CDOIQ Symposium 2024 | Anthony Deighton | Tamr

theCUBE.net

Clips
More from CDOIQ Symposium 2024

Anthony Deighton

CEO

Tamr

play_circle_outline Tamr: MIT-Inspired AI Data Management Solutions for Simplified Cloud-Based Entity Resolution in Enterprises

play_circle_outline The need for a combination of AI and human expertise in complex analysis tasks.

play_circle_outline Challenges in traditional MDM solutions and Tamr's approach using machine learning.

play_circle_outline Maximizing Data Efficiency: Virtual CDOs and RAG Technology's Role in Optimizing Human Capital

Info
Transcript

Anthony Deighton | Tamr

Anthony Deighton

CEO Tamr

Anthony Dayton, CEO of Tamr, spoke at a conference in Cambridge, Massachusetts about Tamr's AI research origins and focus on improving enterprise data quality. AI can efficiently compare millions of data records, addressing scale issues in data management. Tamr's SaaS solution automates entity resolution to publish the best data versions. AI, machine learning, and large language models clean and curate data. The discussion also mentioned the role of a chief data officer in managing data pipelines and how AI can disrupt some tasks. Tamr's virtual chief data of... Read more

explore Keep Exploring

What is the focus of Tamr and how does it aim to help organizations improve the quality of their enterprise data? add

What problem is Tamr focused on solving in regards to enterprise data quality and why is it still an issue despite the existing tools in the BI industry? add

What is the fundamental issue with solving data problems? add

What kind of work are Chief Data Officers (CDOs) currently doing that could potentially be taken over by AI technology in the future? add

bolt Powered by CUBE AI

Anthony Deighton | Tamr

search

>> Hi everybody. Welcome back to Cambridge, Massachusetts. We're here at CDOIQ. This is theCUBE. I'm Dave Vallante with my cohost Sanjeev Mohan. Paul Gillin will be here tomorrow as well. Anthony Dayton is here, he's the CEO of Tamr. Anthony, welcome.

>> Great to be here. Nice to see you again.

>> Yeah, ditto. We newly minted relatively started in January. We first talked to Tamr when Michael Stonebreaker came on at this very event, which was at the Tang Center at the time, in late last decade. Give us the update on Tamr, what's happening there?

>> Sure, and the fact that you brought up Mike is a great point. So Tamr comes out of academic research from literally around the corner here at MIT, Mike, the academic co-founder. In a way, the idea behind Tamr is very, very simple, which is what if we could train AI to do common data management tasks, things we're good at as people. If I show you two customer records and I say, "Are these the same?" As a human being, you're like, "Yeah, those are the same," or "No, those are different." If I show you a phone number and I say, "Does this look like a valid phone number to you?" As a human we're very good at saying, "Yeah, that looks valid," or "No, that doesn't look valid." What if we could train a machine to do those tasks? What would that open up? How would that allow us to improve the quality of our enterprise data? And the one thing that we can be a hundred percent confident, the one theme that I see repeated constantly and with certainty at this conference is my enterprise data is a post-apocalyptic dumpster fire, and I need to get a handle on it. And I got great tooling for visualization. I have great tooling for accessing the data, but the data itself is terrible. So Tamr is focused on the problem of helping organizations improve the quality of their enterprise data by using AI to identify the entities that exist in those data and allowing them to publish the best versions of that data at any given time.

>> So why is that the case? I mean, the BI industry was created ostensibly to solve this problem or maybe to create a veneer, false veneer on top of the problem. It's about a $30 billion business. We've got beautiful dashboards and visualizations. We have these hyper specialized data pipelines that serve the business. Why is data still such a mess?

>> Because fundamentally data is a scale problem. And by the way, I couldn't agree with you more. The BI and analytics companies have solved an extraordinarily important and frankly difficult problem, which is how do I give people easy to use access to that data? But I hear this consistently from customers, "I love my dashboards. I don't believe the data that's in them." And so to your point, why is it that we haven't solved the data problem? It's because it's a scale problem. So if I had a simple example, if I had 10 million customer records and I wanted to understand if there are any duplicates in that customer record, I have to take the first record and I have to compare it with the 9,999,999 other records that it could be. And if you do that work, you're going to have around 50 trillion comparisons that you have to do. It's what we call in computer science an N-squared problem. And so it's an overwhelming problem, and as a result, you don't do it. But AI hacks the scale problem. So now we can actually ask the machine to go do that work on our behalf. It doesn't sleep, it doesn't eat, it works weekends. It's highly parallelizable, and we can actually attack this problem for the first time ever, arguably, we can actually get a handle on this problem.

>> That's interesting. I was just looking, I mentioned Mike Stonebreaker was on late last decade, and one of the things he said, this is obviously well before the AI awakening or the AI heard around the world. He said, "The second thing I'd like you to know, and really stress is machine learning is going to be a game changer for essentially everybody. And not only is it going to be autonomous vehicles, it's going to be automatic checkout, drone delivery. And so you can basically affect everybody and categorically, any job is going to get automated." And that was late last decade. Pretty prescient-

>> That's why he is prescient.

>> Yeah, correct. Yes.

>> And it was clear. And we also talked about master data management, and I had mentioned, don't hate me for saying this, but the database that you basically invented, the data warehouses and the big data, never lived up to the promise master data management as well. Lot of promises made by the technology industry. And his point was, "We're just going to solve that." How are you solving that and why is it going to be different this time?

>> Well, so first of all, go back to the premise of the question. The MDM industry has failed. I've never met anyone that put their hand up and said, Yeah, I just got staffed onto an MDM project. Super excited," right? No, that doesn't...

>> Yeah, it's a very high failure rate MDM projects.

>> And that's fundamentally because we've asked people to go and take on this task, and as we said, it's a scale problem. So we're asking people to do boring work at scale. Well, shocker, they don't like doing that. The opportunity and why is it different this time? Well, to the point you brought up, Mike and the academic team at MIT, they've invented machine learning. They can now take this problem and turn it over largely to the computer. So if we can take a corpus of data and we can have 95, 99, 99.9% of the data cleaned, curated, brought together, the entities identified, so we know what companies exist in that data, what customers exist in that data, what suppliers exist in that. We can do that automatically. If the machine can do that, then the edge cases that we leave on, that aren't perfect, that we turn those over to people, they're happy to take that on. That's interesting work. Let's go figure out where these weird edge cases are, but we need to take 95, 99% of that work and have the machine do it on your behalf.

>> So I have a question earlier on, you were talking about data quality is an issue. And so what is the connection between data quality, MDM, entity resolution, are these all related?

>> They're all related, and in a way they're all answering the same question, which is, what's the best version of this type of data that I have inside my enterprise? And I'm using this term type a little bit in the colloquial. So is it, what is the best customer data? What is the best supplier data? What's the best parts data? What's the best product data? So we all have these common ideas inside our company about the data that matters to me as a business. And the question is how do we identify those, the term we would use is entities or the type of data, identify those types of data, and then within that domain, what is the best version of any given record across every source and internal sources, external sources, structured sources, unstructured sources? I want all of that information brought together and the only way we're going to solve this problem, the only way we're going to solve this problem is with AI. That is not a human scalable problem. That's what we talked about before.

>> What kind of AI? Are we talking about the legacy AI, machine learning that Michael Stonebreaker talked about, are we talking about LLMs and generative AI, both? How are they being applied? What's your perspective on that?

>> Sure. So first of all, just some definitions. AI is a very general topic, which is as the acronym is, it's artificial intelligence. It's really the application of algorithms to any sort of task. Machine learning is a specific instance of AI and large language models are a specific instance of AI. These things work very well together. What large language models are so good at is either generating but also consuming text, and that can be very valuable in the context of some of the AI algorithms that we've developed at Tamr. So for example, if I give you a description of a part, which is a long string of text, and I want to extract from that some piece of information. So I give you a description of the part and it says, "This is a seven millimeter ball bearing." And I ask, "Well, what's the size of the ball bearing?" It'll say seven millimeters. It's good at that sort of task. Also in reverse, if I hand it a structured adjacent object of the customer data and I say, "Can you summarize everything we know about this customer?"-

>> Or write an email.

>> Or, "Write an email." These are tasks that large language models are good at. On the other hand, large language models, and Mike will talk about this on Friday at this conference, but large language models are not good at resolving entities. There's a specific domain that we can, a specific type of technology that's really good at that, but all of these things work well together, and they all fit under this moniker of AI.

>> Okay. And coming back to what you're doing, you're applying all these different elements of AI to scale the MDM problem.

>> Yes, to slay the MDM beast finally.

>> Not scale the problem, but-

>> Yes, finally.

>> To scale the success, the outcome, a successful outcome. Go ahead, please.

>> I've followed Tamr for many years as an analyst, and I've seen that every few years Tamr would change its value proposition. So what is now the current state and how is Tamr now getting a product market fit?

>> Sure. So really focused on this problem of entity resolution at scale to solve the MDM problem, entirely a SaaS solution. So you don't need to install anything or configure anything, simply use Tamr Cloud, point it at the sources of data that you have, and the system will identify the key entities that matter and the data within those entities that are relevant to your business. So yeah, in a way, really making it simple. And finally being able to get a handle on this MDM problem.

>> Irrespective of the physical location of the data, is that right?

>> Correct. Yeah.

>> But do your customers have an issue, is the data being moved or is it staying within my account?

>> Yeah, so it's very much, it is very important that it stays within your account and it's your data. The question of how we resolve those entities, clearly we need to apply the model to the data, and we can't expect source systems to do that work. It would put far too much load on those sources. And plus the types of calculations that we're doing are fairly complicated-

>> Compute intensive.

>> Yeah, compute intensive. So absolutely that we want all of that processing to exist in our cloud, and we want it to be encrypted and secure and isolated, et cetera. So those are all important questions, but fundamentally, it comes back to outcomes. If we can take tens, hundreds, thousands of sources of data and show you your customers, your suppliers, your parts, your products, these key entities that matter to your business and deliver those to your business decision makers, maybe through a dashboard that allows beautiful visualization, but finally a dashboard that you actually trust the answers in, that's the holy grail. That's what we can achieve. Finally.

>> What are you talking about at this conference? There's this term virtual chief data officer. What does that mean? How does it relate to the chief AI officer role that's emerging?

>> Yeah, so the question I asked, and that presented on at CDOIQ, I wanted to sort of explore this question of whether AI in a way disrupts the chief data officer's job. And in some ways, this comes back to I think a fundamental concern or anxiety that maybe everybody has, which is, is AI coming from my job. And given this was a CDO conference, I asked the question, "Well, is AI coming for the CDO's job?" And in a good way, the answer is, well, yes and no.

>> It depends. Almost every segment we are recording today, the answer is it depends.

>> It depends. So what does it depend on? So a lot of the work that we see CDOs doing are really answering basic questions. How many employees do we have? What's this customer doing? Can you tell me where this supplier is located? These are basic questions. And frankly, having a CDO team or a CDO themselves answering those questions is a waste of human capital. Those are questions the AI should be able to answer. And using Tamr and our virtual chief data officer capability, the term of art here is retrieval augmented generation or RAG. We can pass to the LLM, the entity data that's relevant to the question that's being asked, and we can give people a natural interface to be able to ask and answer those questions. That we think is extremely exciting and it's a way of making the work that Tamr does of identifying these entities relevant to users. And we think takes this boring work that CDOs have asked to do, it takes it off the table and allows them to focus on the strategy of the business.

>> It would seem to me that what gets disruptive, I'd love both of your input on this, the data pipelines today are highly complex. They're comprised of extremely specialized individuals, data engineers, quality engineers, data scientists, business analysts that rely on each other and in many cases are very linear, you can't do anything until the new data gets in and it's clean and so forth. A lot of heavy lifting. I would think AI disrupts that quite dramatically, but-

>> Yeah, so my point of view is that more than 50% of CDO's job, daily function should be in orchestrating all these different teams, shepherding what they're doing, building consensus, getting buy-in, understanding where the problems exist, fix and directing LLMs towards that. If 50% of a CDO's job is people related, that's not going to be disrupted by LLM, but the other 40% or 30%, which is the boring heavy lifting of go run this algorithm and find-

>> And can you share this, extract this data for me and share.

>> Yes, that should be disrupted by LLM, but not the people-

>> Okay. All the boring stuff. So Anthony, you gave some examples of pretty straightforward examples that the machine should be able to take care of, but if I want to understand how a variety of pricing actions are going to affect my revenue, you're saying you're still going to need the people to do?

>> I would not give that critical job to an LLM today. Maybe a few years from now. I am not going to put my company's revenue at stake.

>> But fast forward a few years.

>> But the key to answering that question, so you're asking the question, what would a change in price do for my gross margins or my revenue? The key to answering that question is the underlying data associative answer in that. If I have all my customer data is a mess or if I'm only looking at customer data from North America because I haven't figured out how to link it to customer data in Europe, I'm answering that question with one hand tied behind my back. And so we believe that the key to seeing the success of any kind of analytical process inside a company and the one you building a model for pricing is a great example, is clean, quality, curated, trusted entity data at the core of that model. Otherwise, you're at risk of hallucinating. And this is a really important point, which is in our personal lives, if we go and interact with ChatGPT or Gemini from Google and it hallucinates an answer, we go, "Ha, ha, ha," we laugh. We tell a funny story. If we're in our company and we rely on one of these models to do an analysis and it hallucinates an answer, it can be catastrophic. We could be pricing incorrectly, we could be giving away a discount we didn't need to give. We could be saying the wrong thing to a customer. So the key to all of this is the quality of our enterprise data and this problem, this quality of data problem is fundamentally a scale problem. There's too much data to take it on manually.

>> So you brought up RAG. We know a little bit about TAG. We developed our own RAG, theCUBE AI, cubeai.com released it earlier this year in sort of a little beta. I have data from our survey from ETR. They just did a survey, 1400 respondents. I was somewhat surprised at this. "Which of the following best describes your organization's plans for incorporating retrieval augmented generation into your organization's Gen AI strategy?" 28% of those 1400 said no plans. 22% currently experimenting, only 4% said they're going to roll out in the next six months and only 3% said they have RAG in production. 23% said they're not sure. I was surprised given-

>> That's very little overall.

>> That it was pretty trivial for us to take. We have good data because when you come to theCUBE, there's not a lot of racial bias. You're not swearing. So it's pretty tight dataset, but it really is a semi trivial request response type of model. First of all, do those numbers surprise you at all? Only 7% plan on rolling this out or have it in production.

>> I'm actually quite surprised. That's a really low number.

>> So I wondered if this is just talking to the wrong people but-

>> And the fact that you are saying that you've got RAG, and for you it's like what's the big deal? You should have drag. I think it also depends upon the kind of data and questions you are asking. So it's the data quality and what's the use case? Your data is mostly videos and text-

>> And transcripts, yeah.

>> Right. So putting an LLM and asking similarity kind of questions and probabilistic is okay, but you were saying there could be catastrophic issues when you are asking data that questions, they should really go to a deterministic model on structured data. So I think we need to understand what is the usage.

>> But Anthony, you were saying you actually use RAG.

>> Sure.

>> And so how do you address that problem?

>> So I think it's because this strategy and these terms frankly are very cutting edge. Tamr is still a relatively small business. We have yet to dominate every single Fortune 500 company. We are thrilled and excited about the opportunity to give everyone in the world the best RAG models on top of their enterprise data. But it's early days and we're going to get there. I think most companies in the world are just experimenting with frankly large language models and seeing what the possibilities are. And that's why you see the data that you-

>> So that doesn't surprise you as much as it does Sanjeev?

>> The other thing I want to mention is that, why do we have RAG, because we are limited by the context windows of the model, and we haven't yet developed the full breadth of technology for AI. Once we develop that, then the value of RAG may actually go down. I don't think it's going to go away, but there will be other opportunities like million context length window for example, may not need RAG. Right now it's too expensive and latency.

>> All right guys, we got to leave it there. Anthony, thanks so much for coming on. Sanjeev, my cohost for today and hopefully tomorrow. Tomorrow as well. We got some other guests coming on as well. So you'll be here. Paul Gillin will be here as well. Keep it right here. My name is Dave Vallante, you watching theCUBE's coverage of CDIOQ 2024 from Cambridge, Massachusetts. We're right back.

>> Thank you.