Dev Ittycheria, MongoDB | MongoDB.local NYC 2024

Welcome back to NYC. We're here with Dave Ittacheria, the CEO of MongoDB. We were just walking down from the analyst event. You're like a celebrity. Everybody's taking pictures, signing autographs. Welcome back to the Cube. Good to see you again.

Great to see you, Dave. Thank you for having me.

Yeah, you bet. So I liked what you were saying in the analyst program about these MongoDB Locals. They used to be 2-day events. Now you go, I think, 23 cities around the world. How many of those do you go to?

A couple, you know, but the team here is really does a great job. And the reason we did that is that we realized our buyers are very decentralized. We're not like a Workday or Salesforce, which is typically one buyer. You know, our buyers are developers all around the world. And rather than forcing them to come to us to a couple of shows around the world, we said, let's go to them. So we went from 2 or 3 shows around the world to 23 shows. So we're in Europe, in Asia, and obviously in North America. And it's been a great success. We just were in Toronto 2 weeks ago. We're going to be in Mexico City. A couple of weeks from now, and then we're going global. And it's a great way to reach out to developers, help them understand who we are, where we're going, what's new, get their customers, get local partners involved. So it ends up working out really well.

Go to where the developers are. So I want to get right into the AI stack as you see it, the MAP program that you guys announced earlier this week. But the stack you mentioned on one of the previous earnings calls, 3 layers of the stack, infrastructure, sort of the tools, the LLMs, the building blocks, and the apps. And you play in 2 and 3. And your premise is that, yeah, a lot of the action now is in infrastructure, certainly NVIDIA, Anthropic, the LLM players, the hyperscalers. But the real value where the rubber meets the road is ultimately going to be up the stack, app layer. So explain that in a little bit more detail and we'll get into it.

Sure. So I use analogies like telephony and the internet to kind of make the point, right? So when the internet was being built out, you saw all the infrastructure providers like the Ciscos of the world make money hand over fist. But besides high-speed internet access, people, it wasn't really transforming their life. It's when the new transformative apps and services came to play. Obviously household names that we know now, by the time they weren't that well known, Google, Facebook, Uber, Airbnb all built these amazing businesses on the internet. And now it transforms how we live, how we work, how we collaborate. I think the same thing's happening in AI where, why right now there's tons of investors going into the infrastructure layer, But AI has not really transformed my life, Dave, right? I mean, yes, I can maybe write a better email. I can get research done more quickly. I can automate certain parts of our business, but it's not really transformative. But I think it will be. And the way it becomes transformative is that these next generation of apps that really fundamentally change how we do business, how we drive cost out of the business, how we engage with our customers, how we find new ways to drive growth, that's going to be the way it really transforms not just our business, but any business.

And you think about the real winners of the internet you had in your slide earlier today that, you know, the pictures of Cisco's and Cisco was the most valuable company in the world at one point. They're still a valuable company, but, you know, not number one anymore. And a lot of the names that we remember, we don't— many people don't remember the Lycos. I mean, AOL is irrelevant now. And CMGI, which nobody's ever heard of, was They were the hot companies back then. They're all gone. And it was those guys who came in after the initial enthusiasm. And you described that very well. So I'm trying to understand your thoughts on the parallels, 'cause one of the other parallels is you had all this VC money going in to fund people to buy Unix servers from Sun or storage from EMC or databases from Oracle, and then advertise on Yahoo and Lycos. And it was this like echo chamber inside of the tech world. And then when the funding dried up, all those guys kind of disappeared. You have a lot of CapEx now going into the hyperscalers, obviously benefiting NVIDIA at that lower layer. But one of the other parallels is the real winners of the internet weren't even tech companies. It was the people who applied the internet to their business for e-commerce and others. Do you see the same thing happening here?

Most definitely. So I think as these models mature, as it becomes easier, as it becomes less costly to use these models, As compute becomes, you know, the price performance of compute gets better, it's going to just unlock so many more opportunities. And I think every company, when I talk to customers today, a lot of customers are a combination of overwhelmed and fearful. They're overwhelmed because the rate and pace of change in AI is happening so quickly. Facebook came out with Llama last week. A couple of weeks earlier, Mistral came out with a low-cost model that had great performance. And a couple of weeks before that, Anthropic had, you know, came out with their 3 models. Right. And so For customers, they're like, what do I do? And oh, by the way, GPT-5 is rumored to be around the corner. So that's— they're kind of overwhelmed. At the other stage, they're also a little fearful because they're like, are my competitors doing something I'm not? Are they using AI in ways that I'm not? Am I at risk of being disrupted by my competition? So there's a lot of focus on AI. We see experiments in AI, but we don't see a ton of applications in production just yet in AI. I think they're coming. But we just don't see that just now.

So we see the same thing in our data. And that's what the MAP program is all about, is 2 things. One is helping customers, those, those 20% that actually aren't even pursuing GenAI because they're, they're concerned. And then a big chunk of customers who aren't maybe getting enough value out of their AI. That seems what the MAP program is all about, also to bring people into layer 2 and layer 3 of your AI stack.

Yeah. So the MongoDB AI Application Program, which you call MAP, is really designed to address 3 problems. One, give customers reference architectures to address a certain set of use cases. 2, have a prebuilt set of integrations that they can use out of the box so they can get up and running quickly. And 3rd, bring some technical expertise so that we can help them go from idea to production as fast as possible. And customers want that. And the reason we feel we're well positioned to do that is one, we don't have our own LLM, we don't have our own cloud. So people can say MongoDB is going to be very thoughtful. They're independent. They know how to build modern applications. So they can be a real thought partner to me and really help me pick the right choice for me. And then what we found is we, one, all the hyperscalers are part of this program. They want to be part of the program because they know how influential we can be. Two, we have a bunch of really interesting startups as well, who, whether they're offering their own models, whether they're fine-tuning models, whether they're offering orchestration layers, whether they're providing other tooling to help build these apps. They want to be part of this. And then over time, we'll add more and more partners as we vet them and make sure that they can really add value for customers. And that's just the essence of the MAP program.

Many times you've had to explain, well, we're not like Snowflake and Databricks, we're an operational database. Now, both of those companies have announced an LLM, DBRX in the case of Databricks, and Arctic in the case of Snowflake. You haven't announced an LLM. Do you see a need to have R&D that is funding an LLM, your own LLM?

I think I personally believe that these LLMs coming out from these smaller companies will not stand the test of time. Why? Because it takes so much money to build LLMs. And what Facebook announced with Llama 3 last week The performance they're delivering, it's open source. So why would I go build my own LLM when I can take Facebook's LLM and fine-tune it for our own use? It just doesn't make sense anymore. And I think frankly, a lot of VCs are going to wonder, how can we keep funding new foundational models when Facebook is basically open sourcing the most powerful LLM and essentially, you know, will, some people say, will basically take share from all the other LLMs out there. Now, I do believe there'll be some combination of closed-source and open-source LLMs. There'll be smaller LLMs for particular use cases. So you're gonna have a variety of LLMs, but you're not gonna have hundreds of LLMs. And I think companies, other companies trying to build their own LLM, I don't think is a good use of their time or money.

I kind of agree with that, but I'll tell you, I'll share a story. Yesterday I was meeting with a large financial institution. They told me they're gonna build their own LLM. They're actually gonna work with a semiconductor, a silicon manufacturer, And I was like, why would you do that? They had specific use case reason. Okay, maybe. Can't you get that from Llama 3? And their concern was that Meta could pull the rug out from under them with the terms of the open source license. Do you worry about that?

I mean, that's a fair concern. Llama is not a complete pure open source solution. They have obviously Meta controls the IP. And the other thing that people realize is you can put a lot of money and focus on one version. There could be another version that comes out 6, 12, 18 months later that'll be vastly superior. And all of a sudden all your investments could essentially be for moot. So there's, you know, people need to be cautious, but I don't think Meta's in the business of alienating their customers. And I think the whole point Meta's saying is that we want to build models that are incredibly powerful that people can use because they're the beneficiary of leveraging all the learnings from the community. They're already getting feedback on how to make their own models better. So I don't see them kind of pulling the rug. Now, they have said that some of the future models may not be open. Maybe they have a certain monetization strategy that makes sense for them. But I do believe at the end of the day, to the essence of your question, for a company like MongoDB or other companies of our scale, and we're not a small company anymore, it still doesn't make sense for us to build our own model. We may use models and fine-tune them on our own data, for particularly internal use cases or if we feel that we need to embed into our product. But to start from scratch makes no sense whatsoever.

This is the other similarity yet difference with '99, 2000, the dot-com is, you know, back then you were building out the LAMP stack. Linux in many ways sort of moderated Microsoft's dominance. Again, LAMP stack got built out. Today you see these open source LLMs kind of pulling the If you have a power law, they're sort of pulling the torso up, you know, which is different and unique. So my question relates to LLM optionality and diversity. Obviously you believe that. Obviously AWS believes that. I don't think anybody says there'll be one model to rule them all. That's marketing.

I actually would argue that even Microsoft believes that. They've invested in Mistral. They're, you know, they're going with a whole portfolio approach, much like AWS's.

Right. So I wonder if LLM optionality is— I guess it's necessary to build apps. Is that the right answer? In other words, it's not necessarily a moat to have optionality and diversity, but it's necessary for the right tool for the right job.

I think the way people should think about LLMs is like the operating system of the future. It's necessary but not sufficient, right? And you don't need to own the operating system to build the apps. And so we think that we're in the business of helping developers use AI to build applications that transform a business, right? Back to the first question you asked, where will people see the value? They'll see the value through apps that enrich a customer experience, reduce cost, find new opportunities for growth. And we're in the business of helping customers do that. And we are LLM independent. So if people want to use OpenAI, they want to use Llama, they want to use Anthropic, they want to use something else, You know, that's fine with us. We want to make sure that we set up developers to build whatever they want. And there'll be LLMs optimized for particular use cases. There'll be other architectures optimized for different use cases. And that's the whole point of the MAP program, is to actually help customers decipher all that.

So if it's the OS, then that would say there's going to be some consolidation, right? I mean, you basically have Linux and Windows, right? Is this like Unix where you had HP Unix and Sun Unix and IBM Unix?

I mean, obviously the question is like, I was, you know, the question is like at the rate and pace of investment, how many people can afford to raise billions of dollars to train a general-purpose model? Now, there may be use case-specific models that make a lot of sense for particular use cases that are smaller, more performant, offer better latency, etc. But to build general-purpose models, I think over time there'll be probably a handful of general-purpose models.

And then I know you get this question a lot, but I'm curious to weigh in here. The whole commoditization of LLMs is, On one side of the equation, it's like, yeah, these things will be commoditized. On the other side is, hey, there's a lot of CapEx or money out there that could go into innovation. And as long as there's innovation, you're going to have at least, you know, a good handful of choices. What do you think?

Yeah, and I think choice also definitely drives innovation because it forces people to compete and better serve customers. I think one of the things that's interesting is today it's very easy to switch from one foundational model to another. But as these, and you're starting to see LLMs start to introduce memory, they're starting to introduce memory into their platforms so that by definition they retain all your history, all the data that you've previously provided. So it becomes, switching becomes that much of a higher switching cost, right? Today it's near zero, but over time that may be so. So that's, I think, one way that LLMs will try and differentiate themselves is by capturing a lot more of the memory of what you're doing on them. So that's another interesting angle to watch. But again, at the end of the day, it's all about what are the apps and services are you building, right? Even something like Perplexity. Perplexity is packaging its search offering by using an LLM and a bunch of other tools to create an app for search, right? It's a very popular app, but there'll be other, you know, consumer apps that do things like Perplexity for other use cases. And then enterprises will have their own apps to serve their own needs. And that's where I think the real value will come from.

I think it was you who likened it to the dial-up era. I think we were using that. So you think we're gonna look back 10 years from now and go, you remember ChatGPT, how bad that was? You remember Perplexity? Wow, what a joke. Remember the interface? I don't know. I mean, it would seem with exponential curves, we're probably gonna see, they're gonna blow our minds with this.

Most definitely. I think we'll look back and laugh at what we're using right now.

All right, last word on MongoDB Local. What's the future look like for MongoDB? What are you excited about?

Well, I think one, you know, one of the things that's interesting is, you know, when you look at every era, Dave, like from the mainframe to client-server to the internet to cloud and mobile, the cost of building apps got lower and lower. So what happened? There were more apps and more data. You're gonna see a step function change in developer productivity with AI. So that's just gonna unleash a whole new generation of apps, apps we can't even imagine today that will profoundly impact our life. And I think we're excited by playing a part in making that happen and really enabling customers and end users to really transform both their business and their personal lives using AI.

Dev Ittycheria, thanks so much for coming back on The Cube. It's great to see you. And check out thecubeai.com. It's running on Mongo. We love it. We haven't done your vector embeddings yet. That's next. And we love it. We love Mongo. We love the community here. Thank you very much.

Thanks, Dave. Thank you for having me.

You bet.

All right.

Keep it right there for more action from MongoDB Local in New York City. Be right back. You're watching TheCUBE.