MongoDB .local NYC 2024 | Sanjeev Mohan | SanjMo & Carl Olofson | IDC

Clips
News
More from MongoDB .local NYC 2024

Sanjeev Mohan

Principal

SanjMo

Carl Olofson

Research Vice President

IDC

Analyst insights: The ‘crucible moment’ of AI integration in database evolution

As artificial intelligence continues its ascent within enterprises, a deep dive into AI integration and database technology reveals groundbreaking dev...

play_circle_outline MongoDB Local event coverage in New York City

play_circle_outline Exclusive Interview with Former Analysts: Keynote Takeaways on AI, Developer Focus, and MongoDB AI Applications Program

play_circle_outline The need for simplification and consolidation in the database industry

play_circle_outline Exploring the Evolving Challenges of AI Technology and the Quest for a Database 'Star Trek Computer'

play_circle_outline The Evolving Landscape: Exploring the Impact of Hyperscalers and Database Vendors on AI Application Development

play_circle_outline Unlocking the Power of Intelligent Data Platforms: A Unified Approach to Structured and Unstructured Data for AI Applications

Info

Sanjeev Mohan | SanjMo & Carl Olofson | IDC

Sanjeev Mohan

Principal SanjMo

Carl Olofson

Research Vice President IDC

search

>> Welcome back to MongoDB Local, everybody. You're watching theCUBE, our live coverage wall to wall. Actually, a full day. This used to be a two-Day event. Now, it's a one-day event. We're here in New York City, the big Apple. It's a great time of year. We saw the Knicks play the other night. They went into overtime. They're now going to go to go invade Philly. The Rangers, they got through, they swept the capitals. You got the Islanders, they got swept. And of course, the most exciting two minutes in sports this weekend. May is such a great month, the Kentucky Derby. But we're here covering MongoDB local, and we're going to do the analyst angle just coming off the keynotes. We're here with Sanjeev Mohan, who's the principal at SanjMo, former Gartner analyst. And Carl Olofson is a principal database analyst, senior DB analyst at IDC, probably a VP if I botched your title, but-

Carl Olofson

>> Research VP.

>> Research VP. Awesome. Which means you don't have to manage a ton of people and do all those performance reviews. You just focus on your research. Gents, welcome back to theCUBE. Great see you both.

Sanjeev Mohan

>> It's great you.

>> All right, Carl, start with you. Same question to you Sanjeev. Carl first. Keynote, what will your takeaways? What'd you think? Let's riff on it.

Carl Olofson

>> Well, I felt that Dev struck all the right tones in his speech. He talked about the things that I expected him to talk about. It was a good talk overview of where the company's going and how the products are developing. And then Sahir got into the details with the various technical improvements that they're going after. I felt that, and this is a common thing MongoDB, they're very much aimed at the developer. So they're talking to the developer, basically saying, This is what we're doing for you.This is all aimed at the developer. So it's not aimed at the high-end strategic planner or the head of the CIO or something like that. They're talking to the development organization.

>> A lot of people, Sanjeev, would say that they need to expand their messaging beyond the developer, start speaking more wallet in order to continue to grow. I don't know if you agree with that, but your take on the keynote.

Sanjeev Mohan

>> So I Liked what Dev said. Two things that really liked, one is this thing that he said Sequoia came up with called the Crucible moment, which is a decision you make today will have an outsized impact down the road that you may not even see. So his whole thing was that AI is that moment. It's really revolutionary, and the fact that most companies don't want Mickey Mouse kind of use cases because they're not transformation, or like chatbot. They want examples of Novo Nordisk reducing their time from 10 days to minutes. That is what... And the other thing I liked was where he said that anytime a new technology comes up, the value is generated at the bottom most layer, which the infrastructure. And we are seeing that with NVIDIAs and the AI model developers, but the real value happens at the top most layer, which the applications. So the fact that they have announced this MongoDB AI applications program is quite a good thing.

>> Yeah, it was interesting. He's somewhat reframing the AI stack. On the earnings call, he talked about the three layers, the infrastructure layer, the trading,-

Sanjeev Mohan

>> Tooling layer....

>> tooling layer, and then the apps layer.

Sanjeev Mohan

>> Correct.

>> And he even said, We don't really play in layer one. We play...But the chart he showed today had Mongo right the middle, which was kind of interesting. The big news here is their MAP program-

Sanjeev Mohan

>> MAP....

>> where they're bringing together the application because-

Sanjeev Mohan

>> Correct....

>> I look at this as a way to bring their customers into that second and third layer. Do you guys see it the same?

Sanjeev Mohan

>> Yeah, because it's a partnership that includes all the major players, all the three hyperscalers, all the integration players like LlamaIndex, model players like Entropic, and then also it includes professional services companies to deliver solutions.

Carl Olofson

>> So it's a tricky thing because you have a bunch of moving parts here. You want to make it sound simple. In fact, he repeated several times, We want to make things simpler.

Carl Olofson

>> Picture is inherently complicated.

>> Yeah. Yeah. I want to ask you Carl about the whole wrap on Vector. So here, it talked about-

Carl Olofson

>> No....

>> and Dave as well sort of referenced it, and Sanjeev, I'm sure you have thoughts as well. And we certainly hear this from a lot of vendors that are integrating Vector into their database, that it's really a feature, not a product. I tend to agree with that, although I see this Tech Crunch article with all these vector search companies getting all this VC money now. And so what's your take as somebody who's watched the database industry since you were in high school? What are your thoughts on... Is there value on standalone vector or would you advise customers? It is probably an it depends answer to really sort of think about integrating.

Carl Olofson

>> Well, it does depend on what problem you're trying to solve. But I do believe that ultimately, Vector technology will be integrated with other AI technologies. It won't be sold as a product in of itself. Okay? But you'll have Vector that's being used for general text AI and that kind of thing where you just ask an English language question and it sweeps through the large language model and pulls together an answer based on your question. And then there's the more specific database centric question or service, whatever it is, that requires for one thing, a more grounded model, because you don't want it to go off in space and look at irrelevant things, and also that it needs to be focused on what is actually there. In other words, when I ask a question about what's in the database, there are two things that have to happen. One is that the response needs to be relevant to my question, and the other is that the answer has to be absolutely positively correct. It can't be approximately correct. can't be vaguely, sort of, yeah, maybe. It has to be absolutely correct because databases are... Data is either right or wrong. There's no in between.

>> And so I infer from that that good enough is not good enough in certain use cases? And that's where you would use the standalone vector database as an example, or not

Carl Olofson

>> necessarily?Well, when you use a standalone vector database, in other words, as I'm looking to gauge... Say I'm asking a question about customer sentiment, how do people think out there in the world about this or that kind of product or concept? It can come back with a textual answer that is close enough to be right that I can use that as information. But if I ask, what are the factors that caused production to increase in the last quarter? And by how much? Those numbers have to be correct. They can't be approximate. And in that case, you really want a vector system that is close to the database that is actually part of the system.

>> How do you think about this, Sanjeev. In security, the recent data from ETR shows the number of security vendor is actually increasing, not decreasing. And when we ask them why, they say, Because we need to fill gaps. We need best of breed.My question, if you take that concept, it apply to databases or are there too many database types? There's graph. There's there's time series, there's vector. Do we need do customers need to simplify, or is it really they need best of breed to fill gaps, drive revenue, drive productivity? What are your thoughts on that?

Sanjeev Mohan

>> It's a very interesting conundrum, because typically I feel the technology sector moves in tandem, what's for the goods is good for the gander, but not in this case. I think security people are very, very particular because it's black and white. But in the database I think you see more consolidation opportunities than you see in security. I don't think anyone wants to have more and more components of the data architecture than what we have today. So the things are different. I don't see... You were right by saying that use cases will determine whether you need a standalone vector search database or not. But more or less, it's better for the vector capability to be closed to my corporate data. And my corporate data is sitting in a document model. It's sitting in relational. It's sitting in grav and spatial. So I think it's much better that they are co-located rather than they're in a separate database, so I'm having to ETL like. Even if there's zero ETL or whatever it really isn't. It's still ETL. It's just automated.

Sanjeev Mohan

>> ETL, a automating and hiding the complexity. You don't... Because every time you go from one stack, one system to another, you've got to deal with some overhead latency security loopholes. So I see there's more consolidation in the database world may not be in security.

>> Yeah. So that should bode well Karl for a company like Mongo. I wonder... IDC gives a lot of thought to markets, of course. How do you think about Mongo's TAM? At the highest level you've got operational database and analytical databases, and folks will... Dave will say, Well, we don't participate in this. We don't compete against Snowflake and Databricks.At the same time, they provide analytics. But how do you think about their TAM? They're roughly 2 billion company, call it. A lot of people think, wow, Mongo's got a huge TAM. They should be growing faster than what they're saying, or they should be a bigger company. What are your thoughts on their market opportunity?

Carl Olofson

>> Yeah, there's no consistent view as to what the future opportunity for document database is. Okay? And you mentioned operational versus analytical, and there's kind of like a water's edge for them with analytical because they're document database. They don't, by nature, combined data in the ways that you need to do analytics, so they have to partner for that, but they can be more clever in the way that they enable people to pull the data that they need and reorganize it into an analytical format, even if they do drop it into a data warehouse or something like that. So there is plenty of opportunity to take them a little bit beyond the space that they've been playing in up to this point.

Sanjeev Mohan

>> So Carl, do you remember, in the last few events, they used to talk about a product called MongoDB Data Lake? They used to talk about data mesh. I don't see that.

Carl Olofson

>> Yeah, they didn't mention anything like that

Sanjeev Mohan

>> here.In fact, last year, a very big announcement, or the year before, was a native SQL driver.

Carl Olofson

>> Right. That's right. In fact, yes. And they didn't mention that either. This seemed like an exciting capability where they could map data into columns and you could execute SQL statements against it. We didn't hear about that this time. It may be that they're finding that they're getting more traction by working cooperatively with people like Databricks and Snowflake than trying to do it themselves.

>> And then what'd you call it? Because there's a SQL, my mind, SQL converter or

Carl Olofson

>> That's a relational thing, migrating from a relational database to a document, David.

>> And when they talk about it, they don't ever mention Oracle, but they're really talking about Oracle. That's the big wheel. Well, maybe everybody is.

Sanjeev Mohan

>> The whole cottage industry around moving from oral repairs

>> Of course. Yeah, and yet Oracle continues to grow. What are your thoughts on RAG, specifically moving beyond RAG? RAG is actually not that hard to do. People are finding, yeah, it's nice, but to the earlier points that were made in the keynote that the ROI isn't really with the chatty stuff,

Sanjeev Mohan

>> Correct.

>> It's really going beyond that. And those examples that you give Novo Nordisk, et cetera, there's a lot of them. When you come to conferences like these. There's dozens. But in the grand scheme of thing, it's pretty small. It's not big enough to move the market today. So how do you see RAG adoption and moving beyond RAG to get higher levels of value? And how long is that going to take?

Sanjeev Mohan

>> One of my favorite topics is exactly what you just said. The topic is, how do I get AI to work on my corporate data? RAG was the first iteration that really took off. And in my opinion, I think RAG, we fixed it. We know how to do RAG. Now, we need to move our thinking beyond RAG. I said, is there something else we can do? For example, before RAG, there a huge topic about prompt engineers, how important prompt engineering is. And I really think prompt engineering is

>> ridiculous.Big snore.

Carl Olofson

>> No, because prompt engineering is basically a new query language. That's what-

Sanjeev Mohan

>> Exactly. And how does one know, what is a good prompt? So what the future is going to be, maybe hopefully later this year, the cost is coming down of inference is some of a combination of fine tuning and drag. So you take a model. It has to be open source because you need to know the model rates. And then you basically train it or you fine tune it with some additional layers, corporate data, and then you do RAG on that, because RAG by itself is getting us to a point, but not beyond. Fine tuning will help a little bit more. Also, in context learning, which is like ChatGPT has introduced this concept of sending a file with my prompt. This is brand new. It came out last week, which means I may have CSV file and I say, Here is my templated SQL. And the response for different questions. Learn from this, so in context learning, and now answer my question. So this is the expanding ecosystem and why MongoDB map is important, because it's not just about vector search, it's about vector search, RAG, fine tuning in context, learning prompting, few short, many short, all of that is needed to be successful.

>> That's a

Sanjeev Mohan

>> curated-Correct....

>> capability that presumably confers competitive advantage to them and takes advantage of all the bedrock

Carl Olofson

>> Except that the technology is still evolving.

>> Yes.

Carl Olofson

>> So because the technology is evolving and is not static, that means that the techniques we know today may not work as well in the future. There'll be new techniques that work better.

Sanjeev Mohan

>> Oh, completely.

>> So are they going through a one-way door potentially, or-it's always the challenge.

>> It has to be managed.

Carl Olofson

>> When you're an early adopter, you're taking that risk, right? Am I going down the right path, or am I going down the wrong path, or is it slightly wrong?

Sanjeev Mohan

>> So you are 100right. The whole AI space is evolving so rapidly. In fact, there's even a discussion these days that have the transformer architecture? Because if you know, transformer came out in 2017 when Google engineers wrote, Attention is All You Need. But is attention really all you need? Or maybe, what's next? There are some new technology. Mamba one of them. So it's going to change.

Carl Olofson

>> Right now in the database world, there are three things that database systems can do. They can help you write a schema, they can help a developer write an application, and they can help an end user find data, generate the SQL or whatever query language is appropriate for the database they're using, whether Cypher or something else, to find the data they need. Even if they're not technical people, those are the three values that they have. But what people really want, and I gave a speech back 20 years ago at Directions and IDC, in which I said, what people really want is the star Trek computer. They want to just be able to ask a question.

Carl Olofson

>> Yeah, and then get an answer. And I said, and we're a long way from that. There you in order for AI systems to actually understand enough about what's in the database to give you that level of capability, you need a lot more metadata. You need a lot more information. That gives it the context because there's no semantics in structured data, none whatsoever. There is a human language, and that's what drives the way these systems work.

>> Because ChatGPT and Perplexity, they have all the data the internet, but it's the that they're lacking. But your prediction is actually... We're kind of there with even search today. Think about how much search has evolved. I think people say, Oh, the early days of dial up.Maybe it is. Yeah, it's good. I to ask you guys this headline that he put, I put this tweet out, it said, internet may be just a passing fad as millions give up. When I crossed out internet and put ai, there's a lot of similarities. I don't want to over rotate on the .com bubble. Every bubble is different. But you remember, you guys remember it Lycos, it was Yahoo, all portals, or CMG, and it was VCs paying companies to advertise on portals and those portals where stocks were going through the roof and then the whole thing blows up and who comes out of it alive and thriving Is Amazon, Google, PayPal, social media years later? Do you see a similar trend happening today where you have all this CapEx?

Carl Olofson

>> Yeah.

>> The hyperscalers They got so much money they know to do with so they buy GPUs like crazy, NVIDIAs. What did Dave say? the greatest quarter in the history of quarters-

Carl Olofson

>> Correct....

>> ever, right? so there's a lot of money changing hands within the insulated echo chamber of the tech world. Do you see that as a similar pattern to ;IDC research has shown in this area that there's a lot of interest out there in the world in AI and in building AI capability. There's a lot of ignorance as to how to go about it, and there's little appetite for spending money on it. So what that tells me is that this is going to be a slowly evolving thing, but expectations will go like this. The actual development process will go like this. People will get frustrated.

Sanjeev Mohan

>> So here's my thinking. All three of us are old enough to have lived actively through the .com.

>> It was awesome.

Sanjeev Mohan

>> It was amazing. please bring it back. We need If you remember, one of the big shocking disaster, there were many of them, but one of was Webvan. in fact, I had a green box for grocery delivery that came out in 1999, and it was an utter disaster failure. But then what happened, quietly, the technology matured.

>> Yes.

Carl Olofson

>> And that's exactly my point. They had a lot of the right ideas, but they didn't have all the supporting technology to make it work.

>> And the economics didn't work. And then when the VC money dried up, they couldn't get the bridge.

Carl Olofson

>> The other thing, by the way that totally changed the picture was smartphones. Because I don't go to my computer to order groceries or to order food.

>> Mobile.

Carl Olofson

>> I go to my smartphone. That's totally changed the-

>> And that came seven years, seven, eight years after the .com bubble. But you're right, Pets.com, now Chewy, is doing great. You got storage networks. Most people never heard of storage networks. That became the cloud. What they were doing was they were taking storage boxes, a little bit of a little bit of Compaq, whatever, just running data centers, hosted data centers. They just didn't have the economics until Amazon came along.

Sanjeev Mohan

>> There was this very famous example just a month or two ago where Air Canada had a fear that they published, and then it was a mistake. And everybody said, Look, GenAI does not work,or it turned out, literally a month later, Air Canada said there was no GenAI.

>> Right.

Sanjeev Mohan

>> So this is the kind of narrative Now, there's also another very famous case where this car dealership, General Motors car dealership put a chatbot up, and some guy went and said, Which is the best truck to buy?And answer was Ford-150, 150. So people said, Look, you should not be doing GenAI.And my thinking is that kudos to the car dealership. They put that ChatGPT API to do this. Now, what they... Instead of...

>> First the sales guy will, say, Oh, it's hallucinating.

Sanjeev Mohan

>> Yeah. But instead of retracting, I hope what this car dealership says, Lesson learned ground the answers in our database.

>> Yes, exactly.

Sanjeev Mohan

>> Then they'll be ahead of the pack. That's the-

Carl Olofson

>> Again, a lot of work involved in making sure that the responses stick to the subject and they're within the range of possible answers. And that is still... A lot of times when you get to the point where you say, Okay, now people have to do some work,then the enthusiasm drives up, right?

>> So do you think, bringing it back to the AI era, do you feel like LLMs get commoditized, that we'll see all the hype around the picks and shovels? And you saw our Entropic up on stage today, their valuations going through the roof. Do you feel like that's going to be suppressed? And Dave premise that the real value is going to be at the application layer, do you guys buy that premise? And how long do you think it's going to take?

Carl Olofson

>> First of all, I think that the language models will be more specialized so that they are focused. If I'm going to do search on medical information, that's the only terminology I need to have in the system. I don't need to have latest the latest sports scores and all this other stuff. So there's that. At the application level, I think really a question of focusing the system on the problem to be solved, and then focusing the application on what problems the system can solve, making those all work in harmony. So there's definitely an application level challenge involved.

Sanjeev Mohan

>> On Hugging Face every week, there are 10,new models, 10,000. I haven't checked the number in the last few days. I think it's about 600,models. So I don't think when language models, and I define those as like hundred billion parameters or more... Like billion for LAMA3 is coming out with a new one. I think that scale should not be done by database complete. That scale should be done by the hyperscaler, Meta, OpenAI, Anthropic,

Carl Olofson

>> What the database vendors can do is the code should fit into that

Sanjeev Mohan

>> frame.

>> Yeah. Coach people how to create a better, more secure, resilient AI application. That should be their job. Oracle, MongoDB, Snowflake, this is what they should be you mentioned security, there is a looming security issue here, which is that with AI systems just combine randomly anything that's out there, including what's in your systems, it could accidentally put together information that reveals things that don't want to reveal.

Sanjeev Mohan

>> Yep.

>> Right.

Sanjeev Mohan

>> Correct.

Carl Olofson

>> And that could be a serious problem. Yeah. I haven't heard too much AI governance today.

>> A little bit on security. But guys, so much, I love having you here. I could go on. in my ear that we got to go, but I'm dying to ask you some other questions about whether or not Mongo should be doing an LLM like Arctic and DBRX, but let's save that for another day. Last question. Plug some research that you guys are doing. IDC guys constantly doing studies. Sanjeev, you're doing some deep work. Carl, what you got going that we should know about?

Carl Olofson

>> Well, we've got what we call the AI council. It's the second year we've been doing it. We're generating... That is not a subscribable service. That's basically an internal service for the rest of us, but you're going to be seeing AI research in all of our services that reflect what they've been finding. We continue to learn a lot from our first research. That's an acronym. I'm not going to tell you what it stands... Well, I can tell you what stands for. Originally, it came out... Never mind. What it talks about is the spending tendencies and trends across the world, broken down by region and country and technology. And that goes on. And those things are all available. Of course, in my own personal space, I'm looking at how GenAI affects everything that has to do with database. And what does this mean? What does it mean to build a database in the era? What do you need to take into consideration?

>> Right. Sanjeev, what you working on?

Sanjeev Mohan

>> I get distracted easily, and you know that. So this year, I started with how do you create a complete intelligent data platform. And you unify structured unstructured data, you unify workloads, whether they are operational analytics and AI agent.

Carl Olofson

>> Metadata.

Sanjeev Mohan

>> Through a common metadata substrate. That is a very important topic. But like I said, I get distracted. So what I'm doing now, next week, I'm speaking at a real-time analytics conference. So I'm very deep into streaming landscape. And I find that very interesting because as I talk to AI, I want my be transcribed into text, into images. So that's streaming data. So these two need to come together

Carl Olofson

>> and-

>> by the way, streaming is giving time series new life, which is kind of interesting also. But one thing I would just like to finish off with from my perspective is you hear vendors at things like MongoDB saying, The answer is simple. Just use our stuff.Okay? Well, the answer is not simple.

>> Yeah, the simpler this industry gets, the more complicated it gets. Guys, two sharp analysts. Thanks so much for coming on Really appreciate your time and your insights. All right, keep it right there. going wall to wall today, all day. Mongodb.local from the Big Apple. You're watching theCUBE.

Sanjeev Mohan | SanjMo & Carl Olofson | IDC

search

>> Welcome back to MongoDB Local, everybody. You're watching theCUBE, our live coverage wall to wall. Actually, a full day. This used to be a two-Day event. Now, it's a one-day event. We're here in New York City, the big Apple. It's a great time of year. We saw the Knicks play the other night. They went into overtime. They're now going to go to go invade Philly. The Rangers, they got through, they swept the capitals. You got the Islanders, they got swept. And of course, the most exciting two minutes in sports this weekend. May is such a great month, the Kentucky Derby. But we're here covering MongoDB local, and we're going to do the analyst angle just coming off the keynotes. We're here with Sanjeev Mohan, who's the principal at SanjMo, former Gartner analyst. And Carl Olofson is a principal database analyst, senior DB analyst at IDC, probably a VP if I botched your title, but-

Carl Olofson

>> Research VP.

>> Research VP. Awesome. Which means you don't have to manage a ton of people and do all those performance reviews. You just focus on your research. Gents, welcome back to theCUBE. Great see you both.

Sanjeev Mohan

>> It's great you.

>> All right, Carl, start with you. Same question to you Sanjeev. Carl first. Keynote, what will your takeaways? What'd you think? Let's riff on it.

Carl Olofson

>> Well, I felt that Dev struck all the right tones in his speech. He talked about the things that I expected him to talk about. It was a good talk overview of where the company's going and how the products are developing. And then Sahir got into the details with the various technical improvements that they're going after. I felt that, and this is a common thing MongoDB, they're very much aimed at the developer. So they're talking to the developer, basically saying, This is what we're doing for you.This is all aimed at the developer. So it's not aimed at the high-end strategic planner or the head of the CIO or something like that. They're talking to the development organization.

>> A lot of people, Sanjeev, would say that they need to expand their messaging beyond the developer, start speaking more wallet in order to continue to grow. I don't know if you agree with that, but your take on the keynote.

Sanjeev Mohan

>> So I Liked what Dev said. Two things that really liked, one is this thing that he said Sequoia came up with called the Crucible moment, which is a decision you make today will have an outsized impact down the road that you may not even see. So his whole thing was that AI is that moment. It's really revolutionary, and the fact that most companies don't want Mickey Mouse kind of use cases because they're not transformation, or like chatbot. They want examples of Novo Nordisk reducing their time from 10 days to minutes. That is what... And the other thing I liked was where he said that anytime a new technology comes up, the value is generated at the bottom most layer, which the infrastructure. And we are seeing that with NVIDIAs and the AI model developers, but the real value happens at the top most layer, which the applications. So the fact that they have announced this MongoDB AI applications program is quite a good thing.

>> Yeah, it was interesting. He's somewhat reframing the AI stack. On the earnings call, he talked about the three layers, the infrastructure layer, the trading,-

Sanjeev Mohan

>> Tooling layer....

>> tooling layer, and then the apps layer.

Sanjeev Mohan

>> Correct.

>> And he even said, We don't really play in layer one. We play...But the chart he showed today had Mongo right the middle, which was kind of interesting. The big news here is their MAP program-

Sanjeev Mohan

>> MAP....

>> where they're bringing together the application because-

Sanjeev Mohan

>> Correct....

>> I look at this as a way to bring their customers into that second and third layer. Do you guys see it the same?

Sanjeev Mohan

>> Yeah, because it's a partnership that includes all the major players, all the three hyperscalers, all the integration players like LlamaIndex, model players like Entropic, and then also it includes professional services companies to deliver solutions.

Carl Olofson

>> So it's a tricky thing because you have a bunch of moving parts here. You want to make it sound simple. In fact, he repeated several times, We want to make things simpler.

Carl Olofson

>> Picture is inherently complicated.

>> Yeah. Yeah. I want to ask you Carl about the whole wrap on Vector. So here, it talked about-

Carl Olofson

>> No....

>> and Dave as well sort of referenced it, and Sanjeev, I'm sure you have thoughts as well. And we certainly hear this from a lot of vendors that are integrating Vector into their database, that it's really a feature, not a product. I tend to agree with that, although I see this Tech Crunch article with all these vector search companies getting all this VC money now. And so what's your take as somebody who's watched the database industry since you were in high school? What are your thoughts on... Is there value on standalone vector or would you advise customers? It is probably an it depends answer to really sort of think about integrating.

Carl Olofson

>> Well, it does depend on what problem you're trying to solve. But I do believe that ultimately, Vector technology will be integrated with other AI technologies. It won't be sold as a product in of itself. Okay? But you'll have Vector that's being used for general text AI and that kind of thing where you just ask an English language question and it sweeps through the large language model and pulls together an answer based on your question. And then there's the more specific database centric question or service, whatever it is, that requires for one thing, a more grounded model, because you don't want it to go off in space and look at irrelevant things, and also that it needs to be focused on what is actually there. In other words, when I ask a question about what's in the database, there are two things that have to happen. One is that the response needs to be relevant to my question, and the other is that the answer has to be absolutely positively correct. It can't be approximately correct. can't be vaguely, sort of, yeah, maybe. It has to be absolutely correct because databases are... Data is either right or wrong. There's no in between.

>> And so I infer from that that good enough is not good enough in certain use cases? And that's where you would use the standalone vector database as an example, or not

Carl Olofson

>> necessarily?Well, when you use a standalone vector database, in other words, as I'm looking to gauge... Say I'm asking a question about customer sentiment, how do people think out there in the world about this or that kind of product or concept? It can come back with a textual answer that is close enough to be right that I can use that as information. But if I ask, what are the factors that caused production to increase in the last quarter? And by how much? Those numbers have to be correct. They can't be approximate. And in that case, you really want a vector system that is close to the database that is actually part of the system.

>> How do you think about this, Sanjeev. In security, the recent data from ETR shows the number of security vendor is actually increasing, not decreasing. And when we ask them why, they say, Because we need to fill gaps. We need best of breed.My question, if you take that concept, it apply to databases or are there too many database types? There's graph. There's there's time series, there's vector. Do we need do customers need to simplify, or is it really they need best of breed to fill gaps, drive revenue, drive productivity? What are your thoughts on that?

Sanjeev Mohan

>> It's a very interesting conundrum, because typically I feel the technology sector moves in tandem, what's for the goods is good for the gander, but not in this case. I think security people are very, very particular because it's black and white. But in the database I think you see more consolidation opportunities than you see in security. I don't think anyone wants to have more and more components of the data architecture than what we have today. So the things are different. I don't see... You were right by saying that use cases will determine whether you need a standalone vector search database or not. But more or less, it's better for the vector capability to be closed to my corporate data. And my corporate data is sitting in a document model. It's sitting in relational. It's sitting in grav and spatial. So I think it's much better that they are co-located rather than they're in a separate database, so I'm having to ETL like. Even if there's zero ETL or whatever it really isn't. It's still ETL. It's just automated.

Sanjeev Mohan

>> ETL, a automating and hiding the complexity. You don't... Because every time you go from one stack, one system to another, you've got to deal with some overhead latency security loopholes. So I see there's more consolidation in the database world may not be in security.

>> Yeah. So that should bode well Karl for a company like Mongo. I wonder... IDC gives a lot of thought to markets, of course. How do you think about Mongo's TAM? At the highest level you've got operational database and analytical databases, and folks will... Dave will say, Well, we don't participate in this. We don't compete against Snowflake and Databricks.At the same time, they provide analytics. But how do you think about their TAM? They're roughly 2 billion company, call it. A lot of people think, wow, Mongo's got a huge TAM. They should be growing faster than what they're saying, or they should be a bigger company. What are your thoughts on their market opportunity?

Carl Olofson

>> Yeah, there's no consistent view as to what the future opportunity for document database is. Okay? And you mentioned operational versus analytical, and there's kind of like a water's edge for them with analytical because they're document database. They don't, by nature, combined data in the ways that you need to do analytics, so they have to partner for that, but they can be more clever in the way that they enable people to pull the data that they need and reorganize it into an analytical format, even if they do drop it into a data warehouse or something like that. So there is plenty of opportunity to take them a little bit beyond the space that they've been playing in up to this point.

Sanjeev Mohan

>> So Carl, do you remember, in the last few events, they used to talk about a product called MongoDB Data Lake? They used to talk about data mesh. I don't see that.

Carl Olofson

>> Yeah, they didn't mention anything like that

Sanjeev Mohan

>> here.In fact, last year, a very big announcement, or the year before, was a native SQL driver.

Carl Olofson

>> Right. That's right. In fact, yes. And they didn't mention that either. This seemed like an exciting capability where they could map data into columns and you could execute SQL statements against it. We didn't hear about that this time. It may be that they're finding that they're getting more traction by working cooperatively with people like Databricks and Snowflake than trying to do it themselves.

>> And then what'd you call it? Because there's a SQL, my mind, SQL converter or

Carl Olofson

>> That's a relational thing, migrating from a relational database to a document, David.

>> And when they talk about it, they don't ever mention Oracle, but they're really talking about Oracle. That's the big wheel. Well, maybe everybody is.

Sanjeev Mohan

>> The whole cottage industry around moving from oral repairs

>> Of course. Yeah, and yet Oracle continues to grow. What are your thoughts on RAG, specifically moving beyond RAG? RAG is actually not that hard to do. People are finding, yeah, it's nice, but to the earlier points that were made in the keynote that the ROI isn't really with the chatty stuff,

Sanjeev Mohan

>> Correct.

>> It's really going beyond that. And those examples that you give Novo Nordisk, et cetera, there's a lot of them. When you come to conferences like these. There's dozens. But in the grand scheme of thing, it's pretty small. It's not big enough to move the market today. So how do you see RAG adoption and moving beyond RAG to get higher levels of value? And how long is that going to take?

Sanjeev Mohan

>> One of my favorite topics is exactly what you just said. The topic is, how do I get AI to work on my corporate data? RAG was the first iteration that really took off. And in my opinion, I think RAG, we fixed it. We know how to do RAG. Now, we need to move our thinking beyond RAG. I said, is there something else we can do? For example, before RAG, there a huge topic about prompt engineers, how important prompt engineering is. And I really think prompt engineering is

>> ridiculous.Big snore.

Carl Olofson

>> No, because prompt engineering is basically a new query language. That's what-

Sanjeev Mohan

>> Exactly. And how does one know, what is a good prompt? So what the future is going to be, maybe hopefully later this year, the cost is coming down of inference is some of a combination of fine tuning and drag. So you take a model. It has to be open source because you need to know the model rates. And then you basically train it or you fine tune it with some additional layers, corporate data, and then you do RAG on that, because RAG by itself is getting us to a point, but not beyond. Fine tuning will help a little bit more. Also, in context learning, which is like ChatGPT has introduced this concept of sending a file with my prompt. This is brand new. It came out last week, which means I may have CSV file and I say, Here is my templated SQL. And the response for different questions. Learn from this, so in context learning, and now answer my question. So this is the expanding ecosystem and why MongoDB map is important, because it's not just about vector search, it's about vector search, RAG, fine tuning in context, learning prompting, few short, many short, all of that is needed to be successful.

>> That's a

Sanjeev Mohan

>> curated-Correct....

>> capability that presumably confers competitive advantage to them and takes advantage of all the bedrock

Carl Olofson

>> Except that the technology is still evolving.

>> Yes.

Carl Olofson

>> So because the technology is evolving and is not static, that means that the techniques we know today may not work as well in the future. There'll be new techniques that work better.

Sanjeev Mohan

>> Oh, completely.

>> So are they going through a one-way door potentially, or-it's always the challenge.

>> It has to be managed.

Carl Olofson

>> When you're an early adopter, you're taking that risk, right? Am I going down the right path, or am I going down the wrong path, or is it slightly wrong?

Sanjeev Mohan

>> So you are 100right. The whole AI space is evolving so rapidly. In fact, there's even a discussion these days that have the transformer architecture? Because if you know, transformer came out in 2017 when Google engineers wrote, Attention is All You Need. But is attention really all you need? Or maybe, what's next? There are some new technology. Mamba one of them. So it's going to change.

Carl Olofson

>> Right now in the database world, there are three things that database systems can do. They can help you write a schema, they can help a developer write an application, and they can help an end user find data, generate the SQL or whatever query language is appropriate for the database they're using, whether Cypher or something else, to find the data they need. Even if they're not technical people, those are the three values that they have. But what people really want, and I gave a speech back 20 years ago at Directions and IDC, in which I said, what people really want is the star Trek computer. They want to just be able to ask a question.

Carl Olofson

>> Yeah, and then get an answer. And I said, and we're a long way from that. There you in order for AI systems to actually understand enough about what's in the database to give you that level of capability, you need a lot more metadata. You need a lot more information. That gives it the context because there's no semantics in structured data, none whatsoever. There is a human language, and that's what drives the way these systems work.

>> Because ChatGPT and Perplexity, they have all the data the internet, but it's the that they're lacking. But your prediction is actually... We're kind of there with even search today. Think about how much search has evolved. I think people say, Oh, the early days of dial up.Maybe it is. Yeah, it's good. I to ask you guys this headline that he put, I put this tweet out, it said, internet may be just a passing fad as millions give up. When I crossed out internet and put ai, there's a lot of similarities. I don't want to over rotate on the .com bubble. Every bubble is different. But you remember, you guys remember it Lycos, it was Yahoo, all portals, or CMG, and it was VCs paying companies to advertise on portals and those portals where stocks were going through the roof and then the whole thing blows up and who comes out of it alive and thriving Is Amazon, Google, PayPal, social media years later? Do you see a similar trend happening today where you have all this CapEx?

Carl Olofson

>> Yeah.

>> The hyperscalers They got so much money they know to do with so they buy GPUs like crazy, NVIDIAs. What did Dave say? the greatest quarter in the history of quarters-

Carl Olofson

>> Correct....

>> ever, right? so there's a lot of money changing hands within the insulated echo chamber of the tech world. Do you see that as a similar pattern to ;IDC research has shown in this area that there's a lot of interest out there in the world in AI and in building AI capability. There's a lot of ignorance as to how to go about it, and there's little appetite for spending money on it. So what that tells me is that this is going to be a slowly evolving thing, but expectations will go like this. The actual development process will go like this. People will get frustrated.

Sanjeev Mohan

>> So here's my thinking. All three of us are old enough to have lived actively through the .com.

>> It was awesome.

Sanjeev Mohan

>> It was amazing. please bring it back. We need If you remember, one of the big shocking disaster, there were many of them, but one of was Webvan. in fact, I had a green box for grocery delivery that came out in 1999, and it was an utter disaster failure. But then what happened, quietly, the technology matured.

>> Yes.

Carl Olofson

>> And that's exactly my point. They had a lot of the right ideas, but they didn't have all the supporting technology to make it work.

>> And the economics didn't work. And then when the VC money dried up, they couldn't get the bridge.

Carl Olofson

>> The other thing, by the way that totally changed the picture was smartphones. Because I don't go to my computer to order groceries or to order food.

>> Mobile.

Carl Olofson

>> I go to my smartphone. That's totally changed the-

>> And that came seven years, seven, eight years after the .com bubble. But you're right, Pets.com, now Chewy, is doing great. You got storage networks. Most people never heard of storage networks. That became the cloud. What they were doing was they were taking storage boxes, a little bit of a little bit of Compaq, whatever, just running data centers, hosted data centers. They just didn't have the economics until Amazon came along.

Sanjeev Mohan

>> There was this very famous example just a month or two ago where Air Canada had a fear that they published, and then it was a mistake. And everybody said, Look, GenAI does not work,or it turned out, literally a month later, Air Canada said there was no GenAI.

>> Right.

Sanjeev Mohan

>> So this is the kind of narrative Now, there's also another very famous case where this car dealership, General Motors car dealership put a chatbot up, and some guy went and said, Which is the best truck to buy?And answer was Ford-150, 150. So people said, Look, you should not be doing GenAI.And my thinking is that kudos to the car dealership. They put that ChatGPT API to do this. Now, what they... Instead of...

>> First the sales guy will, say, Oh, it's hallucinating.

Sanjeev Mohan

>> Yeah. But instead of retracting, I hope what this car dealership says, Lesson learned ground the answers in our database.

>> Yes, exactly.

Sanjeev Mohan

>> Then they'll be ahead of the pack. That's the-

Carl Olofson

>> Again, a lot of work involved in making sure that the responses stick to the subject and they're within the range of possible answers. And that is still... A lot of times when you get to the point where you say, Okay, now people have to do some work,then the enthusiasm drives up, right?

>> So do you think, bringing it back to the AI era, do you feel like LLMs get commoditized, that we'll see all the hype around the picks and shovels? And you saw our Entropic up on stage today, their valuations going through the roof. Do you feel like that's going to be suppressed? And Dave premise that the real value is going to be at the application layer, do you guys buy that premise? And how long do you think it's going to take?

Carl Olofson

>> First of all, I think that the language models will be more specialized so that they are focused. If I'm going to do search on medical information, that's the only terminology I need to have in the system. I don't need to have latest the latest sports scores and all this other stuff. So there's that. At the application level, I think really a question of focusing the system on the problem to be solved, and then focusing the application on what problems the system can solve, making those all work in harmony. So there's definitely an application level challenge involved.

Sanjeev Mohan

>> On Hugging Face every week, there are 10,new models, 10,000. I haven't checked the number in the last few days. I think it's about 600,models. So I don't think when language models, and I define those as like hundred billion parameters or more... Like billion for LAMA3 is coming out with a new one. I think that scale should not be done by database complete. That scale should be done by the hyperscaler, Meta, OpenAI, Anthropic,

Carl Olofson

>> What the database vendors can do is the code should fit into that

Sanjeev Mohan

>> frame.

>> Yeah. Coach people how to create a better, more secure, resilient AI application. That should be their job. Oracle, MongoDB, Snowflake, this is what they should be you mentioned security, there is a looming security issue here, which is that with AI systems just combine randomly anything that's out there, including what's in your systems, it could accidentally put together information that reveals things that don't want to reveal.

Sanjeev Mohan

>> Yep.

>> Right.

Sanjeev Mohan

>> Correct.

Carl Olofson

>> And that could be a serious problem. Yeah. I haven't heard too much AI governance today.

>> A little bit on security. But guys, so much, I love having you here. I could go on. in my ear that we got to go, but I'm dying to ask you some other questions about whether or not Mongo should be doing an LLM like Arctic and DBRX, but let's save that for another day. Last question. Plug some research that you guys are doing. IDC guys constantly doing studies. Sanjeev, you're doing some deep work. Carl, what you got going that we should know about?

Carl Olofson

>> Well, we've got what we call the AI council. It's the second year we've been doing it. We're generating... That is not a subscribable service. That's basically an internal service for the rest of us, but you're going to be seeing AI research in all of our services that reflect what they've been finding. We continue to learn a lot from our first research. That's an acronym. I'm not going to tell you what it stands... Well, I can tell you what stands for. Originally, it came out... Never mind. What it talks about is the spending tendencies and trends across the world, broken down by region and country and technology. And that goes on. And those things are all available. Of course, in my own personal space, I'm looking at how GenAI affects everything that has to do with database. And what does this mean? What does it mean to build a database in the era? What do you need to take into consideration?

>> Right. Sanjeev, what you working on?

Sanjeev Mohan

>> I get distracted easily, and you know that. So this year, I started with how do you create a complete intelligent data platform. And you unify structured unstructured data, you unify workloads, whether they are operational analytics and AI agent.

Carl Olofson

>> Metadata.

Sanjeev Mohan

>> Through a common metadata substrate. That is a very important topic. But like I said, I get distracted. So what I'm doing now, next week, I'm speaking at a real-time analytics conference. So I'm very deep into streaming landscape. And I find that very interesting because as I talk to AI, I want my be transcribed into text, into images. So that's streaming data. So these two need to come together

Carl Olofson

>> and-

>> by the way, streaming is giving time series new life, which is kind of interesting also. But one thing I would just like to finish off with from my perspective is you hear vendors at things like MongoDB saying, The answer is simple. Just use our stuff.Okay? Well, the answer is not simple.

>> Yeah, the simpler this industry gets, the more complicated it gets. Guys, two sharp analysts. Thanks so much for coming on Really appreciate your time and your insights. All right, keep it right there. going wall to wall today, all day. Mongodb.local from the Big Apple. You're watching theCUBE.