Kinetica CEO Nima Negahban on Data Management and Generative AI

foreign speaks we're talking about the role of generative Ai and analyzing and managing data we'll look at some of the challenges involved and also discuss ways to address those challenges and get the most from generative AI to discuss that I'm joined by Newman de gobin co-founder CEO of kinetica Neiman very good to have you with us today all right thanks for having me I think that companies are are really grappling with the idea of of getting more from generative AI specifically managing analyzing their data that's a that's a real dream it feels like it's still ahead of the curve what do you see I mean you're down there in the trenches so to speak I mean as companies try to you know get the most generative AI in terms of analyzing the data managing their data what do you see going on these days yeah I mean uh it's it's really early days just like you said but you know certainly there's um you know a lot of interest and and being able to leverage generative AI around um analyzing it in ways where where it lives but also you know being able to leverage you know generative AI to um really make the developer make the analysts more more productive right and that's really the first phase of what we're seeing is kind of that co-pilot um that co-pilot use case and we're certainly seeing a lot of innovation there to be able to really understand all the data in the ecosystem and then to be able to rapidly generate analytics queries or code um and be able to make developers and analysts you know uh way more productive and that's really what our Focus has been well as you talk to Executives I'm assuming you probably talk with a number of Executives that do sense that there's a is a real eagerness to will open our Wallace now we need to dive into this or is it more like well we need to do some research and we're thinking about spending money on it what is a market feeling it's being you're seeing both I mean um I mean certainly um there is definitely a lot of interest you know and a lot of you know response to the you know kind of you know this kind of been like a you know generational uh you know movement that's occurred right it's it's definitely one of those inflection points and so there's a you know you know there's a lot more open-mindedness in the industry right now to being able to see how to leverage this right so um you know there's there's kind of you know how do I get it to be able to generate you know to really optimize and make you know what we're doing you know more efficient make our developers more efficient our analysts more efficient but also how do we make you know newer products better products you know to leverage this as well so it's it's going in different directions I think right now the the real uh Focus areas how do we leverage this to make everyone more productive right and you know I think there there is a decent amount of kind of uh you know concrete use cases and research already done that you can point to that says yeah this is you know there are some patterns here that you can follow to that you know you know is going to work for you know what your Enterprise is doing right and then there's kind of the more more elaborate stuff that is still kind of Futures right where you know generative might be just doing things wholesale and that is still where there's still kind of a lot of hype and hand waving I think that there are you know clear specific use cases companies aren't just thinking well we need to get on this bandwagon there's actually a reason for us to get on this bandwagon can you point to say one maybe two uh actually use cases where where generative is helping people with their data yeah I mean uh certainly like you know you know the things we're working on Connecticut I mean to be able to understand all the metadata in the ecosystem and then to be able to ask you know a generative model to you know generate code to write a query for you you know from natural language to you know to that SQL uh code you know that's something that um there's a clear use case there you know and clear benefit where uh you know you can have a chat bot that can really take any type of question you know scan all the data that that your Enterprise might be collecting and give you an answer mm-hmm the executive said to you hey you know we're struggling with this anything you recommend you know any any which either best case or your advice you thought thought you might offer to companies about using generative Ai and data what should they be aware of what might help them I think it's just you know uh start backwards right like really understand what is your goal right like because like you know we started this conversation you know a lot of people want to be in it because they think they should be right and not really knowing what is their destination right so you know try to understand what is your end goal use case that you're trying to solve for and then work it back and you know that's really um when I speak to folks who want to understand how do I leverage generator for my data you know I often begin with what are you looking to do right um and you know often there's not a clear answer there and so you know we kind of help help you know bring that out um but you know at the end of the day um this is early days right and there's just going to be a lot of Discovery uh still to happen all right well let's drill down into specifically what is Connecticut doing how is Connecticut addressing the data need of its clients in terms of generative these days yeah so what we do is very specific we'll actually scan you know your your Enterprises data Lake right and pull out all the data necessary um you know if you want pull the data as well and then essentially allow you to talk to your data ask a question you know type out a question um we'll generate the necessary uh code SQL and be able to run that in the background for you either inside of our database or where it lives in another database and then present that to you and then allow you to visualize it in a number of different ways aha okay so you're you're really turning you're turning all that metadata into a scannable scannable repository that we really enable gender today if I'm understanding that right yeah so essentially what we do is you know what we do is we scan all that metadata put it in a form that our our you know uh large language model understands and then uh add it add your question to to the context right and so um essentially you get a conversational experience right where you can ask it anything right it's going to you know understand all of the data assets in your ecosystem and then generate the right code to give you the final answer you know the idea the large language model I think is fascinating and it's a lot of talk about that sort of almost underneath the the headlines you should you mentioned mention of you said our our large language model let me clarify it is specifically Connecticut's I'm assuming companies of course making their own proprietary in many cases explain me a little bit about that yeah I mean you know there there is you know a lot of foundational models and then there's a lot of fine-tuning done on top of those foundational models um and then when you go um even to you know to a customer you do again another set of fine-tuning or you do prompt tuning so there's kind of different types of um best practices but also kind of tuning uh Frameworks that are that are arising so um you know when you go to an Enterprise you know traditionally what we'll do is you know we'll start and work with a customer see how our accuracy is um if needed you know we can do prompt tuning and if we really need to because Enterprises have their own terminology sometimes and um you know some some of that needs to be tuned into you know fine-tuned into the model um whether that's through prompts or through an actual fine tuning do you think that companies are you just this is a general question about llms and how it affects the market you know the companies will in general need to create their own proprietary llms going forward no I think you know most Enterprises are going to be fine with just you know fine-tuning I mean fine-tuning is kind of the last step even before that you can do context or prompts tuning right which is kind of an easier method um and then find you know if you think about large language models it's like a a web of Weights you know fine-tuning you know is the kind of the last step where it's you can actually adjust those weights so um yeah I think most Enterprises are going to be able to get pretty far with you know what's going to be in a lot of products out there um to to be you know automated I think the products that you know we're building and other data infrastructure companies they're building are built around the notion of just making it really simple for Enterprise to do that type of you know of you know extra training or fine-tuning uh through their products and one thing you said you mentioned you're scanning the metadata just my own education so it could be metadata about structured or unstructured data either one for us right now we're focused on structured um you know for unstructured we're actually building stuff as well but um you know that that's not yet out um but you know there certainly is um you know as far as metadata and unstructured there's certainly the easy stuff and you know that stuff you know we're working on but you know I think there's going to be also a lot of a lot of innovation on just how you scan unstructured data and pull out you know not necessarily pre you know you know like uh you know pre uh or deterministically decided or you know human input metadata but actually like uh auto-generated metadata from unstructured but you know right I think for us we're just fully focused on structured and pulling out the metadata that way is I assuming the the unstructured offers a thornier problem and maybe long term there's the greatest amount of growth there or not necessarily um unstructured I mean so you know with unstructured you know there's a lot of focus there right now you know with generative and Vector search right and and you know you know taking images or text blobs and or audio and doing you know a vector uh embedding from that and doing Vector search um but you know there's there's going to be a lot of growth on you know for on the structured and unstructured side as far as like under you know being able to extract metadata um besides vectors right and give that uh to the LM well I think the big question is the future uh what what do you see in terms of the future relationship between data and genitive AI it's a vast topic I'm sure you probably spent four hours talking about it but if you look say you know two three years in the future what kind of things do you see going on I think there's a lot of so I think there's like um one thing is around standardization right and um how are large language models going to speak to other knowledge knowledge bases right you know right now there are kind of plugins or plug-in Frameworks like like Lang chain that you know is um is really picking up but you know there's going to be a standardization period around how you know large language models are going to interact with Enterprise and that I think is gonna um be happening here in the next you know one to two years as these common patterns emerge right um because uh I think large language models and you know large language model agents are going to be kind of these you know always running entities that are doing you know what's you know I think is what's being labeled as Chain of Thought workflows where you know they're actually taking multi-step processes and so that they're going to be able you know be able to do more complex workflows uh and in order to accomplish that they're going to need to be able to interact with systems and services on their own just all right for my education there was a phrase you used a system of thought was that it Chain of Thought chain of thing what what is from information what is what is a Chain of Thought workflow Chain of Thought is you know essentially where um you know you're asking it to do you know an entire workflow and it almost feeds back onto itself so it'll do one step get an answer feed that back into itself and then you know or to another agent and essentially um if you there's a lot of uh projects out there but I think the most popular one is auto gbt where you can actually get it to do a pretty large multi-step workflow by kind of giving it a you know a task and a goal and it you know it's able to actually start to iterate and do all the necessary steps in between so it really does mimic human functionality at a pretty profound level you know theoretically you know when you actually start to get in there you can see that it's early days but yeah like the the bones are there which is the exciting thing that is exciting yeah uh Nema I think you said it it's a lot of fascinating stuff it's going to be a fascinating sector to follow in the years ahead yeah thank you so much for sharing your Insight and please come back and talk with us again sometime thank you so much thanks for having me

This transcript was generated automatically from the video's captions and may contain errors.