#62: AI Predicting the Future. What Happens Next? With Vanja Josifovski

I think that's gonna be extremely important work as we go forward. That's a new vector, and I think that's the one that scares people even more now, right? Is this thing gonna get conscious and malicious and selfish because it's learning from the human knowledge and the human knowledge is full of those things, right? It is human-like, so how human-like is it? Is it so human-like that it will have a bad intent as well? I think that's the key question. However, I think that's preventable. So far it seems that it's preventable. At scale.

Welcome to Startups Decoded, the show that gets real about what it takes to build a successful startup. I'm Andy Walsh, 2x exited founder, strategist, and your host. Each week I talk with top founders, investors, and operators to unpack the plays that drive real traction, not just headlines. If you're building something bold, this is where you'll get the honest insights to scale with clarity and confidence. Let's keep growing together. Hit subscribe and share with your community. It would mean the world Hello and welcome back to Startups Decoded. In the world of AI, we're looking at lots of information all at once and giving you all sorts of things. But what is the next step after that? Is it understanding how you can look at today and historical data to forecast the future? It feels like a world that 10 years ago you would have just laughed at, but now that reality is here. And today I've got Vanya Josifovski. Hopefully I did okay. My Australian tongue does its best. But welcome to the show, Vanya.

Hey, glad to be here, Andy. How are you today?

I am great. I am very good. It's been wild start to the year. We had the obligatory weather chat and I'm, as an Australian, I can't handle the cold here in New York City. But we're not here to chat weather. We're here to talk about what you're up to and your business, Kumo, is fascinating and we'll go into your backstory as to why you're doing what you're doing. But can you just give us in a sentence what, what Qumo is trying to do?

What Qumo does, it has brought the newest advances of AI, Transformers, foundation models to the most common type of enterprise data. That's a tabular relational data the enterprise is collecting to data warehouses. And then we use that data to predict the future and help enterprises increase revenue, reduce fraud, improve service for their customers directly from the most common type of enterprise data.

It seems crazy. And in context of today with AI and the type of work that we're seeing it do now, I just read a report this morning about medical advances and the ability to understand disease and diseases today that have zero cure. And in the next 10 years that some of those could be completely eradicated. It just blows your mind of the processing of data and where AI now is basically programming itself to develop the next phase. You don't need any human. Role whatsoever. So predicting the future in context of that actually feels like it can make a lot of sense.

Yeah. I mean, look, it's a fascinating time to be in technology or just around. It's like, I can't believe some of the things that are happening and I'm so excited about the future. You know, I think it's going to be amazing future for all of us. I'm an optimist. Yeah.

I'm an optimist too. And I, I want a future that's bright. And we're having a complete reset right now. There's a, the ceiling obviously has been broken, but I feel like that we're getting to a point now where the new baseline is starting to form. The foundations are definitely being built as far as what we can expect, how people use it. And I'd say by the end of this year, we'll have a real true representation of what the next step looks like in the role that AI plays day to day, but also any innovation that can roll off the back of that. And I feel like that what you're doing is right in the heartland of that.

Yeah, definitely. I mean, I would say that I agree with that 80%. I think 20% of me is even more excited than that and feel that there could be even, you know, better future than the baseline that we would set this year. There is a little bit of that unknown, you know, like watching a movie and, you know, maybe you're almost sure you know how it's going to end, but, you know, there's 20% chance that it will end somehow differently, right?

That's what a great movie does. Yeah.

Send you down a path. That's the AI movie we are watching right now, right?

In real time, in real life.

In real time.

Yeah. Lots of different characters in that movie. But to hell today, you've got a fascinating background in big tech. You've worked at Yahoo, you've worked at Airbnb, at Google, managing huge teams, but you actually started in research, not in pure tech. Can you take us through a little bit of that background and how that came to life?

Yeah. I mean, growing up, I thought I'm going to be a professor. So I went for a, you know, grad school and PhD, and then I was a professor for a while. And then I was just, just before the dot-com crash, I came to Silicon Valley here thinking I'm going to spend a year in industry and then go back and teach. But somehow once you come back, so many things are happening in this area that I just could never leave and went, you know, first went through the bust and then came up back again. And there's something to learn there about maybe how this AI, uh, you know, wave's gonna play out. And then, you know, as, as internet and, you know, came in into, you know, big force, I, you know, first I worked on databases, then I worked on search, uh, infrastructure. Eventually around '06, I turned into what was called at the time data mining and machine learning, which I would say is probably the precursor terminology to what we call AI today. We used to call AI something else at the time, you know, it's more of symbolic kind of thinking. And then from there, I think I spent most of my career, if I can describe it, in helping companies use data they capture to better serve their customers, users, to increase their revenue, to decrease the bad things happening on the platform. And this is not new. I mean, companies have been doing this for the last 20 years. Everything from recommendation, personalization, fraud detection, you know, product recommendations, ads. That's, that's where I've spent the last 20 years.

So you were in the trenches, like without the advance of the technology that exists now, which ultimately empowers you at scale, but you're already playing in that territory with the future because you're looking at predictive behavior. You're looking at human behavior, but also the things that people do from a functional point of view day to day with how they engage digitally in different platforms through different contexts. So you're having this really sort of future insight anyway in living and breathing that data, I would assume.

Yeah. I mean, companies have done predictive work for quite a while. I think if, if you look at the history of AI, I would say, you know, in, in late '80s, early '90s, if you looked at the newspapers at that time, you would read a lot about AI then. You know, some people like maybe a little bit older, like my age, remember that. But, you know, at that point it was symbolic kind of an AI called Prologue, you know, so everybody thought we'll have doctors in 2 years and it didn't happen. But this is the second goal of AI. But in between, I think the methods became more numerical. And we went through several iterations of these methods. But yeah, companies collected data and used data to make predictions so that they can serve their customers better. The methods were not as scalable. They were not as complex as now. But what we have today in the world of AI is not something that popped out of vacuum. It's evolution of work. In engineering and scaling systems and in algorithms over the last few decades. So there is some continuity there in, in, in that.

Yeah. I had a fascinating conversation recently with a, a gentleman who's been working in AI for close to 30 years. And the early iterations, obviously very different to today. And one of his, um, patients he's working with right now has got no body movement. So quadriplegic. Cut, it's nonverbal as well. And he uses his eyes to communicate and actually write speech. And he's CEO of a real estate company and his eye movement can actually track into words. But now he's created a voice with him as well. So for him, he's being completely empowered to communicate in spite of the way that he actually functionally can move his body. And that kind of advance for me is just incredible because it's very human first and foremost, and his ability to do it. But it can also go the other way. And one of the topics that I want to be able to tap into today with yourself is an element of trust. And as AI starts to show up everywhere across every company, even like my parents who are so tapped out as far as new technology, they're using it and they don't even realize they are. But then there's times when people are using it and they have a real lack of trust. And with anything new that comes out, People are always cautious. You have your early adopters on one end of the scale who just want to be a part of the first wave of anything and everything, but then you have your later stage adopters as well, and they're the ones that you need to be able to win over. As far as where we're at right now, where do you think on that curve, we've obviously got a lot of early adopters, at what point do you think AI will start to get to a stage where people take it for granted, say like getting on a website and putting your credit card details in? Now normal and step above. Like, where do you think we are on the AI version of that?

Every technology, every new technology— I mean, look at the internet— has a potential to do something bad, you know, and be it with intent. And maybe, you know, with AI, that's the new kind of, uh, you know, possibility of things going wrong, is that, you know, it starts being conscious and intentional in, in some bad ways, or by accident. You— when you have technology that reaches so many people, our lives will be changed. You know, cars are there on the streets and some people are killed by cars, but we still have cars. And so the question is now, you know, how do we structure the technology and how do we evolve it to provide acceptable risk to majority of people in the society? And I think that's the key word is acceptable risk. Now, you weigh the risk versus the benefits. You know, if somebody tell you that taking statin, that's a pill, right, will increase your chance of you know, reaching 80 years, for example, by 60, 70%. And you have a choice to trust it or not, because this is something that's based on data. So what we call AI is, you know, predictive AI is like something that's based on data, and then a system makes a prediction about the future. It's up to you if you're gonna take that pill or not. But, you know, this is based on science and statistics, and it's, it's likely correct. I would take the pill. Now, some people might decide not to take the pill, right? And just say, hey, I don't want to do this. But I think if the risk versus benefit is there, then people would accept it. But I think one thing to understand is that, as you said, people are, you know, people are affected by machine predictions today hundreds of times during the day. Like already, every time you swipe your card, something determines if this is a fraudulent transaction or not. And it might block your transactions at will, right? Every time you open a website, something determines what content to show you, right? Every time you go to your favorite social networking app, it's all based on predictions of behavior. And we all know that some of this is good and great and some of it is less, right? We all, you know, I'm not going to repeat the issues with social media and how it impacts our society, politics, and everything else. But it is here for us and it's a fact of life. That our lives are impacted hundreds, maybe thousands of times a day routinely by machines using data to make predictions. And that's just gonna go further. Now, are we gonna be more trusting to machines? Is this gonna increase with the newest wave of AI, what we're doing at Kumo, what's happening in large language models? We can do a whole new way of predicting about our lives, personal lives. And are we gonna use that more or less? I think it's gonna depend on how good we are in containing the adverse outcomes of these things. And to some extent, it will be a personal choice. I believe in personal choice, and I think there should be personal choice about that.

Like anything, two people can take the same technology or same workflow and use it for good and for bad, right? That's, that's the nature of humans in the world that we live today. So you're always going to have examples where you feel like, I can't do this anymore because I saw the bad side. But then the upside of using things as well to make life more effective, easier. And not have to do so many things all the time, like the automation that we're seeing in businesses now rolling out and flowing into humans' lives just broadly. If you've got a family and a couple of kids, that's a full-time job in itself, just juggling all of those things. If you can take pressure off that and it understands you better, then that has to improve life. That's a, that's a given. So there's tension naturally that exists in all of those things and has for a long time with any change and the old adage of of the car when it was first invented and people were nervous about using a car full stop. And you use the same thing before, like now people die in cars, but doesn't stop them using them. There's a percentage of risk in everything we choose to do in life.

I think actually, I firmly believe from our work and what we see out there that in enterprises, most of the decisions that humans make today, we will automate it in the next decade or so. And that I say is because in personal life, People have choices, but in commercial environments, you have a pressure to do what's best for the company. And I think machines will be better in making decisions than humans in most of the cases. And that will change the whole way of how enterprises function. That mid-tier of decision-making is going to be automated. That's my controversial statement of the day.

No, I think it's an important statement to make. And it's interesting that as human beings, evolution over tens of thousands of years of the human species, means that we trust other people on certain things in certain contexts. Say a surgeon, like really serious moment in time. If you're having major surgery, maybe it's a quadruple bypass on your heart, you're trusting someone. But that someone also has a life. Maybe they didn't sleep well. Maybe their two kids were being absolute little shits the day before. Like you come into it with all these different things and that you are trusting that person to do something major. So we get into that tipping point where the mistakes that AI or LLMs are making are getting less and less compared to a human, which is, we're human and we know ourselves, we have good days and bad days. It's just the nature of being human, particularly in the world and the climate that we live in today. So that, like, we'll start to see those things become normalized just by the fact that AI is growing at such an exponential rate.

Yeah, no, I mean, I saw that firsthand. I had a family member that had serious surgery and, you know, those, you know, robots with 6 things and it's just amazing, you know, it's just hardly any, you know, hardly anything visible outside. Mind-blowing. And I think they're, they're going to get to the, they're getting to the point where it can do the whole thing itself. Which is comforting to me. That's very comforting to me. It might be not comforting to other people, but I think, you know, technology You know, in a lot of cases works, right?

And there's things that we've got used to over time. It's just to a point where you trust it. And as I said before, credit card payments are a great example where it's so easy now. You don't think twice about having credit cards saved on third-party sites and there's a level of fraud that happens, but there's a controlled risk to a certain degree as well. So that's always going to exist. It doesn't matter how good the technology is. It doesn't matter how good the team are as human beings that are doing something. There's always going to be gaps in that. That's just, that's just life. So there's an acceptance level that happens over time.

Yeah, I think we, you know, we have seen a lot of movies about AI going bad, right? From Terminator on. And I think this is the next level. You know, it's not a mistake, but it's a, it's an intent. And that's the new boundary and frontier here, you know. And I think when you, when you look at the companies building these large language models, I mean, they're putting a lot of effort to to avoid bad usage of this, either giving bad people superpowers or intentionally, or the models getting into conversations that can be harmful. And I think that's going to be extremely important work as we go forward. That's a new vector. And I think that's the one that scares people even more now, right? Is this thing going to get conscious and malicious and selfish because it's learning from the human knowledge and the human knowledge is full of those things, right? It is human-like. So how human-like is it? Is it so human-like that it will have a bad intent as well? I think that's the key question. However, I think that's preventable. And so far it seems, uh, we can, we can find interesting cases in the newspapers, but so far it seems that it's preventable at scale.

Yeah. And that's the same in life. You're always gonna have people that use things to their own advantage and at the harm of others and maybe don't think twice about that. But to fast forward to today and what you're doing with Kumo. Kumo is Aumo.ai. What you're creating is termed predictive intelligence, and you've touched on the ability to look at volumes of data historically and then start to create models for the future. Can you talk to us a little bit about that process as far as what are your data sources, what are you looking at, and even an example maybe of a workflow that you have?

So the first data that we ever collected was enterprise data. Right before the internet, you know, we had relational databases that collect tabular data about entities like, you know, products, customers, and about events, I would say. That's transactions. So what happened? And relational data is very good in collecting history, and that's why people collect it. And the reason companies collect it, because it helps them operate. And, you know, through the years, what we did is with it, we did analytics. We would analyze data to figure out what happens so that we can then feed it into a human brain that will then make decisions about what to do in the future. Nobody cares about the past. The past is fixed, right? So the whole reason companies collect data is to change the way they operate in the future. So instead of doing analytics and then feeding it into humans at Qooma, we use the newest in AI technologies, specifically the transformer network architectures, as the T in ChatGPT, and foundation models that now can ingest tabular relational data and emit predictions. So think of it, you have your data warehouse. and you can have Kumo, and you know, we can talk to it and say, hey, you know, which are my next 3 accounts that will churn? Or what should I do? What kind of treatment discount and so on should I give to my accounts so that they stay with me? Or what's the predicted spend? Or, you know, hey, Kumo, who will— will this patient be readmitted in the hospital in the next 3 months with complications from this surgery? Or should we release this patient now? Should we release this patient in a week? So you can ask all kinds of predictive questions, and now These agents on the top, they don't only do analysis and then feed it to people. They can make decisions. They can discharge a patient. They can improve somebody's outcome. They can predict sepsis. They can predict anything and do it on the fly and thus improve the outcomes. Yes, this sounds a little scary, but the likelihood that a human will make as good of a decision as a model that's based on the entirety of the data is very small. And these agents will be able with Qumo to use the predictive power in the data to make decisions and make them automatically. That's what we do at Qoom.

So very complex neural network of lots of information there, right? That you are distilling down very quickly on that. What are some of the inputs? Like you're looking at, you're looking at human psychology as far as behavior. You're looking at the industry itself. Like what are some of the verticals of data that goes into that network?

Yeah, so we built the first platform that's able to train these models and then we assembled a large amount of public datasets and across various domains. And one of the first things we needed to decide is how broad or how narrow this model is going to be. Should we build something that works for one vertical or should we build something that's going to work across all verticals? It turns out that we were able to build a model that's applicable broadly to, I would say, any vertical or most verticals and give you predictions about the future, pre-trained, all packaged. So you don't have to do any training. But this is a very interesting finding that we were able— because initially we thought it's going to be maybe vertically focused, but we were able to build a model that applies to, as far as we know, any vertical that we know. And that's a very, very interesting outcome. That means that, you know, because a lot of this is about human behavior and people interacting with things or, you know, it seems like the universe of different functions, it's possible to estimate with a single model. I think it's truly fascinating. But as far as the model, it seems like it's applicable to almost any domain.

It's crazy to think about it. And I go back to old research and understanding the difference between quantitative and qualitative, like really simple terms in a research standpoint. So you've got hard data and quantitative, but articulating that information and you know this as a researcher, you can tell a story 10 different ways based off the same data information, right? So the qualitative part taps into a little bit of maybe the more human side of like the art and the science of how you're forecasting. So those two side by side, you're able to get a sense of one, the information that you can predict as in quantitative data that you can probably forecast with a higher level of accuracy versus how a human would interpret that. But what you are doing, if I understand you correctly here, is actually removing that human interpretation because you're able to glean en masse who that person is or who the collection of people are and get a sense of how they interpret what they're seeing. Would that be a fair description?

I think AI in general went that way. I mean, this happened in computer vision and in text 10, 15 years ago. We used to find features in images like edges and then use those as, you know, to build models to classify images and so on. We used to find nouns and pronouns in text and understand that and then build modeling on the top. Today you largely feed raw data, so you don't use semantics. You use the raw pixels in the image and then you have a layer after layer of a network kind of processes that signal and finds it. You don't need to find by hand what's interpretable. And this is what allows these models to scale, massive scale. And the same thing in our case. You know, we use very little of the semantics. We require understanding the IDs, the foreign primary keys and timestamps because, you know, this is a temporal kind of a structure that we learn over, but we don't require understanding what's in the data per se. That's what allows us to scale this to the point where we can pre-train a massive model to be able to understand. The structural data. And then when we bring that model to your data, this model, we do what's called an in-context learning. It'll grab piece of the, your data, and it'll do that in a very smart way of what needs to, to be taken and then use that context plus the pre-trained model to give you the prediction. But yeah, we learn over the raw data. We don't— that, that's why it's, it's applicable to, to many different verticals.

Yeah. And on that process and workflow, Once you're starting to ingest data through Kumo and you're starting to get predictability, then the longer and the deeper you go, that feedback loop, I assume, then becomes even more powerful because what you are predicting and forecasting, you, you shift from analyzing current data into forecasting. And then I assume at that later stage, you are pairing forecast versus actual and seeing how close you were and that those nuances in that shift allow you to have another data source.

Yeah, that's, that's a, that's a very interesting observation. You can actually build an agent on the top of Kumo that will help Kumo. You know, we have an agent that, that uses large language models that kind of fine-tunes the parameters inside to improve the predictions. And that's a very, very interesting area. It's like you almost have like a, like a scientist who operates the model and improves the performance over time. Or you can use the entirety of your enterprise data by fine-tuning the model specifically for a certain enterprise or even for a certain task in some cases. But yeah, so there's, there's some operation that can be done and then on the top, and, and that is a very suitable task for large language models actually. So we have combined both together to get the best outcomes.

I love it too, because what it allows you to do, and I'm a brand strategist by trade and I'm all about storytelling and part of what you are doing to overcome some of the mental barrier or blockage that a lot of people have with with how far they can trust AI, you can share those success rates and you can start to do really basic stories as far as in certain use cases, the predictability was set at X. We had a 99% success rate based upon what we were doing. A typical human would be at 70%. And you start to give people like real feedback on what they're doing, which has to happen because people need to be able to trust this to get full access to make it as full benefit. If it's only adopted in part, I assume that it's never be like what it could potentially be at its full capacity.

Yeah, one of the key questions in, in trust is explainability, right? And people, this comes a lot, and, and in some cases is regulated, like in a bank setting, uh, you have to have certain explainability. Although, you know, it doesn't in traditional settings. I mean, I'm— my bank, you know, whenever I get the the credit score, when I go for the reason of why it's maybe not perfect, it tells me that I have, you know, that I had some kind of payments that were delayed or something like that, which is not true, right? I, for a fact, I don't have that type of a payment, right? So, but with Kumo, what we do is because we work straight off the raw data and this model is fully differentiable, so we can take a prediction. Let's say, you know, we predict that Andy Walsh will purchase another $300 worth of goods in the next 3 months. And so we want to know why is that prediction, we can take that prediction, we can run the model backwards and it will bring us to the raw data elements to maybe a, a transaction amount or a time between transaction or a product type. And it will give us a score of each of these raw elements and it will tell us how that impacted the prediction. Now, this is purely mathematical. There's no hallucination here. There's nothing like that. And then we can take that. Now, those scores might not be perfectly interpretable or friendly to a human. So now we take that and we run it to a large language model that makes the qualitative part, that gives a super nice explanation, that will explain exactly why Andy is buying $300 more of things. And, you know, for example, you bought a bed and people who buy beds, they, we figure out in some cases they buy more stuff. And so, but the combination of kind of solid scientific, you know, numerical, non-hallucinated explanation and then something on the top like large language models that are able to put the story around it, as you said, will give you something that has both sides. That's real and it's readable. Book.

I equate it to doing academic research in a certain degree in that if you are a doctorate or a professor and you are got a hypothesis of something that you want to test, you are forecasting an outcome at that point. And obviously it's not a bias, it's you have to have a hypothesis to get to a point where you can build out a testing protocol.

Mm-hmm.

And what happens is you run through all the testing and then you get a level of significance about how accurate that data is, but you're retroactively then looking back at like, it went one way. We didn't expect it to go that way. Like, why did that happen? Which is then going back to your core original data to be able to understand those pieces. So it sounds very similar in that way, in that, that you are predicting something based off what you know now, but you are then working backwards and getting a sense of like confidence level ultimately on how you can validate the source of origin of all those inputs. That got you to that point.

Yeah, I mean, that, that is correct, but it's all automated and it's instantaneous, right? You ask a question and, you know, the system does not need to be specifically trained for that question. It's like if I want to ask future spend, or if I want to ask churn, or if I want to ask the supply question, how much of this product will I sell? You don't have to anticipate that in the same way that with, you know, with ChatGPT or with Claude, you don't have to anticipate the, the precise question. The model is pre-trained and can answer many, many different questions. Now it's the same thing, it's just with your enterprise data. And so you can talk to it and you say, hey, what's gonna happen? Which product I'm gonna sell in 3 months, in 6 months? And how should I think about this? And it will give you those decisions. And I think once this is operationalized, decisions about marketing, decisions about discounting, even, you know, at some level strategic decisions in the company can be automated. You know, that's, that's where we are going. Imagine this big decision engine of the world that powers decisions in, in all the enterprises. That's the vision for Kumo.

Very big and bold vision. I did a recent episode with Frank Paz, who's a partner at Morrison Foster law firm, who's a partner of ours. He brought in one of his clients, a gentleman, Craig Alberino, who's the founder and CEO of LangSmart. And a lot of what they're doing is actually in large enterprise, looking at compliance and audit procedures around how AI tools are implemented. And there's a level of risk, right? In that it's not only how you use AI, but it's also the third-party platform in what is their own compliance model. And if you're publicly listed as a company, it goes even further again, because there's a lot of compliance you do to be able to remain publicly listed, but also shareholders, board members, like that legal liability goes a lot further. For you in a very practical level, how do you look at forecasting data and decisions and knowing that you've got enough information for them to have a confidence level to make a decision? And where does that risk lie? Because obviously you can predict to a certain degree, but you can never give anyone 100% confidence that something's going to happen, right? It wouldn't be that crazy. Crazy. So like you start to go into a future state of legal risk, how do you start to play into that forecasting model of how someone makes a decision and do they need to authenticate it or validate it in a parallel path that may be more traditional?

Look, I think some of the more regulated industries, finance, health, and so on, that, that those will not be the early adopters of, of complete automation, right? Let's, let's say that, right? I mean, it's gonna be, and I think rightly so. You know, we have to be cautious in places where the impact would be much bigger. However, I do think, and you know, the way we've prevented bad things from happening is through understanding. You know, I have a detailed understanding of what's happening. However, you know, it's going to be this very similar thing with self-driving, right? Currently, I think Waymo was the one that published that the rate of accidents of their vehicles is far lower. Than the rate of accidents of vehicles with human drivers. And so in this case, AI is better than humans and it saves lives. And in some cases, bad accidents will happen and things are going to go wrong. But on average, it's like it's not just a little bit, it's 10x safer. So what if you can do this for saving human lives? What if you can make medical decisions 10 times faster and better? What if you can make even financial decisions in a much fairer way? Than humans do, and you can measure that. And I think that's going to be the road. This technology will have to prove itself. And again, not just by a little, it has to be by a stretch because little people will not trust, you know, that's, you know, you can manipulate measurements, you can do this, but if it's 10x better, then, you know, it's really hard to fake that, right?

Absolutely. And I love the car example because the irony is not lost on me. A lot of accidents can be caused by people being on their phones, but a cell phone, social media, people's attention spans are in decline. They're so connected, but they're not connected. But when they're driving, they can't be off their phone for half a second, so they're on it, right? And I assume that that has increased the risk of accidents that happen where now you're putting technology into a car itself and it's making decisions for you to make you safer as well. And it's like two different technologies side by side having very different impacts and effects on what is an everyday occasion of getting in and driving a car.

I mean, are you not distracted by your phone in the car, Andy?

I don't use social media. So for me, it's, um, I'm actually okay. I'm quite content to sit and wait and I encourage anyone else, the more, the less time you spend on your phone, the happier and more connected you are.

Yeah, I get distracted. I must admit. And I'm waiting for the day when, you know, it's going to be easier to do other things in the car. I'm not very fond of driving. Driving. It seems it's a little boring for me.

There's a lot of times, particularly if you're stuck in traffic, it sends, it sends me wild. I can't deal with that. On the open road, sometimes it actually can be quite fun on a road trip driving. And like to have choice would still be good. I'm not sure if I have to hold on to an old car in the future just so I can make that choice. Who knows? But we're not here to talk about cars though. But what I am fascinated though, is I look at other industries and I'm gonna look at a classic old movie reference as well in Back to the Future and in the second film, the Sports Almanac comes back in time and you're making future bets on sports odds. At what point, I mean, obviously betting places now are using a range of data and AI to analyze odds and try and protect the house, so to speak. But is it going to also eliminate some of those categories where people who are the ones taking the bets, is it going to be almost impossible to make money because forecasting becomes so accurate? Can it get to that level where you can understand and forecast results of a sporting game?

I think it can. And I think definitely, you know, even now you have models that look at the second by second of games and predict the, you know, the next outcome. I, I think we are already there. It's probably the question of gathering the data, which, which is probably not going to be a problem as well. So I do feel that, you know, that field will probably get harder because everybody will get access to super accurate predictions instantaneously. You know, I mean, they, you know, these companies will probably change the rules of the game and do all kinds of other things that will then make, you know, that will preserve their ability to, to do better than their them, you know, and their customers. I could see ways of how you could do this and change the rules a little bit. But in general, what's happening today, yes, everybody will be able to get a prediction instantaneously. You know, the future of AI in my mind has three kind of features. It's gonna be natural language. You're gonna talk to it. The input of AI is natural language. You know, we had the era of the phones. Now with the large language models, everybody's just used to talking to AI. So I think that's going to be definitely how we're going to interact with it. And it's easier than watching something. The second is going to be instantaneous. You're not going to ask something and then it's going to tell you come back in 2 hours. Everybody expects the answers to happen immediately. And the third part, it's going to be predictive. You're not going to ask what happened, but you're going to ask what should I do? What will happen? That's my prediction about AI. Natural language, instantaneous, and predictive.

It's such a fundamental shift in human behavior as far as how we think about making decisions, the input that we have, the amount of people we talk to, the reference points, like other people's examples, et cetera. This is like a dynamic process that happens quite passively sometimes in our own brain. As far as we're, we're, we're doing that every minute of every day, making decisions, like little micro decisions. Sometimes they're larger ones. But we'll get to a point where, as you said, everyone expects an answer immediately and it shortcuts your own decision-making process and you're bringing external support into that and it changes human interaction. And I'm always fascinated on the human psychology part of it. Like evolution by default takes a long time to shift, right? As far as human beings, our bodies adapt into that. So you're going to have this huge void between the way that we default behave on certain things that we do in our life. Versus where technology is at. And for us to catch up where we're doing that, it's really fascinating social experiment about like, at what point does that become normal? And like we have a cell phone in our pocket, I guess talking on that has become normal and expected. I guess that's where we're going.

I think, you know, because we're going to be less involved, like today, the technology provides you the input, then you have to process that. And, uh, And that's why we have the screen formats and you have to consume a lot more information. I think we're gonna be consuming less information because it's gonna be more about predictions and about the future, and that will reduce the need for input devices as phones. Now, how are we gonna catch up as species evolutionary? That part, I'm not sure.

You're not predicting that one for us yet?

I don't know. You know, that's the fun part.

Yeah, that's the tension that we're seeing. And like the technology has a clear path and how you live, how you work, the type of things that you're doing. There's a lot of data now that's coming through. It gives us a sense of that. But what you're doing with Kumo is a fascinating social experiment as well as a very interesting business model. So I'll be staying close, but Vanya, for today, we're running out of time. I'd love to leave you with a moment And as far as the type of work that you're doing, what sort of folks should be reaching out to you and the Kumo team to start testing some projects and protocols with you?

Yeah, any type of predictive problems in enterprises. I mentioned a few. We do personalization, recommendation, fraud. We do agents for sales teams. We do agents in the medical space, you know, to support doctors. Every time there is tabular relational data and you need to build a natural language agent on the top that's able to instantaneously give you predictions of what you should make decisions out of this data, that will be, you know, what we do.

Fascinating. And one last question on that, actually. I assume then that your strength is on proprietary first-party data of the company that they have. So everything is contained internally rather than relying on external sources and data scraping somewhere else?

Yes, that's correct. We work with enterprise private data.

Awesome. I'm fascinated to have a play. I'm not sure if there's a little example I can do in my own world now with my podcast and everything else I'm doing around that. I'd love to have a play, but today, Vanya, this has been really fascinating to look at predictable intelligence and where we're going in the future. So I appreciate you making the time to have a conversation.

Of course, this was a fun conversation, Andy. Anytime. Thanks so much.

Thanks for listening to Startups Decoded. If today's conversation helped you think differently, share it with the founder in your network. And don't forget to subscribe for real-world startup insights every week. We really value your support. If you're interested in partnering with us, email hello@startupsdecoded.com. Www.coded.com. I'm Andy Walsh. I'll see you next week.