00:55:00 · Rajkumar Sen
[00:55:00] S1: Yeah, and then there is like on a formatting side, LanceDB and others, but there are also like on the optimization thing, uh, which is what I think you know um Rowan's company is doing, and hopefully, you know, we'll get a chance to be part of partnering together. Um there are how big this will go, I don't know, but there's still a lot of innovation left. There is also, I believe, a lot of data management related problems on inferencing. Now, will they be performance features within inferencing companies or some inferencing early employees at inferencing companies will come out and start a product? It's like how Confluent came out of LinkedIn, right? They saw the pattern they came out now. Uh that will also be there. And it's no surprise that you know I'm I'm seeing a lot of data bricks and uh anthropic folks joining inferencing companies now with data management background. So I'm seeing the trend also. Um, and yeah, I think there's event management, even you know, uh, somebody, something like Kafka. I know Redpanda is trying to pivot to some uh agentic. So the all of this will be there. Now, do they go inside the database as part of ClickHouse, Databricks, or they're outside to means to be seen? But there is going to be a lot of innovation uh just on data infrastructure for AI. Um, but couple of things that I but I've my personal uh uh interest right now is uh trying to basically applied AI. As much as I'm using AI, like AI is doing very good for coding, yeah, AI is doing reasonably doing good, like I'm a big fan of Glean inside database. I've used it. I think uh you can point AI to you know models to a set of documents and you can ask questions. Uh I always wondered like why Google actually didn't create the question-answer thing in their search. Why did they stop at search? But you know, eventually it's there. Um I was listening to um uh recorded version of uh Nvidia's keynote, Jensen Huang. I think he brought an interesting topic that you know uh whatever uh you know innovation has happened on the frontier model uh big models on internet data. Yeah, a lot of this innovation has to also transform to business data.
00:57:32 · Rajkumar Sen
Now the challenge with business data, as I see it, is not open, like you know, you can scrape the whole web and train medical, this, that, but uh you know Tesla is not going to share their internal business data with anyone, right? How do you how do you unlock that potential? And I think Ali talked about this in the Databricks Summit also. Uh, I think Databricks created a geniontology and all of it. So a lot of this is getting recognized that you know um structured data, but also like old school structured data. Uh there's a lot to be you know unlocked there. Uh you can keep it as is and still use it for old school BI. But what happens to that data? But one of the big challenges with that data is that you know, um how do you interpret the data? How do you ascribe meaning to that data? It's not very straightforward. The schemas are, you know, I think Atlan is doing some work in that area, but you know, schemas are undocumented or documented in a file, it's in somebody's head. Uh how do you transform the data? It's in some ATL logic. Now, ETL logic can reside in the source, yeah. Oracle, SAP Informatica, or in on on the target in the lake house, right? In terms of you know, like dbt and others. So part of the data is getting changed in the source, part of the data is like, you know, and then most of the meanings are actually federated or tribal, you know. Uh if you think of um any ERP, right? And you you worked on ERP, right? CRM, ERP, everybody has custom ERP implementation. Now, you can't trade a model on everybody, all Fortune 1000 is custom, right? And each company has multiple. So that's I think fascinated me a lot when I was doing my research. Uh because I was trying to get the cloud to reason about my own private data, which was kind of like similar. Like, you know, there's a lot of like um maybe AGI will come and make it redundant. But I feel like um I'm an old school, like, you know, um Oracle SAP guy, so I know that there's a lot of business data trapped there. So somebody will want to unlock that data. But to unlock that data not for BI, right?
00:59:59 · Siddhartha Ahluwalia, Rajkumar Sen, S2
[00:59:59] S0: BI is solved almost. I wouldn't say it's solved. [01:00:03] S1: I think more BI still needs agentic, like text to SQList. The interfaces to BI will change has changed. Like agentic interfacing. BI is a different problem, you know. [01:00:13] S0: Um there is overlap between but when I think of business data, uh to most of the the use cases that I think of unlocking, people thinking only in terms of BI or exactly, exactly. [01:00:26] S1: Which is where I think you know there's some thought leadership. Some I'll give an example to you. Actually, I was talking to a friend of mine who works at an oil and gas company. [01:00:34] S2: Yeah.
01:00:34 · Rajkumar Sen
[01:00:34] S1: Um so oil and gas company today, they just forecast revenue, and the BI use case is you know, some uh, you know, all these rigs are producing this man amount of oil, this is the metric, and what is the forecast for next quarter? But think of an interesting problem that uh you have a rig in Texas, right? And one of your um you know equipment is malfunctioning. Can you tell today at the what malfunction is happening and downstream what is the revenue impact as of today? For that, you need the rig information in some unstructured form, use some you know SCADA historian data, put it on Snowflake data breaks, then figure out which set of suppliers or you know who are depending on this rig. That data is in some Oracle database or some supply chain, like or SF. If you can ask a question that, hey, you know, this rig is 25% of its capacity today, who are getting impacted? If that company can figure out right away that these are the five people who have some, you know, my vendors who are getting impacted, they would want to notify them in advance that this is what you're going to see is like I mean, bad example is the war going on, right? If we had known that state of homes, the problem so these things are important now that uh this is pure like you need to reason about and join things across systems and you know, trying to do a little bit of that reasoning that you know what if at 25%, what is the impact at 25%, and then the nature of the defect, right? There is an English language destitutional defect, what is the impact on you know downstream refining and all of it? Not a B pure B, I think there is some calculation involved, uh, but it's like more like you know, uh, and there are use cases in transportation, logistics. Uh there is a I was looking at uh a company uh where some of these things are mentioned. So I I I feel like there is a lot beyond BI. Yeah, but BI has been there, a lot of companies are trying to table you. But text text to SQL plus BI is a very interesting use case using agents because agents by default will hallucinate. So there's a lot of innovation happening where folks are saying, you know, we will give deterministic answers on every run, you're gonna get the same answers. Very big problem to solve.
01:03:01 · Rajkumar Sen, Siddhartha Ahluwalia
But the other angle to it is like beyond that, people would also want to know if something has happened to your operational things, what is the revenue impact, who do we want to get to? And the other part is like pure, you know, how do you take AI to non-technical people? Sure. You know, the analysts, procurement people, supply chain people. They want to get answered to a question. Right now, they'll be like opening a ticket. Hey, give me this information. Somebody else will answer that ticket. After two days, somebody will answer the ticket. Now, based on the answer of the ticket, they will how do you identify that? For that, you have to bring some reasoning and you know, so there is a fair bit of it, and again, I I mean I'm just looking at the potential of a technology. Yeah, if uh you know the models today can answer questions on documents that I have written and my friends have written, I'm pretty sure they can reason about like business questions, and this is a whole new area, you know. Like I think Jensen referred to it, like you know, nobody knows the potential. [01:04:11] S0: Yeah, so what I think probably as you are mentioning, there could be 10x use cases besides BI on business data.
01:04:11 · Siddhartha Ahluwalia, Rajkumar Sen
[01:04:11] S0: Yeah, so what I think probably as you are mentioning, there could be 10x use cases besides BI on business data. [01:04:18] S1: Or there could be nothing, but uh, we have to go with hope, right? That there will be 10x. Uh and I think the next uh frontier for models is uh structured business data. I think SAP acquired a company uh spun out of ETH Zurich. Uh I think a lot of the AI labs will also start like building maybe exclusive models to train more business data. I know Databricks is definitely doing it. They published a paper recently. Uh obviously, right? If you think of anything new, that company has to be there at the forefront. Uh, but there will be opportunity for others also. And there will be niche, you know, business data is source. There will be finance has their own. Uh healthcare will have their own. Uh procurement supply chain will have their own. ERP is a big on their own. Then on the healthcare medical stuff, Epic has their own. So a lot of this niche, there will be a broad thing, there will be a niche thing. But I think it's a very we are at a very interesting times where um there is a new technology, and then we have to find like where the new technology can unleash new applications, you know, new products. It's a belief, you know, we have to believe, right? [01:05:28] S0: Um, thank you, Samad. I think we can go on and on.