構造化データはなぜ取り残されたのか――グラフニューラルネットワークと Kumo のリレーショナル基盤モデル（Jure Leskovec 氏）

Let's go. You're listening to Making Data Simples, where we make the world of data effortless, relevant, and yet even fun. Folks, welcome to the podcast. I love that I get to do this all the time because I get the great guests and I get to ask a ton of questions and I get to learn. That's what got me here to begin with. Today, I'm with a most distinguished guest by the name of Dr. Yuri Leskovic. Let's go back. He's a chief scientist and co-founder at Kumo, and he's also a professor of computer science at Stanford University. Right. So he is a true pioneer in graph neural networks and network science. His research has reshaped how companies like Pinterest build recommendation engines, epidemics with computational epidemiology, and how enterprises unlock full potential of their structured data. Welcome, Yuri. I'm just excited to have you here. Thank you so much. Yeah, excited as well. Why don't you introduce yourself? Yeah, as you said, I'm a professor of computer science here at the AI lab at Stanford.

I've been here for 15 years. But during my career, I've always collaborated intensely and heavily with industry. So actually, I was part of the team who built the first people you may know recommend arranging for Facebook back in 2008, 2009. And that was a huge change in the slope at which Facebook grew. And then another kind of stint in the industry that was super amazing was I was a chief scientist with Pinterest for six years. Joined there when it was, let's say, a 100-people company and stayed post-IPO and really built basically graph learning platforms that changed the way Pinterest operates, were supporting a number of different use cases and really accelerated the growth of the company so that later it was able to successfully IPO as well. And out of all these experiences and learnings, about three years ago, I co-founded Kuma, where we are basically trying to bring this Gen AI revolution to structure the business data. Fantastic. You should be rich by now, man. You've seen so many things go IPO, and you've been a part of it.

So why are you still working? Do you love this stuff or what? I really love this stuff. And for Kuma, you know, I feel it's a category-changing company, and there is no better people to do this and bring this technology to life than the three founders. And that's what kind of, you know, powers up to wake up every day and keep pushing forward. I'm really, really excited about this. So you've been at very prominent companies before they went IPO in prominent roles. When do you know when it's time to move on? Because I've seen you move on to the next big thing. What's the catalyst that makes that happen? I always ask myself, you know, when deciding what to do, I ask myself three things. You know, who will care? Why should I be working on this today? So why work on this today? Why not three years from now or three years ago? And then the third one is why am I the best person in the world to work on that, right? And whenever this kind of moment comes and all these three questions can be well answered, I know it's the time to move on.

So with Kumo, it was exactly this new exciting technology that we developed at Stanford that can be truly transformative. So why now? Because now is the right time when kind of this AI revolution is happening. And why us? It's because we are the graph people. We are the best in shaping and developing this technology in the world. Cool. I'm going to jump into Kumo here in just a moment. And actually, there's four big themes that I'm interested in. One is why structured data hasn't had its chat deep moment. That's one. Secondly, how GNNs and graph transformers are breaking the bottleneck. GNNs, of course, graph neural networks. Third, high-impact applications, recommender systems, fraud detection, drug discovery, et cetera. And then finally, fourth is the future of AI. But before I do, a couple more questions to kind of lead us in there. You split your career between academia, like Stanford, and industry. Either Pinterest, now Kumo. How do you see those worlds? I mean, for those of us that are no longer in academia, how do you see those worlds operating together?

I think there is immense synergy and benefit by these two worlds kind of being connected and talking to each other. And there is no better place in the world than Stanford, where this culture of this discussion and exchange is really kind of the best out of any university in the world. So we have these amazing relationships with industry, and they benefit both sides, right? From, I would say, generally speaking, from industry point of view, you basically get to talk kind of about problems and you just have this sense where the world might be going. You know that you are a bit unsure whether something is there. And really, you know, young people at the university students who are all of excitement and have a very different, I would say, risk profile and risk appetite can go out of this kind of far out type problems. Right. And that's extremely valuable for companies. At the same time, it's also very valuable for me as an academic because I get to work on useful problems. Many times in my research career, I see two problems and they both look kind of equally interesting.

But then I go to someone to industry and I say, hey, if I solve this left thing versus if I solve this right thing, which one would you care about? And they're like, I don't really care about solving the problem number one, just solve the problem number two. But at the beginning, they look equally kind of researchy and interesting to me as an academic. So having that feedback from industry to say, hey, if you solve this, this will be more useful, more impactful to me, really allows us to do not only great science, but also to be working on impactful science. Do you still teach or is it all research? I teach. Actually, my primary kind of job is to run a larger research group of around 30, 40 researchers, PhD students, postdocs. We always have a couple of visitors from industry actually embedded in my research group so that the knowledge transfer is very fast and efficient. And then a bunch of master's students, undergrads, and so on. And as part of this, I also would teach one kind of advanced course every quarter to also kind of teach the students, but also then prepare them for them

that they can kind of join the lab and get the necessary background they need to be able to be effective at solving these research problems. I know Stanford's your favorite, obviously. Is there any other university you look up to? I think there are many aspects of good. I think every university has its own personality. But in terms of having real-world impact and creating the noise in the real world, like being heard, being seen, and having the impact, in my humble opinion, there's no one that can come close to Stanford. And when I joined, I honestly didn't know all these things. but just the way people are here the way the environment is just makes these things, the way the university is run, the way it's open for any student from any discipline to be able to be let's say mentored by any professor by being able to take any course at the university and also I think our students just being so forward looking and so engaged makes this like a super super super good place

so I'm really, I feel blessed and I'm thankful that I'm able to be here. The Stanford trees, they're the best. I got you. I got you. So one more question, and then I'm going to jump into Kumo, and that is at Pinterest, most people recognize Pinterest. You did a lot of work on transforming recommendation systems. What were some of your biggest surprises or learnings from that experience? Pinterest was a very interesting experience, right? And maybe there are several different ways how you can look, let's say, at the Pinterest ecosystem. and maybe there were two things to say that I pioneered that Pinterest had kind of tremendous impact. The first one was to realize that even though Pinterest is not a social network, right, you don't follow other people or anything like that, you don't interact with other people, just modeling Pinterest, thinking of it as a graph, provided huge leaps forward, right? And the way we started thinking about Pinterest was to think of it as a heterogeneous graph, so we have different types of nodes. So, of course, every user is a node, but then every pin, every image is a node, every board becomes a node, and now these things become connected.

So that was the first realization. And why these connections become important is because they allow you to resolve or to learn about what an object is, not only from the properties of the object, but also what other object it is related to. And maybe the biggest thing, the easiest to explain this in the Pinterest use case, if you look at the example of images, right? Like, for example, we found out that the neural network's computer vision had immensely hard time telling apart a rug on the floor versus tapestry on the wall. Visually, these are so similar, it's very hard to take them apart. But if you then look at how users categorize them, how they put them in boards, and how they interact, these two objects could be more different. Or another example was, you know, some picture of some ground meat versus picture of soil. They also looked about the same. And the neural networks, the AI would be confusing these things when making recommendations and classifications and so on. But if you put in this relational structure on top of it, then it becomes very easy to kind of tell these objects apart

and really gives you another level of detail, another level of granularity from which you can learn to make more accurate, let's say in this case recommendations, forecasts, whatever it is. So that was the first, I would say the first, this kind of graph view, everything is connected, everything is related type of the Pinterest. And then the second thing was that we actually said, we are going to build these models purely informed by people's behavior. So not informed by what things are, but based on how people behave. And this meant that we created this universal embedding space in which everything was embedded. People were embedded. Pins were embedded. Boards were embedded. Search queries are embedded. So now when you search for something, it's not that we are doing work toward matching. It actually goes into the embedding space and says, what's near this query? Those are the results for that query. And all this is learned through this massive, let's say, data that Pinterest is collecting from its users. So those are what I would say to be learned. Thank you for that. I appreciate it. So that serves as a good liaison to Kumo.

So what's your best two-minute pitch? You're the co-founder. I presume that you the techie that they go to when there design and everything else that goes into the technology What your two minutes technical pitch So we have built the first foundation model for business structured data That's one way to say it. Another way to say it is we have figured out how machine learning models, how AI models can learn over the data that is spread across multiple tables. And just this simple observation, this simple solution has kind of profound impact on how models are built, how performant they are, and what they can do. Kind of taking some of the learnings you did mention from Pinterest, it sounds like. Exactly. Exactly. So maybe I can say now I can go deeper and say a bit more. Yeah, like you said, you started, you said, oh, you know, why structured data, basically in some sense, the most valuable business data enterprises have,

why has this thing been kind of left behind in the AI revolution, right? If you want to build any kind of, let's say, predictive machine learning model, any kind of AI model on structured data that is forward-looking, you have to go back and use 30-year-old technology, right? In a sense that you say, okay, I'll create a training data set. I need to join this data that is spread across multiple tables into a single table. I need to do feature engineering. Then I train my model. It can be a decision tree. It can be logistic regression. Or it can be some fancy neural network. It doesn't matter. But, you know, you train that thing. And now you put it in production. You write all these pipelines that update and generate these features, and features need to be point-in-time correct. They must not be stale. They must not be from the future because then you have information leakage. And it's like a total mess. It's like half a year, over half a year of work. So when we were looking at this, right, like then you can say, okay, but, you know, how about Pinterest? At Pinterest, we said, Pinterest, you know, yes, you can think of it as a set of tables. You have a user's table. You have a pins table. You have a boards table. you have a pins-to-boards table and things like that,

we built a platform that would represent this multi-tambular data as a graph and then would just learn models on top of that graph. It took us several tens of millions of dollars of investment and three years to build this platform at Pinterest. With the best people, all the know-how from Stanford, it still was a humongous investment. And that was kind of, I think, the key realization for Pinterest, was that building these graph learning systems is so hard and so complicated that not many organizations in the world can do that. And in the open source, there is a couple of kind of libraries, but no real production-ready system that you could use to do this. So that's so difficult. What makes it so hard? What makes it hard is that in graphs, there is no natural unit of sharding, right? Graphs are connected. Images are not, documents are not. So you can stream them. They are kind of linear. You can just stream them through the neural network. But graphs are connected. So learning over this one big giant blob that's all interconnected becomes very, very hard.

Of course, you can say, oh, we have graph databases and things like that. But those graph databases have been built for a different type of query. They've been built more for this kind of transactional type queries. Give me an edge, remove an edge. They haven't been built for AI. So there are orders of magnitude too slow or inefficient to really build this, to support this high throughput AI model learning over graph structured data. So now what's the innovation coming out of Pinterest is that basically we can connect to any database, to any set of tables. This could be, let's say, a user's table, a product catalog table, user customers order's table and things like that. we for the first time now have a neural network that can just learn over the data that is spreading these tables, right? So the key kind of technical advancement is that now you can truly have a neural network that just learns over the raw data. Of course, there are, you know, other, I would say, these auto ML platforms and so on. But what those do in the background is they would join the tables together and engineer features, right?

But the innovation with Kuhn is that you have now a neural network that just learns from this raw data. And the way you can do this is that you can take these tables, represent the underlying data as a complex, temporal, heterogeneous graph, and then basically learn or train a neural network or a graph transformer, so a transformer architecture that is used for, usually we think of it for tokens, for LLMs, but you can generalize that and extend it so that it can now not attend over a sequence of tokens, but it can attend over this recursive structure of a graph. And if you do this, when you now build models over this type of data, you get two big advancements, benefits. First one is that building these models is now much faster because they learn from raw data. You don't have to feature engineer. And then the second benefit is that these models tend to be actually more accurate than what manually built models are. What gives them the accuracy element? Yeah, so here's maybe the idea, right? Like if you would say, I want to build a user churn model. Everyone knows what user churn is, right?

So you could say, okay, here is Jure, you know, let's predict whether he will churn. So then you say, okay, what features do I use to describe Jure and his behavior? So you could say, okay, I have his age, I have his location, and then I will count the number of purchases he made in the last one week. And then some other person says, no, no, it's not about the count of purchases in the last one week. It's about how much he visits our website past 8 p.m. So now let's aggregate his number of visits in the evenings in the last, I don't know, 17 days. Then some other data scientists or whoever comes and says, no, no, really the signal that predicts whether somebody is going to churn is whether they buy expensive or cheap products. So now you say, okay, we need to decide how long of a time period and how do you define expensive? How do you define cheap? Is this the total number of orders, total value of orders? Is it the median order? Is it the average order? Whatever it is. So essentially what we are doing today is we are computing some arbitrary statistics over this data and then kind of, you know, putting our hands together and hoping that these statistics predict whether you return or not.

But what a neural network that can learn directly over my individual events and individual data, that neural network can attend over each individual event at such a fine-grained level of resolution that it can learn signals that one will never put into a feature. And that's the reason why this becomes more accurate is because it has kind of higher fidelity. It can learn, you know, second-order correlations. It can learn some fine-grained patterns that you writing a SQL query that generates that feature is practically impossible. That's the intuition why this becomes more accurate. Are there other competitors doing this? I would say not that I know of. There are basically these AutoML platforms where you bring in your tabular data, But what they will do is they will first join this data together and then compute, you know, 17 different statistics over each column and hope that some of these brute force computed statistics are predictive. Our approach is different. In our approach, basically, it's the attention mechanism

that attends over all these low-level signals and learns how to aggregate them into a very strong predictive signal. And it's learned end-to-end. I presume you're also going to tell me a vector database or something like that's not going to work because it's more suited for unstructured data like images and text. Graph is going to store relational data. So are you building, you're actually building this LLM based on graph on GPUs, just the same as every other LLM, it's just a graph database with structured data? That's a good question. So let me maybe walk you a bit through the journey. So our immediate thesis, why should I move on? My other co-founder was a CTO at Airbnb. But when we came up with this insight, we said, look, every database can be thought of as a graph. And if I think of it as a graph, then I can bring this AI revolution to the field of structured data that seems to be stuck in the past. So that was the first realization.

That's why we said we have to go fund Kumo. We could easily already retire, but we said, no, we have to do this. It's so important. It's our life mission to do this. So that was the first observation, and we built an industry-scale platform that can ingest this business data, automatically represent it as a graph, so as a user you don't even have to worry about it, and build you this task-specific graph-based models. And we see a lot of adoption and a lot of benefit of this approach across a number of different industries and across companies that are super, super sophisticated to companies that are less sophisticated. So it's amazing that even to the most sophisticated Silicon Valley companies, this technology provides lift and value over what they were able to build internally with traditional approach. So that was kind of the first, let's say, two and a half years of the company. But what we have recently built is something that's kind of unbelievable, right? I couldn't believe when we started to work. And it's really, you said earlier, a ChatGPT moment for enterprise or for structured data.

So basically what we have now is we have a pre-trained foundation model. The same way as, you know, ChatGPT is pre-trained, you point it to some text, you give it a question, and it gives you an answer. And we have kind of an equivalent or analog of ChatGPT, but for structured data. So we have a pre-trained foundation model that you just point to your tables, you prompt it with what kind of prediction you want, and half a second later you get an accurate answer. So we have basically a foundation model for machine learning. So a model that is pre-trained and can be pointed to any database, be prompted with a task specification, and it will give you a prediction without any tuning, training, anything like that. It's kind of crazy once you start thinking about it. It's kind of revolutionary. So you're not using a vector database. You're not using RAG or anything else like that because you're saying you can put any database into this LLM model on the fly, and it'll give you predictions. Correct. And it's not an LLM because an LLM does text in, text out.

So LLM, you know, is great at kind of an LLM like chat GPT like LLM, right? Like it's great at this common sense reasoning, like human-like reasoning, but it cannot do forecasting. It cannot do fraud detection. It is terrible at recommendation. And our model is basically complementary to LLMs. LLMs are great at common sense reasoning, kind of human-like reasoning. Our model is great at prediction It connects directly to your enterprise data to your structured business data and based on patterns in your data gives you an accurate prediction an accurate forecast an accurate lifetime value, an accurate churn label, and so on. Again, without any training. It's a frozen model. I presume map the indexes and everything else? Oh yeah, good point. Good point. What does the model, we call it Kumo RFM, RFM for relational foundation model. We actually have it available. Anyone can go to kuma.rfm.ai, and start playing with this to actually see and experience what I'm talking about.

But basically, how this operates is the following. First, you need to connect it to your database. So generally, this means that you need to point it to a set of tables and verify that primary foreign key connections were correctly inferred. So that's the first step, is to say, here's my private data, connect to my private data. And there is a bit of mapping in a sense to say, oh, this column is a string, this column is an integer, this is a primary key, this is a foreign key. You know, this column in this table matches to this column in that other table. We can infer a lot of this from column names and so on, but sometimes it's not perfect, so let's say a human needs to verify that. And now that the model is connected to your data, the second step is now to prompt the model. and the way we prompt the model, we have basically a domain-specific language that actually looks like SQL, like SQL. But if you think about SQL, SQL is about you aggregating the past. You say, how many purchases did you make last month?

What was the sum of purchases you made last month? And with our model, you can ask it, tell me how many purchases will you make next month? Tell me the sum of your orders next week. Tell me whether this unlabeled transaction is going to be labeled as fraud in the future. And that's how the mod is now basically know what the task is, how it's prompted, and then gives you that prediction. And it'll do that out of the box. It will do that. It's frozen. There's no training. There's no tuning needed. So let me give you the scenario. I've got a buddy that works in industrial, essentially big machines, and he is a franchisee of these big machines. He sells these machines. He talked to me the other day. He says, look out, I'd like to be able to look at my parts and very simply, I mean, there's many questions he'd like to ask, but the question above all right now that if we could answer that question, then he could answer many more, is I want to know which part is requested the most that we don't have in stock. That would be then super simple for your model right out of the box.

Yeah. So, for example, what we could do is imagine he has his catalog of parts, his technicians, customers, and then service visits and this kind of complex schema describing the historical patterns of his business. And, of course, what we could then do is for each of the parts, we could, let's say, forecast what is the anticipated volume needed for that part in the next, I know, one month. and then we could then take those forecasts, look at the current inventory, take the difference, and wherever that difference is negative or close to zero, you order those parts. We've actually had a very similar use case that you mentioned, but in a very different industry. So the industry is banking, right? And the industry was, they have ATMs, and they need to supply ATMs with cash. But you don't want too much cash in the ATM, but you don't want to run out of the cash in the ATM because then people cannot withdraw money. So the task was, can you forecast how much money is going to be withdrawn

from different ATMs so that then, you know, your supply team can go and supply these ATMs with cash so that ATMs don't run out of money but also that there is not too much cash in those ATMs. And it worked beautifully. Like they were able to bring down the cost of resupplying by over, I think, 20%. So similar use case, a very different industry. Seems like that would be a run on ATMs because they only let you withdraw so much, so it would have to be a number of people going there. But I got it. I got it. But the point is that you can basically accurately forecast, let's say, the withdrawals, or in this case, how much of a given part will be needed. And the beautiful thing here in this forecast is because the model is learning over all your historical data, it can learn all kinds of patterns that are geographic, temporal, seasonal. you can learn from correlations that, you know, to service something you need this part and that part and so on and so forth. That's the beauty of this approach is that you can just point it to your raw data and you get out the accurate prediction.

So I know you're the techie on this, but how do you monetize this product? Yeah, we essentially let people use it and then we charge based on the amount of data, based on the amount of compute, right? So it's basically a software as a service. We provision the platform, run the hardware, and then the monetization is through the, in some sense, amount of compute people use. And I will say we also have what is important here, is we have different deployment modes, right? So, for example, we can run our platform inside Snowflake as an application. So all the data stays inside Snowflake, but all the compute is done on Snowflake hardware, on Snowflake GPUs. So this means that it's super secure because we run like an app inside Snowflake, no data leaves. Then, of course, we have the SAS model where we provision the hardware and the data would go to us for our GPUs and then go back. Or we can also install in people VPCs, so virtual private plug. But does that mean you always have to have GPUs to make this run? You have to have GPUs, yeah.

This is an advanced neural network that requires GPUs. It is computationally cheaper and smaller than these large language models that have tens of billions of parameters. So from this respect, it's much cheaper, but you still need the GPUs. By comparison then, by and large, how many GPUs will you have to procure to make this? I know it depends on the application, but say on average, what do you feel good about in terms of a recommendation? Depends on the data size, but for most of what we see, it's a handful. And training, if you tune these models or train them, doesn't take that long. takes half a day, a day, or several hours, you know, like on data sets with millions or tens of millions of data points, maybe like half an hour. So it's very efficient from that point of view. We are not talking about months of training over tens of thousands of GPUs. But then the only negative, I guess, would be that this is really targeted for the enterprise, somebody that's going to, which is fine, that's a big market. But if you're like a small, like maybe just $100 million business,

it could be too small or you're going to have GPUs, it could be costly. No, but then for example for smaller clients you can use our pre-trained model where basically you don't have to build or fine tune specific models you can just use the pre-trained model connected to the data and start getting predictions immediately. I would say the spectrum from I am ready I want to be making a million predictions a second, I want a super accurate small model, using our platform you can build that, or you can say, hey, I'm a smaller business, I want it to be contained, I need some number of predictions per day, per hour, per minute, and you can run that at a much kind of low volume as well. So the platform kind of... Do you still need GPUs in that case? In both cases? You need GPUs, but GPUs are a click away on AWS or on GCP or whatever public cloud you are using.

So we haven't seen provisioning GPUs as being any kind of problem or bottleneck. And we don't need even the latest generation of GPUs, so you can use GPUs that are a few years old, and those are actually quite cheap. Is this all going to be done in the customer's data center or on the cloud, or how do you envision this being laid out in terms of infrastructure? More on the cloud. Either the customer's cloud account or our cloud account. But, yeah, cloud is a standard way to deploy this today. Either, as I said, public cloud, so we provision hardware as a SaaS platform. We can install the platform in people's own cloud accounts, or we can run this, as I said, as an application on top of Snowflake, where now Snowflake is providing the GPUs on which this runs. And then, again, no data leaves your Snowflake perimeter. So if you trust Snowflake, everything stays inside. I've got a couple more technical questions for you, but that was my next question, is how can the listeners or I go out and try this? Is there like a sandbox someplace where we can play with it and see it in action?

Yes. So here's the following. Right now, if people want to try this out by themselves, I would point them to kumorfm.ai, which is a website. There is a sandbox for them to use a pre-trained foundation model. So it's a Python SDK. You connect it to a set of Pandas data frames, and you write a single line of what kind of prediction you want out of those data frames and you get an accurate prediction. And this is the easiest way to experiment and play and really get a feeling for what this technology is. And we are getting, I would say, amazing response and reception from all kinds of AI engineers, builders, which now makes it so easy for them to integrate AI into their applications, into their workflows and things like that. Anything from hospital readmission prediction, to power consumption predictions so you can trade electricity better, to predicting wildfires and building risk reports, and so on and so forth.

So these were, for example, three use cases that people built applications around at the hackathon we did last week. So that's one way. And then what we also have is we also have an enterprise platform that allows then people not only to use the pre-trained model, but also fine-tune it on their own data with their own use cases, which gives you kind of a further boost in accuracy and performance. And for that, you have to reach out to us and you can provision it for people. I think we've answered why structured data hasn't had its ChatGPT moment. I think we've answered how GNNs and graph transformers are breaking the bottleneck. Any more on high-impact applications? Like, what's the best use cases? I mean, we've mentioned a couple of use cases already. I guess we've already dived into that. but any other use cases that are you know top of mind and any use cases that you might think are out of scope here Yeah good question So in terms of use cases so we see great adoption in recommender systems You know, I think it's public, like Reddit.

Reddit is one of the clients who's using our platform for recommendations to huge success. DoorDash is another example, and so on. So a lot of recommendations, I would say, in retail, in media, and so on. Another good set of use cases comes from this customer 360 modeling, so predicting what the user is going to do next, what's the next best action, and things like that. Here, for example, we have use cases. I think actually also I can say this, both Databricks and Snowflake are actually internally using Kumo for some of their machine learning prediction workflows, especially on the sales side in terms of sales lead scoring or predicting what's the next product that the customer is going to engage with. So basically even on B2B type data where you have a few thousand or tens of thousands of data points, you can use our technology to build very accurate models. There was a cool use case with one of these two clients who basically wanted to build a sales lead scoring model,

and they couldn't build it internally with feature engineering because it's unclear what features predict whether somebody is going to become a customer. But our neural network just kind of attended over all this raw data and built a beautifully accurate model. So, you know, I said recommendations. I said this kind of customer-client prediction behavior type thing. And then the third set of use cases is around fraud, any kind of financial fraud or other types of fraud. These models work really well. And the reason they work really well is because they kind of identify these behavioral patterns. And another beauty in this fraud is that, you know, fraudsters are always changing their behaviors. So if you are building these feature-engineered models, you're always a step behind because your model deteriorates over time, and then you have to, because the fraudsters change behavior, you have to engineer new features to capture this new behavior, retrain, and hopefully you get back up. But our approach is just retrain the model, and the model is going to automatically identify this kind of new fraud behaviors

and you don't have to go through these future engineering steps. So we see a lot of benefit in fraud. What is not useful? What is maybe not a good use case? I would say these models are not meant to summarize documents. These models are not meant to do common sense reasoning or something like that. This is really predictive models where you want to be able to specify the task, the prediction that you want the model to make. But you can engage with it in English language, right? Just like in English. You can. If you are a human, you can engage with it in natural language or in more of this like SQL-like language. So you can write these predictive questions on the fly, or you can then fine-tune for a given predictive question, right? If you are saying, I want a very accurate model that predicts what ads you are going to click, I can pre-train that model for that, you know, kind of specific prediction. But I can also ask it in an ad hoc way. Anything we miss on Kumo or anything we miss on GNNs that you want to make sure that the listeners are aware of?

I think we kind of covered it, right? Like, I think the first key point is that any structured data, any multitabular data is a graph. So you don't have to be a social network to have a graph. As soon as you have data in multiple tables, you have a graph. With graph neural networks and graph transformers, you can now kind of attend across the graph of these tables to learn directly from them. So now you can finally, in some sense, apply this benefit of deep learning, representation learning, applied to structured data. And you get out these benefits. Of course, you still need a skilled operator that operates these systems, that is defining what we want to predict based on which subset of tables or what kind of schema and so on. So it's still a tool, but allows, let's say, the data scientists to be more effective and really focus on modeling rather than to be focusing on feature engineering and data manipulation. That, I would say, is the key take-home message. And then the latest exciting technical advancement is that we have this pre-trained model

so that now we don't even have to train models to make predictions. It's pre-trained the same way as kind of JGPTs. And it's ready for people to use and play with and give us feedback. I would really, you know, love to get feedback because technology is just technology. Kind of nothing is magic. And, you know, it works, it works, and it's applicable, but still requires thinking. Where can folks, before I forget, where can folks reach you? And I think you already mentioned the website, but I'll ask it again. So the company website is kumo.ai, K-U-M-O.ai. The relational foundation model for instantaneous predictions out of multithabular data, structured data, can be found at kumo.rfm.ai, RFM meaning relationalfoundationmodel.ai. We have a set of cookbooks. We have research papers on open source SDK, so it's free for people to use and play with it. Hey, so as a professor in the field of AI, you must have a lot of insight.

And I'm not even going to say three years away because three years is completely light years. I should say like a month. But I'm going to say one year. I'm going to give you one year and say, what is your prediction on the most profound aspect of AI you think we're going to experience in one year? Yeah, that's a great question. Let's maybe say how I see things, right? I see that these language models working with unstructured data, human-like reasoning, common sense reasoning, whatever we call it, is here, right? And I think the missing piece in this AI story as we are building these AI agents is the ability for these AI agents to make good decisions, right? If you want agents to become, let's say, autonomous and take actions, they need to make accurate decisions, right? And if you say, okay, what's the best way to make a decision? The best way to make a decision is to make a prediction on what's the effect of the decision going to be,

and then based on that, you take a decision, right? So this, we talked earlier, right, this kind of predictive modeling, this enterprise business data type AI was stuck in the past. But now with the relational foundation model, that becomes on par with the unstructured data AI. And we see that in the future, companies or organizations are going to use both, right? You are going to use this human-like reasoning AI, but this human-like reasoning AI is going to need to talk to this relational foundation model predictive AI to let it make accurate decisions, right? Because even if I'm a customer support agent, I need to forecast how likely are you to churn. I need to forecast what's your lifetime value. I need to forecast what offers are good for you. Because based on these forecasts, I can then take the right decisions to keep you as a customer, for example. Right? So really, I think for these agents to be truly intelligent in the future, you will need this non-hallucinated predictive ability.

And here we now see, I would say, great synergy between these enterprise foundation models, business data foundation models versus natural language foundation models. And I think, you know, where this is going is that even, you know, in the past we were saying data is the new electricity, data is the oil, right? But if you now think about, if we are all using these similar pre-trained models, then the only differentiator between us as businesses is our data, right? The quality of the decisions we can make is based on the data we have. So the importance of data is only going to increase in the future. This is my favorite question. It's a tough one, though. So if you don't like it, we can skip it. It's just what's true about AI, or structured data for that matter, that you know is true, but few people agree with you on? You know it's true, but nobody else agrees with you.

Yeah, maybe I would say just like from my past experience, right, it's either like this, like in some sense, embarrassingly simple insight that relational structured data can basically be, the right way to think of it is in this graphical forum, I either get two reactions. One is, this is so obvious, what are you talking about? Or the other one is, hey, you know, no, no, no, then we wouldn't have databases and so on. So that's maybe one such thing. But over time and over building this large-scale AI platforms in industry, I came to the realization that kind of this viewing the world as connected is the right way to view and model the world. What are you learning right now? What are you on the ground learning, or what's driving your attention? Yeah, what I'm really excited about is agents, and what I'm learning about and studying and doing research is about how to enable agents to have better decision power.

That's kind of what's where. Where do you go for that? Is it like academic research you're looking at, or is it information that's available to the public that we could all be looking at? I mean, where do you learn? I think it's two sides. Like being part of Silicon Valley, a lot is learned through conversations. And then, of course, the research papers. Yuri, thank you so much for being here on the podcast. It's been fascinating. Graphs or DNNs. I'm going to have to do even more research on my sauna. It's been fascinating. Thank you so much. We are quite excited. That's why you asked earlier, right? That's why I'm working with this. I think it's worth the push. Fantastic. I think you're going to have another audio here soon. That's my prediction. That's my prediction. See, I don't need an AI model for that. All right. Thank you. Thank you for having me. I appreciate it. Oh, my pleasure. My pleasure. And thank you, listeners, for listening, as always. Hit us on almartintalksdata at gmail.com. We'd love to hear from you. Until next time, we'll see you on the podcast. Bye-bye.