany in some extent. So
Speaker:I would recommend, there is a very Good book. It is
Speaker:in my opinion, this is very known, but not many people know about it for
Speaker:some reason. It's called the good soldier's make. Okay.
Speaker:Like it said, didn't Not heard of. About the, sort of
Speaker:third world war by Oh, interesting. But this is
Speaker:it's very good. Like, you can actually learn a lot about
Speaker:Czech Republic, Germany, Austria in the beginning of
Speaker:the, Last century. Oh, interesting.
Speaker:Especially now, it's the very thing. It's called in the park. This is a very
Speaker:good thing too. And it's very funny. It's like one of the funniest, books
Speaker:ever written. And also the the second one, I have 2.
Speaker:This called Arc of Triumph, by remark. Okay.
Speaker:This is also about the pre war Europe, pre second World War
Speaker:Europe, like, Southeast, years of the
Speaker:last century. And this is also very, like, you know, you
Speaker:really you really feel like what what was the I mean, living in
Speaker:Germany and, France, during that time, it's very, very interesting.
Speaker:So one of my favorites. So I can definitely recommend both of these
Speaker:videos. Very cool. So audible detecting I'm sorry. I'm
Speaker:detecting a history theme. Yes. Yeah.
Speaker:Yeah. Yeah? Cool. There's a really good book. Since you live in Berlin,
Speaker:you might like it. It's called Faust's Metropolis, and
Speaker:it's about the history of Berlin from, like, you know, Almost
Speaker:stone age time till Okay. Cool. You know, the 20
Speaker:you know, early 21st century is kind of like And the basic
Speaker:gist is, like, you know, a lot has happened in Berlin. Good.
Speaker:Sure. Yeah. We all know the bad. Right? But, like, some good things have
Speaker:happened, kinda everything in between. It's kind of it's an interesting look at, like, the
Speaker:history of the city and how it apparently was built on a swamp or something
Speaker:like that. Like Yeah. It's just, it's it's
Speaker:interesting. And Audible is a sponsor of
Speaker:Data Driven. If you go to the data driven book .com,
Speaker:I think even the data driven book .com might work. Uh-huh. That was
Speaker:a pronunciation joke. You'll get a free, on 1 free
Speaker:audiobook on us, and And we'll get a kickback if you sign up for a
Speaker:subscription. And finally, where can
Speaker:folks find out about you, more about you, and what you're up to at Sumsub
Speaker:And, some of the other things you you're up to.
Speaker:What's up? My my connection was, Oh, where can folks find out
Speaker:more about you and what you're up to? Oh, yes. It's,
Speaker:yes. It's, it's a company. It's called Samsung. So Samsung dot com.
Speaker:Also, like, what we have is, today is
Speaker:with anti fraud. And you have to I mean, It's not about
Speaker:all the product. It's actually about making people helping people
Speaker:learn about, security. So how they can actually navigate the Internet or,
Speaker:like, their life More safely. So we have a portal called
Speaker:some suburb where we actually post a lot
Speaker:of stuff on Making your Internet life,
Speaker:can I say like this, safer? So, actually, I I advise you
Speaker:to take a look, and then probably you'll find something interesting there.
Speaker:We definitely will. And, any parting thoughts before
Speaker:we end the show? Any final thoughts? I just
Speaker:want to say, yeah, Just I was very happy to, to be here
Speaker:and hope, it was Cool. Interesting. This is a great show. It's always good to
Speaker:it's always good to kinda understand The the the intersection
Speaker:of of AI data and security because some people still see
Speaker:those as separate things. But I think as time goes on,
Speaker:we're gonna I'm gonna we're gonna wonder how we ever saw it as separate
Speaker:things. There are so many things to talk about that. Yeah. Yeah. Yeah.
Speaker:Yeah. Well, awesome. Any parting thoughts, Andy?
Speaker:No. Just a great show. Pavel, thank you for, for joining us.
Speaker:It was our honor. Yes. Likewise. And we'll let
Speaker:Bailey finish the show. That was some show.
Speaker:We appreciate you listening to Data Driven. We know you're
Speaker:busy and we appreciate you listening to our podcast. But
Speaker:we have a favor to ask. Please rate and review our
Speaker:podcast on Itunes, Stitcher, or wherever you subscribe to
Speaker:us. You have subscribed to us,
Speaker:haven't you? Having high ratings and reviews helps us
Speaker:improve the quality of our show and rank us more favorably with the search
Speaker:algorithms. That means more people listen to us,
Speaker:spreading the joy. And, can't the world use a little
Speaker:more joy these days? So go do your part to
Speaker:make the world just a little better and be sure to rate and review the
Speaker:show.