The Exchanges

Q I remember going on a walk with you in San Mateo in like, 20, 20, something like that. And you were like, you have people reaching out to you who are like, Hey, we use Cuban production. Like, I just need to give you some money. Uh, even though you guys are not a company. Uh, what, what's the story of Cube then from StatsBot to where you are today?

A We, we built a Cube at StatsBot because we needed it. It was like the whole StatsBot stack was that we first tried to translate the initial sort of language query into the, uh, some sort of multidimensional query. It's like we were trying to understand, okay, people wanted to get a active opportunities, right? What does it mean? Is it a metric? Is it what a dimension here? Because usually in analytics you always, you know, like try to reduce everything down to this sort of, you know, like a multi-dimensional framework. So that was the first step, and that's where, you know, like it didn't really work well, because we all this limitation of us not having foundational technologies, but then from the multi-dimensional query we wanted to go to SQL, and that's what was semantic layer and what was cube essentially. So we built a framework Where you would be able to map your data into this concept, into this metrics. Because when people were coming to StatsBot, they were bringing their own data sets, right? And the big question was, how do we tell the system what is active opportunities for that specific users? How we kind of, you know, like provide that context, how we do the training? So that's why we came up with the idea of building the semantic layer. So people can actually define their metrics and then kind of use them as a StatsBot. So that's how we build a cube. But, um, at so…

AI assessment note: “We, we built a Cube at StatsBot because we needed it.”

Q at least with the, with these models, they have some knowledge of their own and they can also tell you if your data is bad, you know, which I think is like something that before you didn't, you didn't have any cool apps that you've seen being built on, on cube, like any kind of like AI native things that People should, should think about new experiences, anything like that?

A Well, I see a lot of Slack bots. So, you know, like, it's just, uh, it's definitely like, uh, they all remind me of StatsBot, but I know, like, I, I, I played with a few of them, they're much, much better than StatsBot, so I feel like, it just, it feels like it's on the surface, right? It's just that use case that you really want, you know, think about your data engineer in your company, like, everyone is like, and you're asking, hey, can you pull that data for me? And you would be like, Can I build a bot to replace myself? You know, like, so they will ping that bot instead. So it's like, that's why a lot of people doing this. So I think it's the first use case that actually people are playing with. But I think inside that use case, people get creative. So I see bots that can actually have a dialogue with you. So, you know, like you would come to that bot and say, hey, show me metrics. And the bot would be like, what kind of metrics? What do you want to look at? It's like, you want to be like active users. And then it would be like, how do you define active users? You want to see active users by You know, like, sort of cohort. You want to see active users kind of changing behavior over time. Like, a lot of, like, follow-up questions. So it tries to, uh, sort of, you know, like, understand what exactly you want. Because a lot of people, and that's how many data analysts work, ri…

AI assessment note: “Well, I see a lot of Slack bots.”

Q you've seen the ChatGPT code interpreter, which is renamed as like advanced data analysis. Um, so what's kind of like the, the future of like the semantic layer in AI, you know, what are like some of the use cases that you're seeing and What do you think it's a good strategy to make it easier to do now the text to SQL you wanted to do seven years ago?

A Yeah. So, I mean, you know, when it started to happen, I was just like, oh my God, people are now building StatsBot with Q. They just have a better technology for, you know, like natural language. So it kind of, it made sense to me, you know, like from the first moment I saw it. So I think it's something that, you know, like, uh, uh, happening right now and that's Chatbot is one of the use cases. I think, you know, like, if you try to generalize it, the use case would be how do we use, ah, structured or tabular data with, you know, like, AI models, right? Like, how do we turn the data and give the context to the data, and then bring it to the model, and then model can, you know, like, give you answers, make a questions, do whatever you want. But the question is, like, how we go from just the data in your data warehouse, database, whatever, which is usually just a tabular data, right, like in a SQL Base warehouses to some sort of, you know, like a context that system can do. And if you're building this application, you have to do it. It's like no way you can get away around not doing this. You, you either map it manually or you come up with some framework or something else. So our take of that and my take is that semantic layer is just really good place for this context to live because you need to give this context to the humans. You need to give that context to the AI system an…

AI assessment note: “semantic layer is just really good place for this context to live”

Q it comes to like agents, you know, so there's like A lot of people trying to build agents for reporting, building agents for sales. Like if you're building a sales agent, you need to know everything about the purchasing history of the customer, all of these, all of these things. Um, yeah. Any thoughts there? What should all the AI engineers listening think about when implementing data into, into agents?

A Yeah, I think kind of, you know, like trying to solve for two problems. One is how to make sure that agents or LLM model, right, has enough context about, you know, like a tabular data. And also, you know, like how do we deliver updates to the context, which is also important because data is changing, right? So every time we change something upstream, we need to sure we update that context in our vector database or something. And how do you make sure that the queries are correct? You know, I think it's obviously a big pain in this all, you know, like AI kind of, you know, like a space right now, how do we make sure that we don't, you know, provide our own counselors. But I think, you know, like kind of be able to reduce the room for error as much as possible that what I would look for, you know, like to try to like minimize, uh, potential damage. Um, and then, yeah, I felt like our use case, you know, like for cube, it's been, We've been using, Cuban used a lot to power sort of customer-facing analytics, so I don't think that much is going to change, is that I feel like, again, more and more products will adopt natural language interfaces as sort of a part of that product as well, so we would be able to power this business to not only, you know, like charts, visuals, but also some sort of, you know, like summaries. You know, like, probably in the future, you're going to open th…

AI assessment note: “One is how to make sure that agents or LLM model, right, has enough context”

Q that, and how hard it actually is to, to break down some of the monopoly. What else should people keep in mind, Artem? You're kind of like at the cutting edge of this, you know, if I'm looking to build a data-driven AI application, I'm trying to build data into my AI workflows, any mistakes people should avoid, any tips on the best stack to use, what tools to use?

A I would just recommend going through to warehouse as soon as possible. I think a lot of people feel that MySQL can be a warehouse, which can be maybe on like a lower scale, but you know, like, definitely not from a performance perspective. So just kind of having, starting with a good warehouse, a query engine, lake house, that's probably like something I would recommend starting from a day zero, and there are like ways to do it, and very cheap with open source technologies too, especially in the lake house architecture. I think, you know, I'm biased, obviously, but using a semantic layer, preferably cube, and, ah, for, you know, like a context, and other than that, I just, like, I feel it's a very interesting space, you know, like, in terms of AI ecosystem, I see a lot of people using a link chain right now, which is great, you know, like, and we build an integration, but I'm sure the space will continue to evolve, and, you know, like, we'll see a lot of, like, interesting tools, and maybe, you know, like, some tools would be a better job. Fit for a job. I know, I, I'm not aware of any right now, but it's always interesting to see how it evolves. Also, it's a little unclear, you know, like how all the infrastructure around actually developing, testing, documenting, all that stuff will kind of evolve too. But, ah, yeah, again, and just like really interesting to see and observe,…

AI assessment note: “I would just recommend going through to warehouse as soon as possible.”

Q whole, the whole wave of empowering data professionals. I think now a big part of the wave is removing the need for data professionals to always be in the loop and having non-technical folks do more of the work. Are you seeing that as a big push too with these models, like allowing everybody to interact with the data? Um, yeah. Any customer stories you can share? Anything like that?

A I think it's a multidimensional question. That's an example of, you know, like where you have a lot of, a lot of inside the question. Uh, so, uh, in terms of examples, I think a lot of people building different, you know, like agents or chatbots. We have a company that built as internal Slack bot that sort of answers questions, you know, like based on the data in a warehouse. And then like a lot of people kind of go in and like ask that chatbot this question. Is it a, like, a real big use case? Maybe. Is it a still, like, a toy pet project? Maybe two right now. I think it's really hard to tell them apart at this point because there is a lot of, like, a hype and, and, you know, just people building LLM stuff because it's cool, and everyone wants to build something, you know, like, kind of, even at least a pet project. So that's what happening with our community as well. We see a lot of, like, people building a lot of cool stuff, and it probably will take some time for that stuff to mature. And kind of to see, like, what are real, the best use cases. But I think what I saw so far, one use case was building this chatbot, and we have even one company that are building it as a service. So they essentially connect into cube semantic layer, and then offering their, like, chatbot, so you can do it in a web, in a Slack, so it can, you know, like, answer questions based on data in your s…

AI assessment note: “We have a company that built as internal Slack bot that sort of answers questions”