Transcription
[Music] Good morning and welcome back to the Mind Walk podcast. The podcast where we talk about everything new and innovative in the natural resources industries with a specific focus on digital transformation.
As we get started, allow me just a quick shout out as always to MinRP, the primary sponsor for the Mine World podcast. MinRP is a mining software company specializing in the application of software solutions for mining aimed at increasing interoperability between mining technical, financial, and ESG domains, making mining better. So for more information about MinRP, please visit www.mmin.com.
So with that behind us then, uh, let's start today's podcast and we've got a very interesting guest for you. If I read his LinkedIn profile, it says, "I'm a cognitive data scientist and IBM master inventor with a focus on natural language processing and artificial intelligence." So if you, if you don't understand his accent, then uh, maybe you can use his product to, to help you understand. Welcome to MindW, uh, Stefan Fonder Stocks.
Thank you very much. Glad I'm glad to be here.
Yeah. So, so what on earth is a cognitive data scientist and how did you become one?
Yeah. Um, well, I guess, um, data science is, you know, a wide field, right? So there's many, many different, um, you know, just aspects of it and to it and different application areas. So, cognitive data scientist just happens to be one of those that specialized, um, in the application of technology such as natural language processing, uh, to do something useful. Uh, a key example is, um, IBM's Watson system that, um, if you may recall, it won Jeopardy a long time ago, um, in 2011, um, and it was just, you know, a perfect example of how a computer outperformed the best of human ability in terms of, uh, recall of information. That's a, you know, a cognitive system and pretty much my job is how do we build them? How do we measure them? How do we improve them, um, and make them available as, uh, application software?
Uh, you are with IBM. Why don't you just quickly tell us what your, you know, what does your day-to-day job entail at IBM?
Yeah, so, uh, my current role is I'm the lead data scientist and practice manager for the trustworthy AI practice in expert labs. So that's a mouthful, but fundamentally I lead a team of 11 data scientists. Um, and our core focus is on trustworthy AI, right? How do we take the black box of a model and try and, you know, uh, just unpack that and to see and explain to people who aren't data scientists, like you and me, when we're customers of a client or a line of business manager, for example, you know, just what that model does. And there's seven key pillars around that, you know, how do we know the model is fair? How do we know it consistently achieves business goals? How do we explain the outcomes of such a model, you know, in terms of its inputs? How can we be confident in the model's predictions and when should we trust it and not trust it, right? When, when the model isn't sure, how do we know that? Um, how do we know what the key decisions and development decisions and, um, documentation around the model was? Who built what, when, where, with what kind of stuff? How do we know it's private? Privacy preserving it. You can't hack the model. There are mathematical ways to extract your private information from such a model if people aren't careful. So, how do we be careful about that? And then lastly, can this model be protected from what we call adversarial attacks? You've all seen examples where this is a photo of a cat and the AI says it's a cat, and then you add some white noise, and then all of a sudden it's ice cream or something, right? That's an example of an adversarial attack. So, yeah, my my practice focuses on taking that capability to market and helping people avoid those pitfalls.
Right. And and and this is applied. Now, you've used the example of image recognition, you know, there with a cat, but this is equally required for natural language processing, and as it is for image and then video processing. What else? What, you know, what's the scope of this of the science?
Fundamentally, most businesses, you know, they follow a distribution curve of what types of AI they use. And we'll get into some definitions in a bit, but most of the, the huge Fortune 500 companies are fundamentally doing classification, regression, uh, or, um, a little bit of optimization here and there, right? Supply chain, that kind of stuff. Just by sheer volume, most of it is classification-based, like, should I give you a loan or not give you a loan? Should I hire this person or not hire this person? You do get some interesting use cases like, should I pack this strawberry into that packet or not? What is fair? That I've seen in retail. So the definition of fair is usually applied in a social context, but it could also be fair in terms of some constraint, right? So, it really does the, the science goes very deep, very far, but the requirements from businesses are just scratching at the surface of what is possible, and they'll progressively get deeper as the tooling enables us to do so.
Right. Just before we jump into that, I want to just talk about your credentials a little bit. I've spoken to many AI and advanced analytics people who have, who typically come at this problem from a mathematical perspective. Um, you know, you've come at it from a computer science background, or at least your academic background. Is that right?
Correct. Yeah. So, I'm, I'm one of those odd ducks. I joined IBM right out of university, but throughout pretty much almost all of my 18 years at IBM, I studied part-time. So, I did first my master's degree, uh, with the University of Pretoria, way back in like 2006, I think it was when I started. It was all part-time. So I focused on neural networks and how to make that a generically available software package with design patterns, and that was fun. Learned a lot about neural nets. I then took a couple of years off by just working like a normal person. And then I started, uh, my PhD in, um, hyperheuristics and optimization using, uh, something called swarm intelligence and evolution. So, nature-inspired algorithms. For me, that's, that's, that's the pinnacle of AI, AI, right? When we're mimicking nature's ability to make things better and be smart, but we necessarily don't understand how it does that. We just know that it works, right? That really hooked me. So, yeah, I did a PhD in that, um, focusing on hyperheuristics that dynamically change optimization algorithms on the fly as it solves a problem. So, it's like an alt-tab button, if you want to think of it that way. There's change methods if we notice the current one doesn't work. And just to make my life harder, I said the problem also changes while we're solving it. So, truly dynamic problems, I characterized the types of change that are possible. There's only 27, by the way. And then which algorithms work well in which types of environments, you know, sometimes random is a better strategy than anything if it's like very chaotic changes and so forth, right?
Yeah. So, I mean, just on that, uh, well, you said we'll get to defining some terms. So let's talk about those definitions quickly. First definition, heuristics. What are heuristics?
A heuristic is an interesting word, right? It comes from the original Greek. I think it's pronounced heuristikos. It's essentially a guess, a best guess effort. Um, some problems are so hard that we cannot, we as humanity, with all the computing power in the world for millions of years, cannot work out what the correct answer is using pure math and algebra alone. It's just too many steps. Uh, it's going to take too long. An example of that is if you have 100 deliveries to make in your city and you get in your car and you say, "What's the shortest route?" It sounds easy, but the whole internet cannot compute that for you. You have to guess what the best one is. So heuristics are methods that we employ to systematically guess so that we can be reasonably sure it's close enough to good enough, or but it, you can never be sure that it is the best way. So that's the, the essence of the word heuristic. It's how do we guess?
Well, yeah. And so what on earth is a hyperheuristic? And I've also heard the terms metaheuristics used in the in the same sentences.
Yeah. So a heuristic is a very fixed thing. Like you've all heard this one, right? When you're in a maze, always keep left. That's a heuristic, right? Because you're literally just blindly following some truth, right? And most of the time that will be okay, unless someone designed the maze not to work with that heuristic. So a metaheuristic is a parameterized heuristic. So you can tune it, if you want to think of it that way. That's that's a metaheuristic. There are many types and they, there are simple ones like, why turn left? Maybe it's turn left every second time or whatever, right? Yeah. A range of kind of seeding values from which to pick, correct? Yeah. And it gets very complex, like if you've heard of particle swarm optimization or genetic programming, all those are also metaheuristics. They follow some determinate logic, but the conditions that they work on might be stochastic or learned over time, but ultimately they are metaheuristics. We're tuning the algorithm to do the choice based on either what it sees or external state we put in.
So as humans, of course, we we use heuristics a lot. I mean, we use previous experience and learned, uh, background information to, to kind of see the way that we make decisions. But if we jump into the world of artificial intelligence and I guess machine learning, uh, more specifically, then we've got to give our learning model, um, some background data to start off with, don't we?
The context matters, right? So, one of the, the key quotes I've heard that that shaped a lot of my PhD is that is a quote by a person who helped build out the Scala Z library. And this is getting deep in the weeds of functional programming, right? Um, but he said, "Liberties constrain and constraints liberate." And there's a story around that. So a teacher was, uh, running a class of above your seven-year-olds, and every recess the kids wanted to go play outside. So the school building has this massive yard, but it's surrounded by four busy roads, right? And there's no fence. So the teacher always kept all the kids closed, cuz one could just run off and, you know, traffic could be a problem. So because there were no fences, it was actually dangerous for the kids. So the, the mother hen of the teacher instinct just kept them all close. But then someone saw this and sponsored a fence. So they put up a fence all around the area, as large as a football field. Um, and by the time, you know, this was done, she could then let the kids run free, and they had like a hundred times more area that they could play in because there were fences. Right? So constraints liberate, and liberties constrain is, uh, a fundamental quote. And AI is built on this very premise, right? If you bring your data of your business problem, and it could be anything, it could be like a Caterpillar truck, those big ones that, that you guys use in the mines, right? And you just have all the sensor data, all the driving, you know, telemetry, all that stuff. All of a sudden, you're coming out of this wide world of infinite possibilities down to, this is what the truck did in real life. Here's all the measurements. So now you're in a context, and the AI can learn from that because you've given it information. And now that's, that's the heart of machine learning is how do we extract the patterns within that data and try and use it in a predictive sense. So that's machine learning. You literally look at data and you mimic and extract the patterns and use it.
Right. So, I mean, uh, how did you get involved in this? You told us an interesting story, you know, about your summary of what happened just after Watson won the, uh, Jeopardy.
Yeah. So, it was a few years after my master's degree was completed. I was just working, uh, at IBM as a quote unquote IT guy, right? Which was fantastic. It gave me a solid grounding in enterprise software. And then at some point, IBM acquired the SPSS company, which is, you know, a world leader in predictive analytics and, and, uh, data mining, as we called it back then. I think that was around 2010. So I just learned it because I was curious and wanted to, that's what I wanted to do with my life. But at that point, enterprise software wasn't ready for it yet. With the acquisition of SPSS, IBM said, "We're going mainstream with this stuff." So I got it, and, you know, that was, that was good. And eventually, that became my day job in IBM as well. So, lesson one: always do what your passion is, and eventually your real life will follow you. But yeah, then, then a pivotal moment happened, right? Watson won that Jeopardy challenge, where it beat the two greatest humans of all time in that quiz show. For those of you that don't know Jeopardy, they give you the answer to a question, and you have your, your response has to be, "What is the question?" "This river flows through the Orange Free State and the Vaal River flows into it." And then MP says, "What is the Orange River?" "Correct. MP, you have done it." So, do I get to beat Watson now? Yeah. Now you've won some money. Um, and the question is, can you do that question quicker than Watson, right? Um, and that machine just obliterated the world's best champions of all time. So, it's, it's a precision recall discussion, you know, information recall architecture. And I just watched that in 2011, in February, I think it was, and I was amazed at what we could do. It's a grand challenge that IBM, you know, um, built, and people thought we're in a new era of AI, and arguably we are, right? Yeah. I learned everything I could about that system. You know, scanned research, all the materials I could get because I'm on the internet, and I just learned it. And long story short, you, you talk to your colleagues, and the marketing department finds out about it. And then I, I was offered the opportunity to go and present this for an hour at one of IBM's flagship events in South Africa every year, where 700 of the top business leaders were, um, you know, invited to. Um, and yeah, I, I followed the stage after Gary Kasparov was the key speaker before me, and I had to explain what, and I, I could not believe it. Right. That's, that's the power of IBM and why we work at this company. Yeah, they flew him out, and little me followed him to explain Watson, and that was still one of my greatest moments, at least, you know, for personal growth.
But little old you didn't really remain little old you. You know, I, I did a bit of a look up on Google Scholar, and I see that since, you know, for the past four years, you've been cited more than 302 times on Google Scholar in academic work, and and and similarly. So clearly, you know, it's not only IBM that took notice. I want to, I want to just go ahead and talk about this again, bringing in the, uh, the definition of heuristics. So we know that there's going to be background information against which, uh, something learns and that that it uses to, to make decisions. But then just before we move into how it applies that in a in a trustworthy way, you just used a very interesting term. You said, "We give this problem to the AI and the AI does XYZ," as opposed to to an AI machine or something. So talk to us a little bit about an AI agent. Why, why is it, why is it almost personalized?
We'd have to start unpacking the definitions here, you know, one level deeper, right? So I, I hinted at machine learning. And machine learning is actually a very old technique. When I mean old, I mean very old, like 1760, right? And earlier, it's fundamentally math and fitting curves, if, if you want to think of it that way, right? A typical example is what we call regression. So if I give you a house price, or, or rather, let's do it the other way around. The x-axis is the square footage of the house, and the y-axis is the price of a house. There's going to be a general trend upwards. It's not the perfect line, but it is sort of like a fuzzy cloud that's elongated along the diagonal, right? And you can fit a line in the middle there that minimizes the distance to all the points, and that we all did in school. That's, you know, just curve fitting, and that is the basis of something called regression. Now, what technique you use is that one-point regression, two-point regression, are you intersecting hyperplanes in there? How many dimensions are there? All of a sudden, it gets way more complicated, and we start calling it machine learning, um, if we're using heuristics to do that, for example, or, um, deep learning or matrix multiplication. But at the heart of it, you're, you know, it's math that we know, and we're just extracting and compressing patterns into simpler ones that we can reason about. So AI is, is a higher concept thing, right? Um, it's artificial intelligence, where machine learning is about a decision. You know, is it yes or no? Is it A or B? AI is the system that you build around it, where the goal is to actually learn how to make that decision. So that, that starts getting interesting, right? So machine learning is trained. You've all heard of this, you know, data preparation, data understanding, exploratory data analysis, development of the machine learning algorithm and model testing, all that stuff. That's the life cycle of machine learning. It's all manual. Humans tend to do most of that, right? If not all of it. But AI, you start going higher order and you say, can some of these steps be brought into the machine learning process itself? Can I let it teach itself? That starts getting freaky, right?
It does. Yeah. And but, but I guess that's the, that's the whole objective. Because humans are pretty freaky in the sense that we, you know, the kinds of things that we can do, and the choices about which, I almost want to say, which reasoning model to use when we make a decision and so on. That's not so easy to to build into traditional algorithms because you've, you've decided on its behalf. And if I understand you correctly now, you're saying, well, okay, maybe the machine gets to decide how it thinks about something and what it prefers, and which, which approach to prefer to other approaches, etc.
Yeah. And that's the problem, right? Because academics get excited about that. It's like, wow, that's amazing. But then business people get really upset about that because it's like, I don't understand it anymore. And there's the disconnect that we currently face in humanity. Is these AIs are getting more and more powerful, more and more, uh, integrated and self-learning. So things like the product recommenders on large websites that, you know, sell you, uh, products and deliver them to your door. How does it know what you need? You know, we've all heard the, you know, the, the myth, the fake news of, you know, um, "They're listening to my phone," right? And if I say something, then the ads are coming. They're not actually doing that, but they're so good at predicting your behavior that it seems like that in many cases. And, and that's just an example of them reinforcing what they know about you based on your behavior, what you browse, what you buy. And of course, based on what other people like me do.
Yeah. Exactly. Right. It's mimicking patterns. And the pattern happens to be you're the pattern, right? That I may be or may not be similar to. If we do similar things, like mountain biking, uh, or whatever the case might be, um, then because you also looked at running shoes, it would present that to me because it's more likely that a person who mountain bikes and looks at running shoes, uh, will be suitable for another person who mountain bikes. I will also look at running shoes, as opposed to some esoteric thing that's not related, chef's knives or something.
Yeah. So, so I want to then move along. You introduced yourself and you talked about trustworthy AI. Now, you know, trusting such a, well, I guess we cannot call it random anymore, but sometimes it could be a decision-making agent or or or learning agent. It doesn't happen accidentally. It takes deliberate work to thoughtfully expose, you know, those inner workings of an AI system to, to all of the people who use it, because I guess for me to trust the, uh, the suggestion that a system makes to me, it would be so much better if it was transparent. If I, if I knew, well, you know what, Google, as an example, gives me a certain result because of X, Y, and Z. Then if I, if I wanted to make sure that I don't, uh, end up in an echo chamber, I can, I can actually practically take steps to make sure that I'm exposed to all of the possible answers. So how do companies go about making sure that these agents are transparent and, as you said earlier, fair and reliable and so on?
So that's a good question. Holistically, we start decomposing the problem, right? Um, in that data science isn't just the purview of the data scientists and the developers and the data engineers. Uh, you have to get all the stakeholders around the table. So you would have the lines of business people that own the business problem. They should sit at the table. You should have, quote unquote, the lawyers, right? Depending on your industry, uh, you have regulators, compliance officers, lawyers, etc. Some industries are regulated with a very specific, specific set of criteria. Others don't have that criteria, but they set it themselves, and often they actually set it more tightly than a regulator would. Right? It's self-policing is, is sometimes, you know, even tighter. So all these players, when you start adding them to the table, the pattern emerges about how you're going to do this. So the answer lies somewhere between development practices, uh, risk management, and then just line of business, um, I don't want to say education, but but involvement, let's put it that way, right?
Yeah. So you've actually got to, you've actually got to understand the business for which you are trying to create the solution. It's, it's not only, "Oh, here's this fancy computer, let's let it loose on your data and see, you know, what it comes up with."
Yeah. The traditional view, which as little as 2010-ish, right? The thought was, I have a problem, here are some data, like that Caterpillar example, you know, just use them as an example, right? Um, and the only thing people would consider is accuracy. Can I accurately predict something like failure or whatever, right? And that's all that people mattered, worried about in the business was, "How accurate is it? How accurate is it?" Because if you're going to use this, I better know that it works. But, you know, people are starting to realize, ever since 2014 onwards, there's been an exponential increase in academic literature about things like fairness, about transparency, explainability. U, so it's about more than just that, right? You have to, it's the same as hiring a person. You're not just going to hire the smartest person that always gets all these test questions correct, but can't speak, can't relate to people, uh, you know, can't present his findings or her findings in a, um, you know, systematic manner in a report or in a presentation. So, you're looking for a well-rounded individual. So, we're just taking those practices, if you will, from the hiring world to AI and saying, there's more than just doing the job well. You need to understand and have a trade-off between the risks involved. Uh, what are the business factors driving it? What can we push, what can't we push, uh, so forth.
So, we, when we met, uh, I don't recall when exactly it was, 10 odd years ago, maybe, um, we were looking at an optimization problem for a for a South African mining client, um, trying to figure out what the optimal, uh, distances were between, uh, different levels in a mine and so on. And, um, then it was an optimization problem, but, but things have grown a lot since then. And, um, the question that I want to ask you, even though you may not be involved directly in mining, is how do you see the uptake of, uh, of artificial intelligence, as a, as a broad term, uh, in, in real industries today? Is it something that, that is still just an academic concept, or is it, is it real?
Well, I might be biased, no pun intended, but, um, having said that, the reason I am biased is because I do believe it is real, right? From everything I've seen, um, I would emphatically say yes. And there's two, two major driving forces for why I say that. Number one is, we all have heard the, the data growth statistics out there, right? Data doubles every two years. So we are generating more data exhaust than we can possibly do something with with just our own minds staring at the data. Even us running our own structured queries, doing reporting, it doesn't capture the data. It's, and if you thought you have a handle on it, two years from now, it's double that, right? And you're out of control again. So that's the one pillar. But then the, the other pillar is how computing became cheap, how it's interconnected, how we're in the cloud, and how that made the democratization of things like data science, right? Everyone can go train a deep learning model in the cloud using your favorite cloud provider. Uh, and it's the algorithms are open source, and it's easy to get at. It's not in the old days, you had to phone a professor. Now you can just download a GitHub repository and do it yourself in 10 minutes, if you wanted to. So those two forces work together. Um, and the result is that the big companies took notice, and, you know, uh, they've invested billions of dollars each, right? In making the software enterprise-ready, putting processes in place like my department that helps people make this real. So you can't just come in and develop an equation and leave it in a spreadsheet. You have to integrate it with your systems and make sure it delivers, make sure it delivers business value, not just be accurate. So, yeah, I would say we're ready for the mainstream, um, commercially across all industries, right?
So, and this is where interoperability, I guess, comes in. Um, you know, you were talking about enterprise software earlier. It's one thing to know the answer. For example, in a, in a mine, it's one thing to know what the optimal, uh, allocation of resources to a specific schedule is, for example, or a within a specific shift. But it's a different thing to actually know that I'm able to then orchestrate that work, you know. So say, for example, I've got a fantastic schedule for the start of the shift, but reality happens during the shift, and I'm, I'm delayed in a certain task, and now all of a sudden, I must not do XYZ anymore, which was planned to follow ABC, which, which now took longer. And all of that work must be orchestrated digitally, orchestrated in such a way that that the team isn't dispatched to go there anymore, and etc., etc. So interoperability is really the connectedness of these various systems that allows us to act on the fast decision-making that, that your world enables.
Yeah, that's spot on, MBI. So the way we would say it here is AI or machine learning doesn't live in a vacuum, right? It is not the whole application. It supports a larger application or a e ecosystem of applications. So yes, we, we focus on business automation as well, and AI would be part of the whole, but not the entirety of it. It's, it's a decision. So a decision doesn't happen in a vacuum. Where is the decision? Should I service this tire or not? Should I change the schedule to be this way or not? Um, happens in a context, and that context is, is at a higher order level, like your application. And, yeah, that's, that's where the industry knowledge becomes important.
Right. Go ahead.
Absolutely. Yeah. Because, uh, there are also good suggestions and bad suggestions. Uh, even though the, the bad one may be mathematically wonderful, you know, the engineer will tell you, "Nope, this is never going to work." You know, even though the mathematician would be very happy about it.
Yeah, that's a classic example of did not consider all of the context, right? Something seems fantastic in a vacuum, if, or in a subspace, if you want to talk math. Uh, but when you add one more dimension, it's like, "Oh no,", uh, that, you know, the cost curve or something, and it, it's 100 times more expensive to do the right way versus the okay-ish way, or it could be not just cost, it could be any other dimension that the person building the AI did not consider, or it wasn't in the data.
I recently heard someone say, "Change has never been this fast, and yet it'll never be this slow again." So I guess we're, we're in for a very interesting next decade as we look at the maturing of these technologies and the adoption of these technologies in in industry.
Yeah, for sure. I mean, I, I'll, I'll comment on that quickly. Is like, I said, the technology grows exponentially. More compute power, more data, more algorithms, more cloud providers, all that stuff. But change management for, you know, any industry, any company, we didn't change as humans, right? So for us to change our entire way of thinking as a company that has, in many cases, decades of, "We've always done it this way," is hard. And that, that has not changed. So yes, you're definitely right. You know, getting in there and and chaperoning people through the change is a very important function. And then I love the expression that, uh, that, you know, you use the term ecosystem. But in the real world, there's also an ego system, you know, at play. I like that. We're people that, you know, we hardly want other people to tell us what to do, never mind, you know, some clever algorithm that sits somewhere in the in an anonymous cloud, white labeled, to, uh, to tell me what to do.
Uh, listen, Stefan, we're out of time. This was really, really good. I, I'd love to continue, uh, just more and more, but I guess we've got to draw this to a close. And as we do, always, I'm going to ask you also, uh, what's on your bedside table? You know, what are you reading? And what do you propose the rest do?
Yeah. So, I'm currently reading two, two things, right? So, the one is a storybook. It's, um, a series called The Wheel of Time. For those of you who know what that is, it's the, the longest fantasy series ever written, 15 books, most words ever, and it's regarded by many as the best story ever written. So, you know, again, I might be biased. So, I'm, I'm planning on finishing it this time because I didn't quite get through it, and the author was still writing it by the time I got to book nine out of 15. Well, anyway, so, yeah, that, that's the one thing. And then the other thing, more business-like, is this book that I'm looking at. Um, its title is "How to Predict Everything: The Formula Transforming What We Know About Life and the Universe" by William Poundstone. So, it's a really good book. It's, it's an entertaining read. It's not like a, you know, textbook or anything. Um, and it goes, um, about the story of a guy that uses an equation from hundreds of years ago, um, and he's standing next to the Berlin wall, and he, he just completed his studies in this equation, and he tells his friend, "Based on the data, this wall is going to fall in 26 years, give or take, you know, a certain margin of error in months to the day." Is when, you know, in reality, the wall fell.
Over so much so that what he wrote, incredible. Yeah. And it became a Nature article, which is the pinnacle of science journalism, right? Because of what he said. He said, "I just used the super equation and predicted the end of communism," and it was correct. And he then since applied it to other things, like be predicting celebrity breakups, the stock market, you name it. And they all follow what we call the Copernican principle. And this is old math that comes back. I, I have a theme for that, right? Everything we do isn't new. We're just rediscovering what humanity knows. And this book is an excellent read. It takes detours through philosophy and archaeology and all kinds of stuff.
Sounds like one, well, if you give the pun, one for the books. Um, so Stefan, thanks a lot for joining us on the Mine World podcast, and we look forward to, uh, to seeing what you, what you guys achieve and, and hopefully using what you achieve, uh, you know, in, in mining. I know that IBM is already very much involved, of course, in, in heavy industries, you know, doing data-driven analytics and all kinds of very interesting work. Um, so thanks for your time and all the best with, uh, with the interesting projects that you, that you're working on.
Yeah, it was a pleasure and thanks for having me.
[Music]