Ep. #996 - The Limits of AI and ML

Matt DeCoursey 00:01
And we’re back! Back for another episode of Startup Hustle. Matt DeCoursey here to have another conversation I’m hoping helps your business grow. AI and ML. Artificial intelligence and machine learning. What’s the limit of it? You know, a lot of people are talking about how it’s going to either ruin the world or make the world a greater place. I don’t know if either one is true. We’re going to get into that today, in-depth, and before we do that, today’s episode of Startup Hustle is sponsored by Equip-Bid Auctions—your Midwest online auction marketplace to buy and sell stuff. Equip-Bid provides dedicated support to affiliates in Kansas, Missouri, Nebraska, and Iowa. Join the team and sell everything from heavy machinery to home goods, vehicles and boats to restaurant and kitchen equipment, and tractors to patio furniture. Go to equip-bid.me/startup. Man, that’s a lot to remember. So why don’t you just scroll on down to the show notes and click the link there? Once again, there’s a link for Equip-Bid. With me today, I’ve got David Magerman. And he is the managing partner and CTO at Differential Ventures. That’s a venture capital fund management operation in New York, New York. You can also, in the show notes, find a link for Differential VC. David, welcome to Startup Hustle.

David Magerman 01:25
Thanks for having me. I’m happy to be here.

Matt DeCoursey 01:27
You know, I’d like to start my conversations, and, you know, we’ve got a lot to unpack here. AI/ML all in one episode and what the limits are. But, you know, let’s get a little bit about your backstory. And I’d also like to hear a little bit about what you guys are doing.

David Magerman 01:44
Sure. So I actually started out as a computer scientist. I got a Ph.D. from Stanford back in the early 90s. And my focus was on AI and machine learning, data-driven approaches to solving the natural language parsing problem, part of the speech recognition problem. And I was very much focused on academia and looking to become a professor. But, you know, faculty jobs were hard to get back then. I didn’t get my first choice and ended up going into this new thing called quantitative hedge fund, which ended up using some of the same engineering and software development skills as speech recognition did. And I ended up spending 20 years doing quantitative trading. Building software systems and particularly using data science and AI, and machine learning to solve a real-world problem to predict price movements in financial markets. And also to come up with optimal portfolios for deploying those models and making money off of them. And then, when I left Renaissance, Renaissance technologies came to work. When I left there in 2017, I looked at the landscape of where the industry was using AI. And I found that the AI community was doing what it’s done since the history of the industry, which is taking technology that mimics human intelligence and promising the singularity that machines we’re gonna become sentient. And like humans, solve all the world’s problems without people involved. And I thought it was a great time for me to get involved in trying to influence founders, startups, and early-stage companies that were trying to use AI. I’ve always felt that there’s a ceiling to performance and the power of AI. And if you use the technology within the scope of what it’s good at, you can get a lot of value out of it. But if you presume it’s going to be more powerful than it actually is, you’re inevitably going to lead your company and your customers, and your employees down a bad path. So my partner, Nick Adams, and I found Differential Ventures to try to help founders who were starting up companies in this kind of applied AI data science machine learning space. And try to guide them to come up with use cases and applications that are within the scope of the power of the tools so they can be successful with their businesses.

Matt DeCoursey 04:05
So once again, there’s a link for Differential VC in the show notes. So let’s get let’s turn the Wayback Machine on here and go back to, you know, you’re talking about, you know, working at Stanford, and you know, this is that’s 25 plus years ago at this point, not to put an age on ourselves, but I was I am equally experienced in life. Let’s put it that way. So, you know, I feel like the outlook on AI and ML at that point, you know, like, and well, there are still people that are freaked out about what it can do, but you know, when you look back at like 2530 years ago, this is kind of the birth of a lot of this stuff. I want to say it’s the very it’s wild west or the genesis of a lot of this. How has the overall outlook of the capability of AI and machine learning evolved or changed over that, you know, a quarter of a century it’s been I used to work Genesis, given that it was the ogre machine in one of the latest Terminator movies.

David Magerman 04:59
Yeah. But, you know, like AI has had this history of mimicking human intelligence. As I said, back in the late 60s and early 70s, there was a series of programs, one of them was called Eliza, which was a psychotherapist application, like a chatbot. And it gave the impression that it was, you know, interacting as a human being. And, of course, it was just a bunch of little hacks and scripts and pattern matching that had no intelligence whatsoever. And it led to the development of what they call the Turing test, which is this test that Alan Turing came up with that was if you had a, you know, behind the screen, and if you had either a human or computer could a human interacting with one or the other, differentiate between the two, could you have a computer program that fooled a human to think they were human. And that’s kind of been the litmus test for AI as long along the way, but it’s, it’s missing the point that ultimately, the technologies we’re using now for doing machine learning are somewhat akin to memorization. They’re memorizing data, they’re counting things, they’re computing probabilities and statistics, they’re doing some extrapolation of the data. But fundamentally, it’s learning from examples. And then regurgitating responses related to the patterns that have been seen. And that can only get you so far, like human intelligence. I don’t think it’s simply, you know, pattern matching and regurgitation of memory. So, you know, you can, the deeper repositories of knowledge, you can encode in an AI system. So we have these things called Deep Learning models. Now, deep learning means that they have these deep networks of hidden states that represent more and more information from the training data. So, you know, we used to have neural networks that had, like, dozens of states or hundreds of states. Now we have neural networks with billions of states, you know, trying to mimic the human brain. But fundamentally, they’re just kind of memorizing data. So if you’ve ever heard of GPT, three, this NLP chatbot, that is chatting, you know, natural language processing system, that allows you to do a lot of understanding from sub-level understanding problems in NLP, it’s basically just remembering an enormous amount of data. And then, if you ask it a question, it looks back, and it’s effectively what’s back. And it’s the history of all the things you’ve seen that look like the question you’re asking, and then take the kind of average best response based on its data. And again, that’s not the way human beings, you know, do innovative thinking. So you know, the, where we, where we’ve come today is just an extension of where we’ve always been, which is taking systems that just remember things and trying to make them look like they’re doing intelligent reasoning, but they’re really not.

Matt DeCoursey 08:03
So you talked about the training model. And you know, I’ve had conversations on the show with different people about, you know, AI and ML and, and when we talk about training, it does require a human to, at some point, say, Hey, this is right, this is wrong. In certain cases now, like, if you’re training AI to play go, you know, or chess or something like that. I think that it’s a little more straightforward that this is a winning or losing move, but how do you go about training? Ai? NML? Cuz, like, your impression of what might be the right course of action, compared to mine, could be very different. Am I right? With that?

David Magerman 08:44
Yeah. Um, so there are different kinds of learning methodologies. There is supervised learning, which is like what you’re saying, where you give it some data, and you give it an answer, and you say whether the answer is right or wrong. And you label all the answers with some categories. And then, you train models to predict those categories in new cases. That’s a purely supervised model. And that’s really useful in a domain where you know the answer. If there’s like, you know, if there’s, you’re taking images, and you have 10 categories, you want to say is one of the 10 categories, then supervised learning is a great way to go. But if you’re doing something like, let’s say, you try learning in order to like go when there are so many possible moves, it’s just it’s a comedy computationally impossible to compute all the moves in a game like go or chess there, there are unsupervised learning methods, and semi-supervised learning methods where either you just let the game you let the games play out randomly. And eventually, they lead to positive or negative outcomes. And then you have what’s sometimes called reinforcement learning, where you take the choices that lead to good outcomes, and you reinforce them, and you’re down to weight, the ones that lead to bad outcomes. And eventually, you hope that the models converge on models that can generate positive outcomes as opposed to negative ones. So those are purely unsupervised learning. And then you have semi-supervised learning where you might have a human in the loop, occasionally weighing in on whether, you know, guiding the learning. So it might be useful if someone has some unsupervised learning, and then the human will take a snapshot and say, Okay, let’s wait for these outcomes and equate these outcomes, and then continue the learning. And there’s lots of research going on, on semi-supervised learning because that’s a way of getting the best of both worlds where you don’t, you’re not biased by human intuition because human intuition is limited. You know, one of the best reasons to do machine learning is because machines will think of things that humans won’t, but you still have the limitation that they’re dumb. And you want to add some intelligence to it. So the hybrid approach of semi-supervised learning can be really powerful.

Matt DeCoursey 10:47
So if we discuss the limits of just AI and ML in general, I mean, as those are our models that rely heavily on supervised learning, do they become immediately limited?

David Magerman 11:00
Well, they’re only limited in the sense that there’s a cost to generating the training data. So if you have to annotate all your data for the supervised learning process, first of all, that means you’re coming up with the categories that you’re labeling, which, you know, people are not always good, as you said, your opinion of what’s right is different from micropayments. Right. And that comes true for ontologies as well, that if I’m coming up with a system of annotating it back in the 90s, late 80s, early 90s, I was a part of this entry bank project, where we were supposed to be labeling, like millions of words of data with grammatical structure, sort of like saying what the part of speech would be words, or each word was, what the noun phrases were, where the verb phrases, prepositional, phrases, and so on. And it turns out that people have different opinions about the structure of sentences. Sometimes sentences can be ambiguous. But also some people have different opinions, even about what the labels are for sentence structure. And so you have, so when you have someone coming up with ontology, a labeling scheme, that’s already limiting yourself to that decision-making. So if you’re doing supervised learning, you’re starting out with the limitations of whatever labeling scheme the person who came up with the labeling scheme put into it. And whereas if you’re doing something, something self-organized, there are information-theoretic clustering algorithms, which can come up with categories that humans would never think of, but which actually represent the data better than the brittle categories that one creates. And in addition, as data grows, as you know, we’ve seen with language usage, if you look at a Twitter feed, the language that people use today versus what people were using 30 years ago is very different. So if you limit yourself to grammatical structure, as defined in 1980, you wouldn’t be able to analyze most of what goes on on Twitter today. So there’s definitely a tension between the rigidity and structure of supervised learning versus the flexibility and fluidity of unsupervised learning. But if you’re solving a problem with clearly defined semantics, if you know the answer, let’s say you have, you’re trying to identify like from an image, you’re trying to identify a circuit board and say, what what what it contains, then there’s an answer. And there are specific components that the circuit board is made up of. And you don’t have to worry about the labeling problem because the labels are predefined. And so, in that case, a supervised learning situation is going to be much better than an unsupervised one. Whereas if you’re looking at something more like trying to determine sentiments in the Twitter feed to inform investment decisions, well, then really, a supervised process is probably very limiting because, like a European Union sentiment versus my thinking assessment might be very different. And you want something that’s gonna be generated by the data.

Matt DeCoursey 13:45
So the author of three books during the final editing process, I realized, you would think that the the English language was quite defined, and it was unbelievable how much you talked about sentiment or different opinions people had, like, for example, and there are things like, like the Oxford comma, for example, like people legitimately get upset about that. They’re like, Oh, my God, I can’t believe you would put that comma beforehand. And they get really upset about it. No, technically they grew up in America. Grammatically, it’s correct either way. But when you do, how do you train a machine to get past that kind of stuff? And really, like for real people, like calm down about the Oxford comma, it’s right, it’s right. Either way.

David Magerman 14:30
That’s a really important issue, the bridge between academia and the real world, which is that, you know, academic research tends to focus on problems that no one in the world ever needs solved. You know, whether you like that one, right? Yeah. So the first thing is finding problems that actually need to be solved by real businesses, and then figuring out if there are machine learning data driven AI based solutions that can actually do better than true traditional software. And the answer to that is, in a lot of cases, no. A lot of cases, traditional software does really well. And maybe data driven AI adds maybe five or 10% of performance improvement on top of it. But if you try to throw away the software, and do something completely machine learning based, you’ll get a really interesting system that will do interesting things, but it might not solve the problem nearly as well as the traditional software solution.

Matt DeCoursey 15:28
So when it comes to, you know, when I think about computing power, and just all that, and you know, 20 Here, wait, once again, jumping into that Wayback Machine. I mean, how much more difficult it was to run AI and machine learning models with limited computing power that we had 25 years ago. It’s like, I mean, that my iPhone that’s in my pocket right now is way more powerful than the computer I had 24-25 years ago.

David Magerman 15:54
Yeah, that’s a great observation. And that actually was made the main reason why I left academia, because I was actually doing my thesis research at IBM research on, we’re working on machine translation, speech recognition, and natural language processing. And we did these competitions, competing against labs around the world, every every year or so. And we had to train our big models for these competitions. And in the month before, these competitions, we would have to like steal time on hundreds of computers all around IBM Research, some of which we weren’t really supposed to be on, just to be able to terraform the environment to create enough computing power to run for weeks training our models, so we can be ready to do the computations for these competitions. And then management said it was okay, can you fit your speech recognition models on a desktop, and like, here we are on hundreds of computers, massive computers run IBM, and they’re saying squeezed onto a desktop, which now we’re talking excuse to get onto like a, you know, a cell phone or do it in the cloud. It’s, you know, a different world. But back then we realized that the mathematical techniques were using, first of all, there wasn’t enough data to train them. And there certainly wasn’t enough computing power to process the data. And even the disk space itself for storing the data. And the network saturation that we generated by trying to access the data was really onerous. So you know, a bunch of us at IBM left the field because we realized we were too early. And it turns out that when we fast forward 20 years and look at what solutions were working to solve AI problems today, they were frankly, the solutions that we rejected as inadequate back in the 80s. And 90s. Neural networks were a failed approach to AI back in the 80s, and 90s. Specifically, because there wasn’t enough data to create the deep networks, you needed to memorize enough information to make the models powerful. So really, it is Moore’s law, and the availability of the computing power and data and the internet that’s enabled us to solve problems today that we had no hope of solving 2030 years ago.

Matt DeCoursey 18:05
Is that data shared amongst people in the world on any level? Or is it all in compartmentalized proprietary servers that IBM or whoever, like whatever company has, I mean, is there any open source nature to all of this, oh, there’s a lot of open source data for doing general research.

David Magerman 18:22
But the really powerful applications are based on proprietary data. So you know, if you’re going to be working in a particular domain, as a startup, let’s say the first thing you need to do is develop a relationship with design partners that actually have data that they’ll be willing to barter with you to let you use to build your models or can sell to you or sell access to you or whatever. But you know, the the, the open source data is, by its nature, not powerful enough or not conclusive enough to convince to convince you that you have an application that can work that can sell that can solve a real world problem, because public data is, by its nature, not the same as the proprietary data that enterprises keep to themselves to do their core business functions.

Matt DeCoursey 19:11
I want to talk a little bit more about that and neural networks here in a moment. But as a reminder, today’s episode Startup Hustle is brought to you by equipment auctions, an online marketplace dedicated to growing small auction businesses, they’re solving problems and providing a fun re commerce or liquidation shopping experience to valued betters, go check out their incredible offerings and sign up at a quip dash bed.me backslash x I think that’s a Ford slash startup. So there’s a link in the show notes for that. Now you do not need AI or ML to sell your extra equipment, adequate bed. It’s funny with a good little nice little marketplace that’s dedicated to helping business owners get rid of all the crap that we have stacked up. And I’ll tell you what I’ve got a whole bunch of myself talking about the extra things in our business. Alright, so I first had my I was first introduced, like in depth into, like actually beginning to try to understand what a neural network was through the show. And there’s a guy in Kansas City who is an open, open CV or computer vision developer, who has a company that makes all of the technology that paint companies use in their app, to help you understand what color it’ll want your wall to be. And I was like, oh, man, that sounds pretty straightforward. You take a picture, and it’s gonna cover the wall, and this space is in green, and he’s like, oh, man, you don’t even know. And like, like, I’m looking at now, the listeners can’t see our videos, but I’m looking at them at the wall behind you. And there’s many shades of beige and brown, and that neural network, the way that your eye, your eyes, and your brain began to see shadow shades, depth and everything. And I was like, Oh, my God, if this is the amount of complexity that it is involved in, telling my wife, what color my kid’s playroom will look like, how are we going to figure out the rest of this? Now, you know, so when it comes to a neural network, is, you know, how can you try to break down and explain to listeners like what that is, and how that kind of computer hive mind in many ways will look at it and decide and come up with data or output.

David Magerman 21:33
Yeah, I mean, you know, neural networks are a way of encoding models, which we call nonlinear. So if a relationship is linear, so if you have, like, you know, a certain input, and that leads to a certain outcome, and if you have twice the input, it leads to twice the outcome. That’s a linear model. And that’s easy to represent with just traditional statistical models. But when you have relationships, which have nonlinear lot nonlinear components to them, where if suddenly reaches a certain threshold, and suddenly becomes much bigger, like if there’s a, if just in sort of the world, I used to work in stock prices, like if there was something where if the stock price goes up a certain amount, then that’s, you know, it’s got a linear response. But if it jumps more than three times, that becomes like a much bigger response beyond triple the response, then that’s a nonlinear relationship. And you need to have models that can include nonlinear relationships in them mathematically. And so neural networks are a way of building in cascade effects, which just mathematically will represent nonlinear relationships. It’s not, it’s not like, I mean, there’s a lot more to it than that. But basically, if you have a problem where you think things are mostly close to linear, you don’t need neural networks to model them. But in a lot of world problems, you either don’t know what the relationships are, or you know that they’re nonlinear. And that way, you need to have some mechanism that can learn relationships that aren’t linear. And so neural networks do a good job of that. The one downside of it is that they are very unstructured. So if you know something about the structure of your, of your domain and the problem you’re solving, it’s difficult to add that information to the URL neural net. So basically, a neural network is discovering everything from scratch, and learning everything, even things that you know. And so a human being can prime a statistical model with what we call prior knowledge of the relationships so that you don’t need as much data to train them. But with a neural network, it’s kind of free to discover any relationship at once. So it might come up with mistaken relationships that you don’t want it to come up with. And in a neural network environment, it’s hard to steer the model to include the human knowledge that you have as a prior.

Matt DeCoursey 23:52
I’ve had a couple conversations with, you know, AI and ML startups over the years. And it’s my understanding as well. But one of the limitations with some of this stuff is if it starts getting trained, or learning the wrong direction, getting it to uncheck its course and maybe change it, is that still a thing?

David Magerman 24:15
Yeah, for sure. And that’s why you can retrain models over time, you know, down waiting historical data and upgrading new data. So you effect there are there are now companies or startups forming up that are specifically designed to track the drift in data so that you can see whether your model that worked really well on data a few months ago, is still lining up well with the data that exists today, or whether whether the data is drifting, or the model is drifting in these wrong directions. So it’s definitely a significant problem. If you’re deploying Cisco models in the real world solving mission critical problems, you’ve got to have tools that track the drift in your models and the drift in your data to make sure you don’t end up in one of those situations.

Matt DeCoursey 25:00
Once again with me today, David Magerman, and he is the managing partner and CTO at Differential Ventures go to differential.vc To learn more, and that’s what I want to talk about next. So you’re fun you differential.vc You talk about one of the banners that says straight up a seed stage fund founded by data scientists and entrepreneurs for data focused entrepreneurs. So what’s with the fund? Let’s talk about that for a minute. What are you looking for, what you’ve done in the past, and like what you, you know, like what you see the future for the fund doing?

David Magerman 25:36
Yeah, so I mean, we’re focused on using my experience as a data scientist, and my partner’s experience as an entrepreneur and a sales and marketing guru to help guide early stage startups into the next level to get to the series A Series B round and hand them off to, you know, the the larger VCs to help them get to the next level. But for us, we’re trying to use our experience to keep the founders from making the kinds of mistakes we’ve made in our careers, when it comes to data science, you know, understanding the limitations of your data, understanding limitations of your models, and making sure that what you’re trying to do with your models, is actually legal, ethical, and consistent with the growth of your company. A lot of times, we find companies that are doing things like scraping data from the internet against their fair use policies, they’re using data in ways which would certainly irritate their customers and other users, if it went to scale, and things that are like, you know, kind of tolerated by data providers at the small scale. But if you scale up, then you know, they’re not going to tolerate it. And I always tell founders to like a plan for success, assume that you’re going to be successful beyond your wildest imagination. And imagine if your business practices are gonna stand the scrutiny of the light of day. And a lot of times, companies’ plans don’t, and you know, we can’t invest in a company unless it can scale to become, you know, a hugely successful company, that’s going to have all the regulatory scrutiny on it that you’d expect. So we try to guide them to do things in the right way, even if it’s legal what they’re doing. You know, legality is important. But it’s not the only barometer of what’s allowed. And if the market won’t, won’t sustain certain behaviors, even if the law does, you know, we have to get people to have a plan of how they’re going to grow their company, to stay within the bounds of what users find acceptable.

Matt DeCoursey 27:26
You had kind of a trigger point for me there. I talked to a lot of people that are, you know, a lot of people that are robbing startups that are well funded, they have traction, and they seem like they’re in the business of preventing the sky from following. And I’d say, what are you going to do? If everything goes well? What are you going to do? If everything goes, well? Are you prepared for that too, because, you know, we’re sitting there trying to prop things up and hold them up all the time. And, you know, and that’s the reality as an entrepreneur is, you know, you do have to avoid a falling sky, but what happens if everything goes, well, people? Are you ready for that too, because you can find an equal amount of failure from from that side of it, you know, I kind of, I don’t want to say failure, I ran into the same thing, my company Full Scale, you know, we’re over 300 plus employees four years later. And, you know, some of that stuff is sometimes you got to, we have known times in our past timeline, where we’re like, Hey, we got to just slow this down a little bit. You’re not gonna bring in any new clients yet, we got to get some we don’t want to we the ball of rubber bands is big enough, right. Now, let’s take a few rubber bands off of that. And, you know, I want to talk about the use of AI and ML within software startups. So yeah, it was before the pandemic, we went out to TechCrunch and had some people on the podcast and visited some clients and went to the show. And by the time we that first day had left, our joke was our machine learning model will. And it was like, it was like 90% of the companies we talked to had that line in their pitch. And we laughed and mentioned that being kind of a joke, because you know, some of these things never will need a machine learning model. They are never going to need AI. It’s just not a real thing. But yeah, that seemed to be the buzzword and the and the and the hype that everyone wanted to use to power their rocket ship to the moon. So at what point do you realize as a business that you may or may not be a software business, or I guess a business in general, that you may or may not benefit from, you know, use of AI and ML on your platform?

David Magerman 29:42
It’s a lot like blockchain, you know, if there was a point in time where like every company would fixture it was using Blockchain, even though it wasn’t just your blockchain was somewhere in the pitch. And your biggest question I’d ask is why? What can you do instead of blockchain and why are you doing that instead? But When it comes to, you know, machine learning, it became a big buzzword in data science and AI. But fundamentally, there’s this really important fact, which is that we have, one of the biggest assets most businesses have today are under utilizing is data. Companies have proprietary data that no one else in the world has. And that could be an advantage. It could be advantageous in a lot of different ways. And it might not be but you know, it’s something that needs to be explored. And so, you know, a good company will look at what data is being underutilized by companies. What problems could that data help those companies make better decisions about? And can they build software systems that can use a combination of traditional software development techniques, and data driven machine learning based models to combine their efforts to solve the problem better? It’s rare that there’s a problem. That’s not being addressed at all. So let’s say that you have to start from scratch and all this data science is probably the wrong place to start. But if you take a problem that’s been solved poorly, like, you know, project management is one, which is being solved. Every industry has project management software or not, but they have managed projects. And typically, the impression is those project management software systems are underperforming the poor, they’re hard to use. They’re, they miss things, they don’t send the proper alerts. And so those systems generated enormous amounts of data, which is ignored. There’s logs of what’s going on in projects, there’s deadlines that are being passed, there’s historical behavior by different employees about how well they’re doing on finishing projects, and what a warning sign is for that they’re not going to finish the project on time. And so you can take existing project management systems, and then throw data driven models that will look at historical data that’s being produced by those models by the systems and try to predict when managers should be concerned about the successful completion of a project, which might cost the company money if it doesn’t get finished on time. And so that’s an area where you can take a problem that is being solved traditionally, you have underutilized data, and you can build systems that can harness that data to solve the problem better. So I think that’s the, and I wouldn’t call that an AI solution. Because the fundamental backbone of that solution is software. It’s just a system, which is enhanced in his performance by certain AI driven tools, which are going to make the system better. And I think whether whether someone’s pitching you an AI solution or not, almost all in all cases, that’s what it is, it’s a traditional software system with a layer of AI or data driven components to it that’s hoping to take the parts of the system that were working most poorly, and trying to make them work better.

Matt DeCoursey 32:50
The key point there is creating data or insights that are actionable, though.

David Magerman 32:56
Exactly. And you need to know and be actionable in ways which can actually do better than the traditional systems aren’t. Because if you predict something that has already been predicted well by your system, those models aren’t valuable, they’re just becoming more of a source of complexity.

Matt DeCoursey 33:10
Well, maybe, you know, and, and, by the way, for those of you that, that have approached me personally, and asked me what the hardest part about being a host on Startup Hustle, as it’s sometimes keeping up with a Stanford PhD that specializes in AI and ML, so I try to unpack these things. If you don’t get everything he’s saying, don’t feel alone. We’re not You’re not alone. But yeah, but the key thing here is I mean, looking back at your data from the past, and being like, hey, these three things happen. And we lose a client or a user, who cares. Like your goal needs to be to get out in front of that, and be able to recognize when two of those three things have occurred, and saying, Okay, this is when we need to, we need to take action we need to do whatever it is that we do, and, and intervene or change the course or do something because if you’re just looking at it from a historical point of view, who freaking cares?

David Magerman 34:06
Well, so if your company we invested in comparative, which is showing why if you care.

Matt DeCoursey 34:12
But what was the name again? Let’s give him a shout out.

David Magerman 34:14
It’s called Parrot IV, okay. And so paradigm has as the Socrative student of developing, which will look back and see what events happen to your, to your customers that are precursors to them, either churning or reducing your services or not deciding not to buy something or misusing your platform in some way. So if you can find the triggers that are typically associated with negative customer satisfaction, then in the future when you see those combinations of things happening, you can trigger an alert to have a customer service representative. Reach out to that customer and try to address the problem before they churn before they decide not to buy your product before they get themselves in trouble by mistake. During the product, there are ways in which needed data analysis, historically analysis can produce insights which are actionable. But that’s the key point it’s a, it’s a great observation you made is that these things have to be actionable, just because you’re building a model that predicts things accurately, if the predictions don’t lead to something valuable as an outcome, you know, if you if you have a model that like predicts which stocks to buy, and you have another model, which predicts it better, but tells you to buy the same stocks, those two systems are not going to give an appreciable difference to your overall performance as a fund manager. So the idea is you have to have actionable insights from your models that cause you to do better actions and different things which make your business better. And that’s what we’re investing in companies that are finding ways to do that.

Matt DeCoursey 35:48
Yeah, and that’s not an easy task. People once again, this episode, Startup Hustle was sponsored by our friends over at quit bid auctions, join, sell, earn, it’s that easy to quit bid auctions, become an affiliate, and start to grow your independent businesses by visiting equip dash bid.me, forward slash startup today, even easier, head over to StartupHustle.xyz site and click on our partners page, you can also suggest future guests, or even yourself as a guest, that’s okay, we’ll accept that too. But if you are wondering how you become a guest on the show, it all starts there. That’s the top of the funnel for us to see the equipment, bed founder Andy has everything set up for you to go make money. So go build your business within a business. All right, so here we are at the end of an insightful and complex episode. And thank you for that. I like to let you know I’m starting my week out with a challenge. Fortunately, I’m gonna go to the Virgin Islands in a couple of days. And I might forget everything that we talked about here, but I’m okay with that, too. And this has been interesting, because I don’t get a chance to talk to people. I’ve had a lot of conversations with people that are newer in the space. But, you know, having seen all these different changing things over time has been really interesting. Now, you know, I think as we, you know, come to an end of the show, I mean, what advice do you have for entrepreneurs that are looking to break into the data science or, you know, machine intelligence, startup space?

David Magerman 37:25
Yeah, I mean, I would say that the important thing, most important thing is to have some unique knowledge for yourself, that gives you an advantage to help solve a problem that other people don’t know how to solve, you know, that if you’re going into a space that you know, if you’re going to space that everyone else is working on, and just going in there because you think it’s interesting, but you don’t have any any kind of experience that gives you an advantage, you really, you know, you know, fighting an uphill battle, you know that the AI is an unknown entity. Now, it’s not like a surprise, like 30 years ago, when people didn’t know the power of it. Data science is a well understood, useful tool. And you know, the question is what, what new problems that AI is not solving that you know, a lot about, can you figure out a way to adapt, soft AI software systems, machine learning data science, to add value to an industry that’s not currently benefiting from it. And I think if you find there’s a lot of Greenfield out there a lot of opportunities in spaces that are underserved by data science, and AI. And if you have an expertise in those areas, where you can develop an expertise in those areas, I think that’s a great direction to go in. Because you want to broaden the ways in which data science is being useful, not concentrate in the most focused areas.

Matt DeCoursey 38:39
I think you hit on a key point there. And when I’m listening to you talk about going into a field or an area where everyone is, what I’m hearing is you’re way behind from the beginning. Like people aren’t. They’re already way ahead. You’re not you. It’s easy to think in those cases that you’re ahead, and you’re not even on the lead lap. Like yeah, I’m winning. You’re like five laps behind. Now, I think the easy way to look at any kind of software startup, and if you want to attract you, obviously, the goal is to generate revenue, not just do science, like there are other avenues you can pursue, like academia, where that might be the case, but figure out how to how to help a business sell more or spend less, or maybe both, and you are on to something I mean, that is the crux of what so much of software and software entrepreneurship revolves around so like you talked about actionable data, you know, seeing things in the space that help people understand you know when a customer or client is going to churn or maybe why someone’s not buying something. You know, seeing things in space. You know, I work in the field of staffing and recruiting technically at Full Scale. You can go to fullscale.io To learn more, where we have 300 software division OPERS, most of which have no experience with machine learning and AI because there are not a lot of people that have that out there. And you know, one last thing on the way out like, and you know, I’m not saying it, you know, we help people build some really complex stuff at Full Scale. And like I said, AI and ML aren’t, aren’t they aren’t the things that I find people with a ton of experience with? How do you do if you’re an early-stage founder? How do you even fight? Do you have to create your own people with experience? Because I feel like you might need to?

David Magerman 40:33
Well, I mean, there are a lot of people graduating from colleges and graduate programs with these skills, looking to get into startups. There’s, I mean, it’s actually the labor force is kind of strange right now, although, with layoffs at big tech companies, the market will probably be a little easier to find talent.

Matt DeCoursey 40:52
I’m not sure they’re letting go of developers. So if you look at it, like I mean, I don’t see a lot of developers coming out of these layoffs. It’s mainly like sales and marketing type people that just laid off 11,000 people. I’m sure there were chunks.

David Magerman 41:01
Yeah. Guys, if I do a lot of AI and ml, but, you know, you really need to find the right balance. I mean, there are some open-source and off-the-shelf tools that do AI, but you really need to hire someone who’s got hands-on experience with data science if you’re gonna build a product that has data science as a major component of it. You know, it’s not commoditized. Yet, it’s still something that is brittle and can fail if misused. So you really want to find someone with experience. And I think it’s important to have balance, especially in a founding team, where if you’re a very deep tech technical founder, you want to find a co-founder who’s busy and business savvy. And if you’re a business focus founder, you want to find a co-founder who’s got some deep technology background because the best founding teams we found are the ones that have a good balance in those two areas. The rest you can fill in terms of sales and marketing and, you know, HR different things. But if you don’t, if you have too much focus on technology, you’re gonna have trouble building a product, and if you have too much focus on products, you’re gonna have trouble, you know, building the right technology, so you need to kind of find balance in both areas.

Matt DeCoursey 42:11
Listen to what he just said, people, it’s a key thing eventually you have to eventually have to stop and sell something to many people, our product product product product, and then next thing you know, it’s about being broke. Broke. Did you get it? Yeah, my co-founder at Full Scale handles more of the technical details on many days. I’m the one that looks at it and says, I don’t think we’re ever going to make any money doing this. And then it becomes a hobby.

David Magerman 42:42
Hey, you can’t have one without the other.

Matt DeCoursey 42:43
Yeah, if you’re not going to generate revenue or sell something eventually, then that sounds like a hobby, many times an expensive hobby. Once again, folks David Magerman, man, managing partner, and CTO at Differential Ventures. There’s a link in the show notes if you want to learn more about his fund or reach out to them. David, thank you so much for joining me and challenging me on a fun and interesting topic.

David Magerman 43:09
My pleasure. I enjoyed the conversation.