Subscribe
Share
Share
Embed
Born & Kepler is named after the mathematician and scientists Max Born and Johannes Kepler. This bilingual podcast, offered in both German and English, dives into the expansive world of Artificial Intelligence (AI), exploring its foundations, evolving technology trends, academic search, and its impact on businesses and society.
Born & Kepler will feature a diverse lineup of experts from academia, venture capital, private equity, journalism, entrepreneurship, CTOs, and policymakers. Each guest offers unique insights into how AI is reshaping their sectors and what we might expect in the future.
Our goal is to provide a deep understanding of the core principles and breakthroughs in AI, enabling you to stay updated with the latest advancements in AI technologies and how they are transforming industries. During our episodes, we will explore how AI is influencing business strategies, optimizing operations, and driving innovation. We will also explore the ethical, social, and regulatory aspects of AI in everyday life.
Andreas Deptolla (00:00.27)
We don't eat, right? So we're really good. Good morning, Greg. How are you?
Gregory Wheeler (00:01.805)
OK.
Gregory Wheeler (00:05.41)
I'm doing well, Andreas. How are you?
Andreas Deptolla (00:07.435)
Perfect, very good. Looking forward to our conversation today.
Gregory Wheeler (00:12.27)
Likewise, likewise.
Andreas Deptolla (00:14.187)
So Greg, you're originally from the United States. Tell us about your background from an academic perspective and how you ended up here in Germany.
Gregory Wheeler (00:24.526)
Yeah, so it's kind of a long, strange, winding path. I I have this foot in academia and philosophy and also computer science. And that started from the beginning. And then I've spent part of my career in the United States, academic career in the United States, and then also in Europe. And for the last several years, I've been at the Frankfurt School of Finance and Management, doing computer science mainly.
Andreas Deptolla (00:56.351)
Maybe on a more personal note, how is it different, how is it for an American to live in Germany? What are the differences on a day-to-day life basis?
Gregory Wheeler (00:56.825)
Thank you.
Gregory Wheeler (01:06.552)
Yeah, I I enjoy it very much. But I think I've found that over time, the differences of I don't notice them as much in my day to day life. It's going back and forth or having a head really in both spaces. It seems to have all blended together more or less. Of course, there's differences between the countries. And maybe I'd notice a little more on academic culture and also business culture, these kinds of things that.
that I spend most of my time engaged in thinking about.
Andreas Deptolla (01:43.103)
When you look at the academic world, specifically in Germany versus the United States, where do you see differences, specifically as we're talking about research and teaching of artificial intelligence?
Gregory Wheeler (02:00.238)
Yeah, I think one of the things, and this maybe is a bit broad, but it's an example that instantiates this, that, or it gives an instance of this, but it's the, so in the US, there tends to be more of an emphasis on say ideas and where, and it matters a little less where they come from or who's pushing them. so for instance, the leading labs, there's a track now where,
Andreas Deptolla (02:16.863)
Mm-hmm.
Gregory Wheeler (02:30.03)
Basically, they're hiring very smart undergraduates. people that are 22 years old are having a choice between, say, going to Stanford or Carnegie Mellon or MIT for their PhD. And they just go right into these labs. Some of them do. And are driving research. They're not interns. And I think even a few years ago, anywhere, this would be a preposterous idea and might even still seem a preposterous idea here.
But that's a way that things can be driven forward. Now, on the other hand, there's all kinds of crackpot ideas that come up and wild goose chases. And one has to have a bit of a tolerance for this as well, because it's not every idea is a good one. So that's on the one side. And on the other side, think having a bit more of conservatism.
making sure the fundamentals are in place. I do appreciate that as well in the German context.
Andreas Deptolla (03:36.385)
So the German perception of America is often you can achieve whatever you want, right? You can be in one class and then move to the next class and move up in society. Well, this is now from a wealth and monetary perspective and academia. Is that what you're seeing as well? How would you compare Germany and the United States in terms of upward mobility?
Gregory Wheeler (04:03.854)
Yeah, I guess I'm one of those examples. My father drove a truck. I was the first one to go to college. And then I had no idea, really, about this world. And then I was studying philosophy, of all things, or got interested in this. I had a background in engineering also, kind of this practical hedging thing. But that I was able to pursue this. And I didn't know enough to.
know that this is a very risky endeavor. And there's survival bias. It worked out for me. I think there's also the downside of there is real risk in people who fail. And it's a noisy place as well. I think it's, well, one of the things that happens when you move abroad is the thing you learn probably the most about is where you came from.
because you have the opportunity to compare and you wonder, well, could we have a little less of this turmoil that is of public now, or maybe it always has been, and have a little more of this sort of stability? the longer I go, maybe these are just emergent properties that are deeply set in each of these societies, and it's not something you can pluck sort of the best.
Andreas Deptolla (05:04.074)
Hmm.
Gregory Wheeler (05:30.488)
or easily, I should say. I don't want to give up hope that there's no changes in these societies. Of course there are, but it's a more complicated process. Yes, there is a more upper mobility, I would say, opportunities, or there used to be at least. I think that's also becoming a little strained in the US just because of the expense of going to university.
Andreas Deptolla (05:33.537)
Mm-hmm.
Andreas Deptolla (05:39.595)
That's interesting.
Andreas Deptolla (05:53.035)
Yeah, I think it's expense in the US, but then also the, I think that there was a recent article in the Wall Street Journal about these organizations that prep students for, you know, to get into the Ivy League schools and whatnot. And the expense, you know, I think was anywhere between like $30,000 to $200,000 just for the preparation, right? And that, you know, obviously limits the pool of students that can even think about that.
Gregory Wheeler (06:04.058)
That's right, right, right.
Gregory Wheeler (06:20.868)
Yeah. But then again, things could turn on a dime. And insofar as one gets into an area where it's about performance and something that can be measurable, and it does at least allow more opportunities for, for enterprising young people to find their way, I think in the US. And, and I would like to see that
Andreas Deptolla (06:25.174)
Mm-hmm.
Gregory Wheeler (06:50.276)
just as a kind of a norm everywhere and do my little parts here to try to encourage that with my students and also the people that were hiring.
Andreas Deptolla (07:00.965)
It seems very much like, so that in Germany, it's also a cultural difference, right? If you now talk about like, you know, the top 10, 20 % of your students, what are they doing? they, do they take a more risky track? Are they going into entrepreneurship, right? Or do they do more traditional things like, know, working with McKinsey or Boston Consulting or going into the industry, right? you know,
What do you see there with your classes? Is there a desire these days more for entrepreneurship, or do you see that it's more the traditional choices that your students are making?
Gregory Wheeler (07:43.672)
It's a mix. main things that I'm doing here, I'm kind of the founder and academic director of the data science program here. So we've of got all that up and running, recruited faculty and that kind of thing. And so we've had a few generations of students that we've graduated and it's been a mix. mean, some of them have started their own companies, some have gone off to do a PhD, many I'd say most are.
are entering into consulting, banking, finance, as you would expect. But yeah, it really is a mix. And some are going to work for startups, so not just starting their own, but taking also, this is also a risky thing to do. I think that there's an appreciation of how much you can learn at a small startup. You have to be very, very resourceful and you may get opportunities you just wouldn't get in a large organizations.
anyone else small this part of startup is and it could be you know very very wide range of things that we're exposed to and I think there's an appreciation for that in the younger generation and you know moving between the two worlds.
Andreas Deptolla (08:57.921)
And Greg, you are teaching at the Frankfurt School of Finance and Management, which is a private school, right? Maybe you can talk a little bit about how it's funded. maybe for the, I mean, most Germans, if you go to public university here, the tuition fees are very, very small compared to the United States, right? So how's...
cool funded and then maybe how do you justify so to speak, how do you differentiate to justify these tuition fees?
Gregory Wheeler (09:32.494)
Yeah, so it's principally tuition-driven. We have a couple of other sources of income. There's other units in the house that will contribute to the Frankfurt School's education mission, know, been an international advisory service and so on. But it's principally tuition-driven. I also, this is probably one of the cultural differences that it's totally normal and not strange at all.
I went to some public universities and private universities through my education. so this is a normal concept. I'm also aware that it's not, it's more, little more controversial, let's say, in Germany. And the way I've thought about it what attracted me to the Frankfurt School and this idea is that because we have competition right down the street that is for free,
it forces us to differentiate ourselves, to offer what I try to do anyway, to offer programs, to offer education that is worth it. And that what I would like to see most is that the ideas that we try out that work then are picked up. And that kind of game, if you will, that using competition to move things forward, that
honestly is what attracted me to coming here. I spent some time at a public university, which I enjoyed. It was a great place in Munich. But I also had a guess that this would be a little more fast moving and dynamic. And that's turned out to be true.
Andreas Deptolla (11:25.023)
And what is the range of tuition fees that somebody would pay if they would, let's say, go through the undergrad or getting their master's degree?
Gregory Wheeler (11:34.488)
So the master's program is a two-year program. And I think our tuition for the entire program is about 30,000. I will get this wrong. It's 37,000 thereabouts, between 35,000 and 40,000. Something along those lines, which is I think it's still less than a semester at Stanford. But not that I think that we should be charging.
Andreas Deptolla (11:37.525)
Mm-hmm.
Andreas Deptolla (11:46.977)
I'm sure it changes every year with inflation.
Andreas Deptolla (11:55.733)
Mm-hmm. Totally.
Gregory Wheeler (12:03.288)
these rates for reasons we mentioned earlier, they become onerous, the debt that students will incur. it seems to be, I mean, this is similar to what I did when I was younger. So this all seems very familiar to me, manageable, and insofar as we are, which I hope we do, provide education for students where they can
leverage this to achieve their goals. This is why I got into this business.
Andreas Deptolla (12:42.465)
Yeah, I guess the question that somebody would have to ask themselves is like, okay, I'm paying whatever amount is 35, 36K now for the education or maybe I choose a less expensive option and put this money into a S &P 500 index and see what it does over 30 years. And then I get to compare what did I get from my education.
And maybe to kind of like reframe the question, where do you see the value for the student? it in the content? it in the smaller class sizes? The prestige of the school? Is it the network? What is it specifically?
Gregory Wheeler (13:29.914)
Well, I would say all of those, but maybe I give a concrete example that I can speak to. So if you go to practically any statistics department in the world, they're Bayesians. But still, we teach undergraduate statistics, basically it's the frequentest statistics that students would generally see. And sometimes that's all that they'll see.
You really have to stick around a little longer either in the stats department or machinery and computer science before it kind of comes around. You're like, there's something that's conceptually much easier. It's just super complicated to compute. And there's good reasons why that change has happened 20 years now. So one idea that I had, and there's other people around the world that are
attempting to do this is to say, what would an introductory statistics course look like if you started off from a Bayesian point of view? What does that look like? And then offer that, tweak it, and then introducing frequentist methods, which are very important, but they're driven by computational constraints. That was historically why that was the case. And so that's something that you're kind of reversing the order.
as it were, and there's a lot of inertia and pedagogical inertia throughout academia, still in the United States for that matter. And that's something we can try here. And that's been very, very successful. It also adjusts then the machine learning courses that we're doing. And lo and behold, now that we have generative AI and probabilistic methods are thrust at the forefront, it fits very well with the curriculum that we're
adapting for that next stage. So I expect all of this will realign along those, that would be my bet anyway, and that's just one example in a vector. And if we're doing our job, we, I mean the faculty at the Frankfurt School or other institutions that are trying to compete, that's exactly what we will be doing. And then I really hope that these successes are picked up.
Gregory Wheeler (15:57.242)
I'm not interested in resting sort of on our morals, there's always new things that are.
Gregory Wheeler (16:08.147)
to stay on the edge to push things forward.
So Greg, it sounds like being a private university institution, you also have more flexibility, right? In your curriculum, in how you do things. How does this apply to AI and how you use AI to teach?
Are you incorporating any AI tools into your teaching? And are you students allowed to use AI in their day-to-day studies?
Gregory Wheeler (20:41.582)
Yeah, so I encourage my students to use AI. so I teach a machine learning course, so the learning fundamentals about this. And as I mentioned earlier, perhaps a Bayesian statistics course, where this is all hands on, where they're doing probabilistic programming. So they're steeped into these topics anyway. I think the thing that's changed over the last, say, year or so is
the use of co-pilots for coding. so I'm the academic director for this applied data science program. So I tell them in their first day when I welcome them that each professor is going to have very, different ideas about the permissibility of using this. I'm probably more on the permissive side, and others are no way.
Andreas Deptolla (21:14.187)
Mm-hmm.
Gregory Wheeler (21:36.102)
and that they should, it's a responsibility to really pay attention to what is permissible and impermissible and follow that. Because part of that's going, and I think the mix is actually right. They should be adaptable and be able to perform under these different constraints. Because the thing that's really important to convey to them is that they shouldn't use this tool as a corner cutting.
Andreas Deptolla (22:00.033)
Mm-hmm.
Gregory Wheeler (22:00.172)
And it's very, hard to do that. But those temptations have, you know, students have confronted those for millennia. So this is just the latest generation of tools to be a temptation. But the difference between someone who is using this as a corner cutting tool and someone who spends the time at the woodshed building up their fundamentals.
is what 10x, 100x, and so don't blow the opportunity, don't be dumb. But of course, know, human nature doesn't change and so one of the challenges we often have to do is being able to discriminate between students that know the material and those that are using these and other tools for shortcuts.
Andreas Deptolla (22:33.569)
Mm-hmm.
Andreas Deptolla (22:53.441)
What does this mean in practical terms for you to differentiate whether somebody cut the corner, used an eye tool or not? Is it verbal exams? What do you do in order to really see if somebody understands the concept?
Gregory Wheeler (23:12.098)
Yeah, so we're also on the faculty side figuring this out. And so one thing that we did is this year is going back to paper exams, you know, until.
Andreas Deptolla (23:15.328)
Yeah.
Andreas Deptolla (23:23.361)
Mm-hmm.
Gregory Wheeler (23:25.838)
until we get our hands around this. And another thing also is having opportunities, contact points for students to quiz them. Because I think the thing from a faculty point of view that we need to maybe readjust is that what we're interested in are the latent variables. So do they know the material? Have they mastered a skill? And the main vector that we've used to assess this, or one of the primary ones, is writing.
generating text or generating code for that matter. And not to confuse the method that we've used to become accustomed to for measuring the thing that we're interested in with the thing. And then my courses are in part to build models, to write prose and also to write code.
So it's a little tighter and more intimate to figure out how to disentangle these things. But it's experimental. I tried some things last year that actually I may have overshot the mark a bit. I was expecting them to have a little more of this problem than I did. And then this year, we're making adjustments and seeing where we come out.
Andreas Deptolla (24:27.499)
This
Andreas Deptolla (24:46.587)
So you talked about coding, right? And there are all these predictions now about, should you even still learn how to code, right? Because in five, 10 years, AI will do all the coding for us. What's your take on that? Is it still worthwhile learning how to code to learn the fundamentals? Or do you think this will really shift?
Gregory Wheeler (25:11.982)
Yeah, I think it's still important. In fact, it would be more important that I think this idea got a lot of attention. A little video clip you can find from Jensen Wang, who's the CEO of Nvidia and whom I admire tremendously. I have great respect for him. It's just on this one I disagree. And I think the.
Andreas Deptolla (25:23.709)
Mm-hmm. Mm-hmm.
Gregory Wheeler (25:36.916)
way I try to explain it to my students is to say, look, the difference between not knowing anything and using these tools to build up, let's say, some software, and a pretty impressive software stack, actually, maybe that's 10x, and so you'll crush people that are not using this tool now. Well done.
But the difference between someone who knows some computer science and knows, can even just work through the abstraction of a stack. I don't know what that would be. 100x, I mean, it's, and those people will really kill it. It's a broad prediction, but I'm already seeing it in my work. I, you know, whereas, so I also have a company and so I will, you know, one of the workflows is that I would sort of sketch things out in a notebook.
And there'd be pieces of this that I would write in prose. Like, OK, this thing, here's how this algorithm I think will work.
And then some text like, okay, then we have to do, which usually is very difficult, do some things, scaling it or looking at more than n equals one, and then a little bit of code. then, know, it's kind of as a first pass to present ideas to my engineering team. And what I've found over the years, I'm doing more filling in more of those details, because as you know, those details will really matter in whether the idea works.
and getting things closer as a first pass to something you push into production, it's clear. The iteration cycle is tighter. Because I'm not a day-to-day software engineer, right? I I have other things going on. So I think that's an anecdotal example.
Gregory Wheeler (27:34.36)
tend to think from the near future, people that are really interested in computer science should not be dissuaded from studying it. think there'll be plenty of advantages.
Andreas Deptolla (27:43.969)
And you mentioned as one of the tools that the students are using co-pilots, right? GitHub has a co-pilot, Microsoft does, right? I think I read a statistic that about like 25, 30 % of the programmers now use these co-pilots and mostly for certain, has like new documentation, maybe bug fixing, right? Documenting the code, these kinds of things.
Gregory Wheeler (27:49.196)
Yeah.
Andreas Deptolla (28:13.249)
it maybe gains 25%, 30 % of the speed. Is that what you're seeing? And how do you think will that develop over the next, say, three to five years?
Gregory Wheeler (28:25.422)
Yeah, I mean, I think that's sort of the nice use case for seeing the potential improvements in productivity. And I've seen it in my own experience, my personal experience, and also in my team. I would say the increase in productivity is more than 30%. And this technology is improving.
Andreas Deptolla (28:36.5)
Mm-hmm.
Gregory Wheeler (28:56.47)
weekly, you know, maybe even daily. mean, it's, it's, it's the pace is, is, I haven't seen anything like this. So I think this will continue and it's exciting.
Andreas Deptolla (29:13.153)
You mentioned, I want to take us on a slightly different, you mentioned the CEO of NVIDIA a couple of sentences ago, right, with a statement that I think he said something along the lines of like, you know, I don't recommend my kids to learn coding anymore, right, to make a statement. But NVIDIA is interesting, right? As a ship manufacturer, like, you know, it's one of the most valuable companies now on the planet, right?
Gregory Wheeler (29:29.474)
Right.
Andreas Deptolla (29:42.049)
How do you perceive the valuation? that justified? From a finance perspective, what's your view on that?
Gregory Wheeler (29:46.264)
You
Gregory Wheeler (29:50.286)
Yeah, I'd be right, right, right. I'd be hesitant to give stock advice. But I would say here's the bull case for NVIDIA. It's not a chip company, any more than Apple is a chip company. You think of Apple as like a platform ecosystem company. That's what NVIDIA is. But instead of B2C, your little iPhone in store and stuff, it's B2B.
Andreas Deptolla (29:56.383)
Hahaha
Gregory Wheeler (30:18.958)
And so NVIDIA is like IBM in 1964, right? So this is the year when they introduced this 360 family of mainframes. And what NVIDIA is building now is a platform for parallel computation, which is the main engine for AI. mean, parallelization is...
even bigger than AI, which that's maybe hard to imagine. So that's the bull case. And the bear case is also, well, it's like IBM. So these massive opportunities, and so the rest of the field is going to come for them. And AI is the main driving engine here. So.
And the just the kind of capex we're talking about is it, you know, it introduces dynamics and uncertainty that, yeah, there's tremendous upside, but there's just a lot of risk. if it, you know, I, there's a case I less, would say of hedging.
But if you look at the history of Nvidia as a company, they've gone close to dying two or three times. So it's very, very resilient organization that this is in its first rodeo. And so they've built a pretty resilient organization that can weather the storm that's coming.
Andreas Deptolla (32:01.835)
What's interesting, you mentioned the CapEx, the billions of dollars, Microsoft, Meta, all these companies are spending and Nvidia obviously is profiting from that. And it seems like a lot of the value creation right now on the pure profit, the EBITDA side is on Nvidia's side. How do you see that developing in the...
Gregory Wheeler (32:09.016)
Yeah.
Andreas Deptolla (32:31.443)
you know, as the cycle continues, right? Like where will the money be made, so to speak? Is it on the infrastructure side, on the language model side or the application layer or somewhere else?
Gregory Wheeler (32:45.59)
Yeah, that's the, know, billion used to be a lot of money. But that's the big question. So let's take the language model, the frontier model. So this is the Anthropics, the OpenAI.
Andreas Deptolla (32:50.125)
Ha ha.
Gregory Wheeler (33:05.614)
of the world. so the bet is that, their bet is that it's a race and that, you know, it kind of like Google, right? So all these search companies and, you know, only one survived really and out of the nineties and fantastic business model, right?
And they're on that race. you know, like infinite profits for the winner. Maybe some leftovers for them. So that's the belief of the investors that are, and they just close this record around OpenAI did.
The other story is that, and this is the bet that Meta is making with their llama models, they're the lead of this, but it's the open source community, is that, well, this will be a commodity. so basically this will be more like the PC business where...
there aren't margins and whatever advances that are made, I these could be spectacular benefits from consumers and developers building on this. But at this model level, is it going to be a single?
capture of value or is it going to be a commodity that basically the cost of that is driven to Well, it gets driven to zero or gets driven to the energy cost of running these things so I mean that is that is a stark of different kind of bets as you can imagine and it's Really uncertain how that's going to play out but these things will be built and I think another area of value creation was
Gregory Wheeler (34:57.816)
building businesses on top of these services. So business verticals on top of LLMs. so it can depend upon where you are looking at this thing. If you're in the large language model race and game, could be stressful if you're thinking from the financials point of view. But from a consumer point of view, it may not matter so much. And you may even prefer the commoditization.
Andreas Deptolla (35:29.653)
Where do you see, so it seems like whether you're now talking about the ships or the infrastructure layer, if you talk about large language, you just need massive amount of funding to really make an impact. And that's why we see a lot of these companies in the US, you have different structures, venture capital and whatnot.
Where do you see the opportunities for Europe, for startups, right? We talked about your students now, right? What kind of maybe niches or opportunities do you see in the world of AI?
Gregory Wheeler (36:14.274)
So I would turn the question around. I think they're everywhere. And then the question, I guess there's different questions, so opportunities for students or opportunities for investors or different stakeholders that are looking at these questions. But.
But one thing to look for that maybe intersects with students as well as maybe investors is you're going to need, and I think to Jensen, going back to the Jensen mind quote, he continues further. I think the point he was trying to make is that this, insofar as it exists, this idea of sort of a siloed software engineer, that is probably going to go by the wayside, or at least diminish in...
the opportunities for that profile are going to diminish because this is enabling more more people who have domain expertise to build software or to build companies really. And my reply would be you can 10, 100X up from that if you combine deep domain knowledge with computer science.
And that would be kind of the recipe if you're looking at a company. So maybe I'll give you a concrete example of this. So I think there was in the news a couple years ago, because I think it involved, well, I don't remember who it involved, but there was a notable person who was in the courts in the US and his attorney said through just a...
Andreas Deptolla (37:30.059)
Mm-hmm.
Gregory Wheeler (37:47.254)
Zero-shot prompt for cases and it spat back cases that didn't exist. So it hallucinated horribly and this attorney was reprimanded by the judge. Of course, made the press. There is a, and so then the conclusion is these things are terrible. You can't trust them, blah, blah, blah, blah.
There is a company, it's called Case Text, and they've been around for a while in San Francisco. the interesting thing is their model is basically built on making an agent take attorney.
Andreas Deptolla (38:08.193)
Mm-hmm.
Gregory Wheeler (38:21.728)
that's an associate that helps attorneys and law firms. And that difference between sort of the experience that you saw was very unreliable and was kind of dumb to actually a usable software service that's giving reliable results. There is a lot of software engineering and know how to stitch that thing together. And that's where the value is created. So if you're looking for opportunities, be prepared to find people or a problem that you understand.
Andreas Deptolla (38:39.873)
Mm-hmm. Mm-hmm.
Gregory Wheeler (38:51.662)
and you're committed to doing all of the work, which is a lot, that's where all the value creation is, of building vertical. And I think that's a great example. And there's that sort of a playbook or template that is not bad to study in looking at how to apply AI across a wide spectrum.
Andreas Deptolla (39:15.657)
It seems like in that example, you still need the human being to look at the results. It probably gave that lawyer now the 10x advantage that you mentioned. Now he didn't have to read all the cases, but you still have to double check things, maybe draw certain conclusions and whatnot.
Gregory Wheeler (39:19.37)
Exactly.
Gregory Wheeler (39:38.306)
That's right. I mean, in this, I don't know this product so well, because I'm not an attorney, it, it's basically what they've engineered is to do what
a top level associate would do. So all the things that come back are documented. go through all these boxes and boxes, is there any fraud that was committed in this company? And then there's a report and then it points to particular documents so that the checking would be not starting from scratch, but it'd be similar to checking an associate's work in a law firm. So it's very much understanding the workflow and then
learning how to program. This is a new kind of computer to program and we're using natural language to code with it and interact with it and building that software up so that the user is not dealing with those details but is dealing with it like a sort of a hybrid of a software tool but also you talking to it as if it's an associate.
And testing, testing, testing, testing, testing. I mean, it's, yeah, so that's really, really important.
Andreas Deptolla (40:52.985)
So we talked about, how can Germany and Europe maybe as a whole benefit from AI, right? From a value creation, from a startup perspective. And you bring up an interesting point, right? Where are we good at right now, right? In the law of times, it's engineering, right? It's kind of like these hidden champions, small organizations, right? That are world class and whatever they're doing. Maybe the question is how can...
you know, are we in those verticals, so to speak, look at like, you know, what role can AI play and how can AI advance those companies even further and keep the advancement, you know, compared to other players.
Gregory Wheeler (41:36.076)
Yeah, so I think Germany is going through a period now that's not unfamiliar. mean, this happened in the 90s, and this has happened in different countries where the ingredients that were making the economy, the country, such an economic power.
you know, those things have a life cycle. And there's a, I think a need that's broadly recognized that there needs to be some updating and some changes. Great. I mean, there's opportunity there. So I think there's plenty of things to chase. There might be some things that we could focus on to improve the odds of
the next generation of companies being born and seizing a commanding position in the global marketplace. So some of those things are, let's say venture capital, just to take that as an example. Because companies can start themselves on the cheap, it's never been easier, but probably it's going to be...
Andreas Deptolla (42:53.665)
Mm-hmm.
Gregory Wheeler (42:56.974)
paradoxically that much harder to stand up a business that's a going concern. So if you think of venture capital, I think the US economy is something like six times Germany thereabouts, but the investment in venture capital is like 24X.
Andreas Deptolla (43:22.656)
Mm-hmm.
Gregory Wheeler (43:24.034)
thereabouts. So I think it was 6 billion in Germany in 2023. And so that's something that is going to need to be closed or increased. That gap is going to need to be closed to be competitive. There are also, that's also functioning IPO market. So that will influence
the risk taking of the entrepreneur. If they have a really good idea, there's a really strong incentive to not do it here, but to do it in, say, the US.
Gregory Wheeler (44:08.546)
There's a function, better functioning IPO market and, and more funding. And we talked earlier in the conversation just about, being comfortable with risk and also just being comfortable with failure. mean, it's, it, it's really painful. It really hurts and, but it shouldn't over and above that to be some, some kind of black mark on.
Andreas Deptolla (44:22.753)
Mm-hmm.
Gregory Wheeler (44:36.806)
it is part of the game. I mean, it's very, very rare for, and there's huge survival selection bias if you hear these stories about coming right out of the gates and your first, everything's just a boulevard of green lights. That's just not how this works at all. But even through these failures, you learn a lot. so this...
I think the younger generation has this kind of appetite, but I think these other roadblocks that they will run into are a little more serious. I mean, they will kill companies that may otherwise.
Andreas Deptolla (45:19.205)
I think a lot of what you mentioned as being risky might be a perceived risk in society. I started a company, I fail and that's it. Certainly the right employer will see this as, this is somebody that has drive, that learned a lot, that's willing to do things maybe differently. But I think that's exactly what's needed.
in society in order to grow industries. You said something interesting about Germany. You compared it to the 90s, right? Where Germany wasn't in a great economic place, right? You mentioned Germany had to reinvent itself, right? Do feel like we had a similar position right now? And if so, what are...
What are maybe your top three recommendation to the government, right? To say like, those are the things that like we really, you can't do it all, right? I think if we now focus, I mean, it's the same as if you're a CEO of company, if you have a hundred different initiatives, you're not doing anything well, but like, what are maybe two or three things we would say, okay, those are the things we gotta get right to turn this around.
Gregory Wheeler (46:36.342)
Yeah, I guess I would think in, you know, about, you know, venture capital incentives to it. So it's about incentives for these main choke points that will allow young people, because I think there is awareness and there is appetite I see with young people wanting to wanting to start out and try to build something.
and build things, or not so young people, enterprising individuals. So I think focusing on, could be tax law, could be just the ease of starting, ease and expense of starting a business. It's...
I mean, it's common to complain about bureaucracy and there is this, we were talking about the US and Germany, there's sort of difference between, you know, so yes, I don't know, I go to the store and.
Germany, I want to buy a rope and they say, wait, you sure you want to buy this rope? You could hurt yourself. You can hang yourself with this. No, no, I really want to buy it. Could you, you know, we need to show you how to avoid hurting yourself with the rope. You know, there's, and it's good intentions, but there's a kind of...
sense of, are you sure you want to do this? Are you sure you know the responsibilities and putting up a little bit of friction to doing that? Where in the US you go out and go, yeah, here's the rope. And by the way, I'll show you how to tie on news if you want, but and you're on your own. But so I think there's some of this is cultural and and I don't know how to and there's many
Andreas Deptolla (48:11.136)
Yeah
Gregory Wheeler (48:24.566)
reasons, benefits for that, at least to stability, it leads to cutting out some of these harebrained ideas that can be destructive for, you know, a real risk to society is not a dumb reason for why these rules are in place. But when the things are breaking down, this needs to, there needs to be a loosening up to find that next, that unlocks a lot of talent here. It's just to reconfigure that and oriented towards
the next growth opportunities and it's unclear a priori exactly what that will be. And so in so far as we can encourage conditions for experimentation, it's an empirical question. It's not something that's gonna be very, very clear and planning in Berlin or Brussels. We'll put the conditions in place where we can experiment.
Andreas Deptolla (49:11.829)
Mm-hmm.
Gregory Wheeler (49:16.03)
and recognize that experiments end mostly in failure. That's just the nature of empirical things. They will mostly fail. And to make that less punishing for folks. And that's a generic advice other than maybe focusing a little bit on the...
venture capital pieces of this and all the incentives around that of why that's entrenched the way it is and try to loosen that up a bit because that's going to be an important ingredient.
Andreas Deptolla (49:50.753)
Greg, you yourself were involved or are involved in a startup, right? Maybe it's a scale up now, maybe you can, you know, X alone. Maybe you can tell us about, you know, the mission of the company, your involvement, what motivated you to start it.
Gregory Wheeler (50:10.55)
Yeah, there's a worldwide there's a five trillion gap between lending gap.
There's folks that qualify for loans but they can't fulfill them. In rich economies such as ours, there's capital that's needing to be deployed and in developing economies, these underfilled loans. One of the things that's been developed to fill that has been the digital lending market. A bottleneck there is capitalization and there's connections to institutional capital.
And the reasons for that are that it's a very fragmented market. It can be opaque. And basically, just technically, it's very difficult to connect to these different loans. So what Exloan does is we run and we've created a marketplace for digital loans. So we're connecting institutional capital and making these digital loans an investable product or investable. So it connects.
Capital to the SMEs, consumers, and other lenders worldwide. And I am one of the four co-founders.
Andreas Deptolla (51:29.249)
Is this mostly like a B2B play or B2C? Who are B2B?
Gregory Wheeler (51:36.362)
Yes, B2B. So we're connecting lending platforms so they have loan origination that they've approved and they just need funding. And then we're connecting institutional investor, institutional capital to those individual loans. We provide a third party screening and also credit scoring and standardize this across the market.
Andreas Deptolla (51:44.875)
Mm-hmm.
Gregory Wheeler (52:03.774)
So it's like Amazon for loans and the marketplace part and then also Bloomberg for analytics and Shufa for credit scoring, put all these things together. And I'm one of the four co-founders and I do the AI and machine learning in the company, which is at the heart of the company.
Andreas Deptolla (52:05.387)
Mm-hmm.
Andreas Deptolla (52:25.249)
And how, give us a sense maybe of the size of the company in terms of volume of loans that are processed on a yearly basis or employees.
Gregory Wheeler (52:39.438)
So we're about 10 employees and so we're running very lean and we're up in all, let's see, we're up in Europe, certainly US.
Southeast Asia, parts of Africa. We have about 42 platforms. let's see, the volume is, I'll forget now it's, of the volume of loans that we're processing through our market is, I wanna say 60 billion, up to 60 billion the first quarter of this year.
So it's a growing market. But we're also creating a new investible, it's an alternative investment and it's a new investment class, if you like. It's opening up this digital lending market as to institutional investors. so there's...
it's back to the risk appetite. So there's different appetites for depending on where the institutional investors are located, whether they're maybe here locally or in Singapore or United States.
Andreas Deptolla (53:58.681)
And what is the business model behind it? How do you make money? it on the spread of the loan or fees?
Gregory Wheeler (54:06.284)
So it's fees. So it's the fees to the institutional investor that, so the assets under strategy, we say, and so it's the flow that they dedicate a certain amount of capital, our software, they can set their risk profile or what they want to invest in basically, and screen through the flow of loans that are coming through the marketplace. And then those orders are placed.
we take a fee from that flow.
Andreas Deptolla (54:39.421)
Is there an ultimate goal with the company? mean, are you running this to IPO one day or is this meant to be a cash cow for the founders? Like, where do you want to take it?
Gregory Wheeler (54:55.776)
I think we're really liking the business and we're still growing the business. so I think we're open to...
Gregory Wheeler (55:12.178)
We're open to all future positive opportunities, but we're not in really thinking about selling, I would say. think it's not a bad business. And it's also interesting, I find it intellectually interesting to work on the, to develop the technology and actually building the company.
So I'm not in any hurry for my one part of it. We have three other partners and 30 some investors now. So it's a going concern and those decisions will be taken when and if they come. But my little part, I really enjoy building and making the software better.
Andreas Deptolla (56:09.057)
And I'm sure it's interesting for you to combine, so to speak, the startup or scale-up world with academics, right? And bring some of these learnings and examples to the classroom.
Gregory Wheeler (56:24.93)
Yeah, that's right. It's a mix. It's now also blending where it's, you know, I'm coming from, so I've been in this game a while and, you know, when I was in graduate school in the 90s, there was logical methods in AI were dominant. And there's actually an old idea, goes back to Leibniz and George Bull from Boolean logic and the book was The Laws of Thought.
And in AI, this was really the dominant paradigm since the 50s. so, you know, to give an example this, we had two courses. One was in the computer science department, one was on natural language processing, which was using statistical methods and listeners may be familiar with that name.
But there's another course too, was a natural language understanding. And this was a logic-based approaches. And the questions that were addressed there were, you we just don't want to have sort of pattern matching for doing language translation or text to speech. We want to understand the meanings of things and how children understand stories. And these are deep questions. And it was a lot like philosophy, right? So.
But it turns out that the natural language processing, that was the route to go. so seeing this over this 20 plus years, because we're going on 30, has been a real privilege and be having the opportunity to really adjust and adapt to this, what actually works.
I think the combination of industry and being an academe has allowed that feedback loop to be a little tighter than it may have otherwise been.
Andreas Deptolla (58:22.721)
You said what actually works if you look at AI and Exelon. Are there certain AI methodologies, tools, frameworks that you are applying right now to the service, maybe to better match the lender? Or how is AI used on a day-to-day basis?
Gregory Wheeler (58:47.51)
Yeah, so I mean, there's some proprietary things that I wouldn't discuss so much, we always have a piece of our budget time to be looking at new technologies. And this is certainly something that we're doing. I mentioned earlier about the use of coding copilots or coding assistants.
And these have been tremendously helpful and useful. mean, there's been a big value unlock there that may sound more mundane, but that's been something this year that we've really doubled down on. There's also different data sources. So there's a...
You know, it's a kind of a running joke and we have all these scanned PDF documents and things and people are talking about, you know, AGI, but, can you scan a financial statement and read a financial statement, right? And so, I mean, these are also problems that we confront and, you know, having a go at unpicking value from that problem for our specific need.
Andreas Deptolla (59:38.571)
Mm-hmm.
Gregory Wheeler (01:00:02.618)
and re-evaluating tools that are comparing what we can build internally to benchmarks that are in the field and what the price comparison is, this kind of thing. But this is augmenting a more traditional supervised learning stacked ensemble methods that we're using that's still at the core of our various...
Gregory Wheeler (01:00:28.514)
risk matters for predicting the fall.
Andreas Deptolla (01:00:30.015)
And if you talk about like those practical, I think that the co-pilot is pretty, of people use it, know what it is, right? You mentioned scanning massive amount of PDFs that have been scanned with financial reports. How helpful, how good is AI right now in doing those things? Are we at the point where you could say like, hey, you provide AI with a...
a big cube of unstructured data PDFs, it provides meaningful results or do we need more time to really make it usable?
Gregory Wheeler (01:01:08.27)
So it's getting better, but you need to think about how it's used. So if you or I are doing this to Claude or OpenAI or...
Gemini or whatever you're trying out, you may notice over the months it's getting better, but also notice what you're doing, right? You're sort of clicking a button, you're uploading a PDF, you're sitting there and you're fiddling with it. That doesn't scale. So it's building the pipeline around this that's where the real work comes. And so...
Andreas Deptolla (01:01:30.325)
Hmm.
Gregory Wheeler (01:01:48.514)
The trick is to build different technologies where you're anticipating the LLM, the generative AI part, to get better. And so the question to ask is, is that going to be a flywheel? Does this thing gets better and then my process is going to get better? Or is that getting better going to crush what I'm building or worse, crush my business, right?
Andreas Deptolla (01:02:11.371)
Mm-hmm.
Gregory Wheeler (01:02:14.666)
And that's really the question that we will wrestle with, the engineering team and I.
So for some questions, the answer seems to be yes, that we would get crushed, so we shouldn't do this. And others, it's finding that right pipeline that will work for us and unlock value, and that we try to anticipate how it could get better as the models are getting better.
Andreas Deptolla (01:02:46.879)
Now, if you take a step back and look at the entire ecosystem from a bird's eye view, we had the dot com bubble where there was a lot of hype. then obviously, stock market crashed. And it took a good amount maybe for technology to catch up. And obviously, now a lot of like the
some clear winners, right? And the world has changed. How would you compare the dot com bubble, so to speak, cycle with AI? Where are we now in the innovation growth cycle?
Gregory Wheeler (01:03:29.792)
Yeah, I mean, I've used the dot com explosion as a
earlier as an example because I lived through it like that started and I was in graduate school and I remember distinctly in late 90s and I remember some people leaving the graduate program to join Amazon, you know, which was, you know, it only a hundred some employees. was like, what is it? I remember thinking at the time, nobody's going to buy books online. I mean, it's a little bookstore. gotta look through this to make any sense. So, so I was on, you know, maybe
Andreas Deptolla (01:03:59.542)
Mm-hmm.
Gregory Wheeler (01:04:14.489)
Not at the lead, I guess, of thinking that this was going to be as revolutionary as as momentous as it's turned out to be and that the bubble was a blip. mean, there's most of the things that were started then, they just too early. would have survived later when we had our phones and all the infrastructure, cloud infrastructure to run these.
these businesses online. That's not to say that there was, it was a huge losses and some people really got hurt and burned and wiped out during the 99, 2000 but into 2000 but
I think one of the differences, so yeah, we talked earlier about OpenAI and it could be that they're Google or it could be that they're what, ask Mr. Jeeves. It's possible and it's possible that they're no Googles, right? It could be that this is a commodity.
Andreas Deptolla (01:05:08.235)
Mm-hmm.
Gregory Wheeler (01:05:15.82)
So there are big stakes, there's bound to be winners and losers, and it's a very difficult environment, I think, to pick. So there'll be plenty of things for critics, do-mers, enthusiasts, it's just gonna be a big enough thing going on where everyone can have something to talk about. One difference, though, I think that also I think contributes to the uncertainty of this is that
some of the the internet as momentous as it was, it was stitching together, it was a network, was stitching together a network of computers and building all the infrastructure and the computers were things we understood and we understood also the scaling that was happening and you know you could once you sort of sold all this these pieces it fell into place and you're paying attention which obviously it wasn't in the the 90s.
You think, I don't know how long it is, but this will sort itself out. You can see where this is going to go. With AI, we're not stitching together existing computers. You think of it as a new computing platform. So it's not these von Neumann machines. It's a non-deterministic, probabilistic kernel.
that does things that are very, that are like, you can solve math problems, but it struggles counting R's in the word strawberry, you know, like so. Of course you can prompt it to call a script to run on a traditional computer to count that and return the answer, but. So I think the, the.
The thing that makes it difficult to see where we are in the cycle is that I think we're towards the beginning. But it's a difficult thing to predict of where this is going to saturate. How much more can we get out of scaling? And then with this strawberry model, this reasoning model that OpenAI, that's another vector.
Andreas Deptolla (01:07:18.39)
Mm-hmm.
Gregory Wheeler (01:07:34.317)
for scaling, for improving. So the picture is that you're having sort of a stack of technologies. Each of them have their own, if you like, Moore's law. It's actually quicker than Moore's law was, and the mechanics are completely different. But the kind of scaling that you're seeing, and these things are stacked on one another. And so, yeah, and they'll do silly things, and you can, there'll be plenty of things you can find cases where it fails spectacularly at, but.
It's adjacent to things that are pretty amazing, these capabilities, and that's coming at a speed. So the floor is raising weekly. So if we were to stop things right now, so suppose OpenAI says, that was it. GPT-4, that's all we have. And O1, the reason why, that's all we have. That's all we have for you guys. That's it. We've hit the limit. You still have a lot of economic unlock building on that platform.
Andreas Deptolla (01:08:30.891)
Mm-hmm.
Gregory Wheeler (01:08:33.122)
Like five, 10 years, I don't know. mean, there'd be plenty to do. And the thing it's kind of, it's dizzying is that there's that, and then the floor of the capabilities is going up and it's going up at a rate, even by computers, computational standards, faster than Moore's law or faster than this growth pattern we've seen for a century of how much compute you can do for a unit of dollars.
say. So I think people get enthusiastic about this because there's, well, different reasons, but I can just maybe speak for myself. I am enthusiastic about this because I'm thinking about these drivers and I haven't seen anything like this and it seems reasonable to me and I'm using the technology but where it's going to go
risks that we need to confront the disruption that's going to come. Yes, that's all needs to be contended with, but it's coming.
Andreas Deptolla (01:09:42.027)
So change is coming, right? What do you recommend the business leaders to stay up to date with the change, right? Are there certain books that you recommend? Are there conferences, maybe podcasts, certain blogs that like, know, thinking like, those are the things that you would recommend to...
read, consume in order to stay up. Because there is so much information about this. I think it's also important to get the relevant information in a succinct way.
Gregory Wheeler (01:10:16.568)
Yeah.
Gregory Wheeler (01:10:24.59)
Yeah, and there's a lot. there's, I think my recommendations are skewing a little bit more technical. And my sense is that there's more of an appetite for that than I think we sometimes will believe. So I'm offering this, also mindful of, so things that are accessible from a public point of view, but that are a little more technical. One of them is a very short,
Andreas Deptolla (01:10:31.606)
Mm-hmm.
Gregory Wheeler (01:10:54.51)
paper. It's really a page and a half and it's by Rich Sutton and the title is The Bitter Lesson. And so what basically that spells out is that the main drivers of advanced and AI over 70 years haven't been clever algorithms. They haven't been the kind of
logical approach or trying to discern rules for reasoning, like thinking of AI like physics. And then what we need to do is find out what the corresponding laws of motion are, what are the laws of thought, and then putting those structures in place and then allowing the computers then to, once we find those out, that we can engineer an AI.
I'm simplifying, but that was sort of the dominant idea. And you say all that, no, it's basically cheaper and cheaper compute.
It's just this massive scaling of compute and allowing that, how to leverage that. That's what's driven the success. And it's a bitter lesson because it really kills the romance, in a sense, from an intellectual point of view. And it does this really elegantly in, like I said, a page and a half. And that paper is like a Rorschach test still. So at some point,
you know, you can, if you're around long enough, can sort of see, I remember myself, I was on one side of this and looking at this, like, nah, this is, this is something wrong. You can pick holes in this. And then at some point you, you look at this like, yes, yes, this, this is it. And I mentioned that paper because some of the debates that are in the public discourse, that some more public intellectuals, know, Gary Marcus is, is, is a, is a very prominent critic for instance.
Gregory Wheeler (01:12:48.654)
Some of the intellectual underpinnings really are down to this sort of debate, which is, you could label it as a nativist. So you really need to impose some kind of structure on this. Otherwise, you're not really getting learning, or what it's doing is not really intelligence, it's something else. And then it gets a little fuzzy about, and really good on the criticism. So, know, it doesn't do this, it messes this up, this is a big risk, you're gonna mess it up, blah, blah, blah. And then what that structure is, or how it would work, or.
like it's a little fuzzy. And so I think that's something to maybe keep in mind and it may help at least in digesting some of these intellectual debates. So that would be Gary Sutton's bit on the bitter lesson. That's available online.
And then another thing is a website. So it's a YouTube video. And there's a series of visualizations of mathematics that Grant Sanderson does. It's called Three Blue, One Brown. They're outstanding. But he has a video on how large language models work. And that's just done through visualization.
Andreas Deptolla (01:13:54.911)
Hmm.
Gregory Wheeler (01:13:57.166)
And so that's this introduction, these transformer architectures. That's what's underpinning on the GPT, the large language models that everyone is building on now and just how that works. And I think that's like a 20 minute video or something like that. And it's aimed at visualizing higher level mathematics, but I think it's digestible to a broader audience and would go, you know, and also you can go down a rabbit hole. mean,
his videos I think are outstanding if you're interested in mathematics. So those are two things I would
Andreas Deptolla (01:14:31.723)
Well, I also appreciate 1 and 1.5 pages, 20 minute video. I think that's very consumable for the audience. So yeah, thank you so much for these recommendations. Thanks, Greg, for your time today, for all your viewpoints. Perfect. Thank you so much, and enjoy the rest of your day.
Gregory Wheeler (01:14:50.606)
It's Andreas.
Gregory Wheeler (01:14:56.312)
Thank you, bye.