Emil Eifrem, Neo4j | theCUBE + NYSE Wired: Mixture of Experts

Palo Alto studio connection Silicon Valley and Wall Street. I'm John F co here with Dave Volante my co-host here at the cub's NYC studio of course of our Peloto studio connecting Silicon Valley to Wall Street. This is our New York Stock Exchange wired program MC wired program and community with the cube. Of course, we've been covering the mixture of expert series. Of course, graph talks in time with Neo4j. And we're here with the CEO and founder MLFM is a friend of the cube. Great to see you. Thanks for coming on. Been a while. >> Great to be here, John. >> We've known each other since 2007 when I mean Facebook was just starting to talk about social graph and they would have events. Um, but such a great success for your company and your team. Um, graphs are now infrastructure, data is infrastructure, AI needs real time. You guys have been great success. So, first start with some of the momentum of what you guys have. Obviously, you're in town for the graph talk. I'll be hosting a bunch of interviews from a lot of graph practitioners who are leading the market by the way, but get into what the momentum is you have right now. >> Yeah, you know, like you mentioned, we've known each other for a few years now, 20 years. So, we were, I think, 25 when we got to or something like that. Four, four years. >> Uh, and it's been an like just fantastic journey, right? We started out by saying, hey, there's got to be a better way to represent information and work with data, not force it to be squeezed into square and static tables. The world is dynamic. The world is always changing. Figuring out how things fit together felt like a worthwhile, important place to pursue in terms of using data. So that's how we started. Right here we are now 20 years later, hundreds of millions of dollars of revenue, billions of dollars of valuation and all that kind of stuff. But really what's exciting now is what's happening with graphs and AI. And honestly, Don, it's like this match made in heaven. >> I mean, it's computer science match too. Recursive nature of the graphs. AI speaks graph, math is graph. >> It's exactly right. And I think what's happening is that the graph representation of data, so organizing your information in nodes and then relationships between them and key value properties on both ends up being this very information dense or knowledgerrich way to represent your data. This is why in 2012, 5 years after you and I met, Google launched their knowledge graph. And this is the thing that we now take for given when you search for New York City. Yeah, on Google you're gonna get a side panel with information about New York and then you can click through to the mayor. You can click through to it. It's in the state of New York State, which is the country of the US. That's all on the back end powered by their knowledge graph. They call it the knowledge graph just because it's this amazingly rich way of representing information, right? And so that is this heavenly match with the stochastic transformer-based AI models which really is a 1 plus 1 equals at least three and we see that happening up and down the AI stack and I'm very happy to talk you through some of the key technical patterns and use cases. Yeah, I want to get into I want to get into the tech, but I also want to keep zoomed out because a lot of people who are seeing graphs now for the first time are like, "Oh, magic's happening." But there's a lot of people who have worked hard over the years to to get here. And we I mentioned Facebook at the beginning. You mentioned Google a few years later. If you remember the social media evolution, social graph was the concept. We all use LinkedIn. Some people still use Facebook and others. Uh but that was based on graph concepts. And the power that came out of those monster companies, specifically Facebook, is now available to every company. So most people think, oh, Facebook has this graph. They were highly efficient in targeting uh with advertising obviously, but they own the data. But now that it's opened up, it's become democratized. You guys really drove that. Talk about that impact of the market because that you like you said the AI world is now going AI native. Everyone I talk to that's AI native have multiple databases and a graph I won't say overlay but I think that's the wrong word but graph connected talk about that dynamic because now everyone could have the power of Facebook >> right >> and done done right basically >> yeah yeah yeah and I think maybe the first 101 15 years of this company was taking what is in Silicon Valley you have a generation of remember web 2.0 0 was the term right at the time right so this is call it Facebook and LinkedIn but also some web web 1.0 companies like eBay and PayPal, you look inside of the machinery of them, it's all based on graph technology, right? But if you're in the enterprise, outside of Silicon Valley, it was very hard or or at small startup, it was very hard to get access to that technology. We democratized access to that. We gave the same platform that Google was built on like on tap for the big financial services, the big telecom companies, the big life science companies, right? >> A lot of people talk about AI as being bad and there's bad narratives out there. Oh, it's going to kill us, the Terminator, Skynet. But if you think about like where we are from a tech perspective, there are the giants that built the AI generation besides the, you know, the algorithms that came out to make AI work were pioneers. Uber, Neo4j, they had to build the stuff from scratch, right? So, but there was a lot of work done. So, this is building on the shoulders of those giants. This is this next era. There's a lot of work that's gone into it. Okay, that's known. So, there's a lot of domain body of work done to that's in the AI era. So, now explain why that's important when people start thinking, okay, how does AI work better? How do I get contextually relevant information as context? How is it safe? because people just don't understand that it's it's there. They think two kids in a dorm room started AI and it's going to take over the military complex and kill everyone which you know technically war games is a scenario but we don't know but it's kind of a fantasy. Talk about the the reliability and the efficacy of where the AI generation is. Yeah, at this point if you look at the current generation of AI models, but even the prior generations like even like the ones released, call it a year ago, which feels like stone age these days, right? >> Honestly, they're phenomenal. The intelligence in those models is just it if that's all that we did as a society, it would take decades to just diffuse that technology throughout the world and we would get a ton of value from it. When I look across our customer today, customer base today, we're not bottlenecked by model intelligence, we're bottlenecked on what context do we give the models. It's that classic thing of the model is like a PhD in some like narrow area that goes into work every day and has to teach and learn everything from scratch, right? Because they don't know anything about your company, right? And so the game now is about how do we give that model the right contextual information? And this is where everyone is circling around the same thing. It turns out that context, what is context? Context is my context is how I relate to the rest of the world. I grew up in Sweden. I'm the CEO of Neo Forj. I have three young kids. I'm married to Meline. That's I drive a Volvo stereotypically since I'm Swedish. Volvo, of course. right? You know, so that is my context. It's very graphic. It's like how I relate to the rest of the world. So, everyone is converging on this approach to giving the AI models the right information at the right time. >> I said on the cube, you got to feed the beast. Um, in a way, you got to feed the AI and but it's you don't know the context. Static web was, you know, database, search it, results. That was a search paradigm. We're not living in a search par but we are living in a discovery paradigm now with agents explain this because search and discovery has been one of the categories go back to the early days of the web search engines had a discovery mechanism keyword get results whatever click on a link navigate to a page now with AI you still need that discovery layer but you need it fast and accurate talk about that piece because people see agents going off the rails hallucinations they might not understand that there actually there is a solution Yeah. So, let's tee off of what we just said, which is the real game here is giving the super smart AI models the right context at the right time. If you think about that, that in computer science is called IR or information retrieval. It's a retrieval problem, right? And so on some level, it is let's say that you're a human being, you contact customer support. Let's take the classic AI enterprise use case, customer support, right? So, John contacts some like provider of yours, right? And you say, "I need help with my Wi-Fi." Let's say your Wi-Fi isn't working. And you describe it. Say, "This is the one that I bought." Or maybe the system knows it already because it knows who you are. Right? And it says, "These LEDs are flashing yellow and my Wi-Fi is shaky." Right? Okay. The agent has to take that natural language that you typed in or that you said over voice. It needs to go from that intent to let's find the call it top 10 documents across my entire corpus of support documents inside of this company. Give that to the model at the right time to answer that question. Now if you think about that problem it turns out if you take a step back as humanity we've solved that problem before. The problem is the same as you log into a search box on the web, you search right, find me the top 10 blue links and you and I are, you know, sadly old enough to remember that there was a world pre Google where Alta Vista, Leas, Yahoo, Exite. >> Yeah. Yeah. >> These days people don't even know those names. So there's tons of search engines, right? And >> except for Yahoo basically. >> Exactly. And one of the problems was people called it the Alta Vista effect. You search for a result and you're going to get a million results, but the top 10 best ones were maybe on page 99 or 433. Right. Then Google came along and they said, you know what? I'm going to search exactly to your point. But then I'm going to rank the search result based on what? >> Based on the graph. That's the page rank algorithm. How the documents are are linked on the web. So that exact same approach is what the enterprise is now using to get the reliability of the retrieval and this >> and the results of page rank by the way the ranking technology called page rank created massive wealth for Google obviously we know the search they created the whole ad for online but what does it mean for enterprises because this comes up a lot I've done a lot of interviews with some of your uh practitioners as well as other graph enthusiasts they're all having an experience of almost like a superpower but they weren't in the organization pecking order and now they're moving the needle on the business and they're being elevated up because they just discovered it's like a caveman discovering fire. It's like they get pushed right to the top. This is where you start to see the grass. Why is that happening? Why? >> Yeah, it's spot on. Although I'm not going to describe my customers as cavemen. So that was Oh, the wheel. The wheel was revolutionary. >> Yeah, fair enough. I'll take that I'll take that part of the analogy. Right. So, so the key thing here is what's called accuracy. And accuracy is the inverse of hallucination. Like the higher accuracy that you have in your AI system, the fewer hallucinations you do, right? And there's a threshold that people talk about as the escape velocity for accuracy. And when people use graph as part of their agents, they reach this escape velocity where it actually works in production. And that's the key thing. Every single big organization right now is on this AI transformation journey. But if their agents can't retrieve the right data at the right time, they'll never have accurate enough answers for it to be able to be used in production. So that's the superpower that our champions get by using graphs as part of their agenda. >> And by the way, it solves a lot of hard problems they've been grinding on in other mechanisms. >> All right. talk about the impact of the AI infrastructure because you know a lot of database a lot of software you know especially you know control layer software or connective tissue whatever you want to call it glue layer people call it was kind of constrained by the fact that you need a lot of compute so now you were in an era where the new architecture on whether you have an AI factory you unlimited capabilities from a horsepower standpoint you now have engines that can pump out tokens okay which is now currency >> those tokens have changed the game on how data is interacted. How has that affected graph specifically because this is where I think the AI connection to the computer science of of AI with data. >> Yeah. So there's two sides to that coin. The first one we've talked about which is how graphs in Neo forj are embedded in our customers AI systems to help them become better. We've talked a lot about accuracy. There's also an explanability and government sorry governance and transparency angle that is really relevant and important too but we can get to that in a moment. So that's one side of the coin. The other side of the coin is if you look at graph technology there are basically two big areas of friction to adopt graph technology. The first one is how do I get my data in there and how do I model my data. Let's say you have lots of data unstructured or it might sit in your snowflake or your datab bricks or your Oracle data systems. How do you get it into the graph? That's the first one. The second one is how do you query it? Right? So there's a query language now called SQL. It's the first sibling to SQL that has been standardized by ISO ever. Right? So SQL was invented 40 years ago standardized as the one universal database query language. GQ SQL the graph query language is the only sibling to that which is just pretty phenomenal right but it's a new language you have to learn it well it turns out that AI helps with both of those problems and a modern AI model can look at your unstructured data and create a knowledge graph out of that from scratch right without any any human manual intervention and then how do you query it well these days you don't need to type you speak English to it right so that that that barrier to adoption has been just completely collapsed and that's a huge part of what's driving momentum for Neo forj right now. >> You know what's interesting? You brought up SQL structured query language. People don't know the acronym that has been the standard for querying databases. And I want to tie this to intent because one of the things that we've learned here on the cube and we see successful companies doing is they've changed the intent equation. Give an example. I used to do a lot of SQL queries when I was in doing co-op work in the in the 80s. You have to think about the business logic formulation first. Then actually construct the query. The query then is my query to the databases. I need a report you know. Okay. Now I create the logic. Then the query goes in. So I have to formulate the query. Sometimes they're huge queries. Now that intent is in the logic of the AI. So I just say I want the top sales by region or whatever my ask is. It does the logic on the other side. that wasn't possible. Explain this because this is like like a gamechanging shift in user experience but also technical implementation. >> This one is huge. In order to pull this one off at enterprise scale, you require a technology that is absolutely fundamental which is ontologies. And ontologies have been around forever. Aristotle talked about >> I did one in 88. I did one in 88. Taxonomy by hand. >> Yeah, exactly. Tom Gruber coined the the most common definition of ontologies for computer science in 1993 at Stanford. Um he then went on to co-found Siri by the way the same same Tom Tom Gruber but really the company that has popularized this in modern times is Palunteer right and they started talking about ontology being their secret sauce right and what an ontology does it's a graph model so it is exactly 101 what we've been doing for 20 years and a businessfacing ontology is the key asset here and what it is is it takes the world of your company. Let's say you're a financial services institution. You have customers. The customers have bank accounts. They have credit cards connected to them. There are derivatives. There are pair like all the concept that exist in your universe and how they relate a business facing ontology because it turns out that a key part of your job as an enterprise data architect today is that you have to design the world that your agents think in. And your agents think in the terminology of the business. They think customers, they think credit cards, they think accounts, they think insurance plans, that kind of thing, right? And then you require a technical ontology which is all my data sources the physical data layer in my enterprise. I have an Oracle database over there. I have a snowflake database over there. And then a mapping between the two. So all of a sudden you know that a customer first name maps to that Oracle database with a column called F_name. The agent wouldn't know that F name is the first name of your customer. But with these key ontologies to connect them, you can do exactly what you said. Your agents can interpret intent and map that to the data that they need. >> And the graphs make it faster. So talk about the how think about graphs. It's almost like picking a fork in the road. Like you can say, okay, down this lane is a series of graphs, but it makes it very efficient. Um the word recursion comes up a lot in graphs. I mean, graphs are nodes with an arc connected to another node and there's data in these things, right? So, like that's computer science principle 101. You traverse the nodes and you go see what's in there. So, take us through why that works in AI now that you have intelligence >> and horsepower. You have compute and and all the GPUs and all the vector embeds and all the other data. >> So, the key secret sauce here, if you take, we talked before about the knowledge graph of describing all the key concepts and how they relate and all that, you can take that data structure and you can store it in anything. You can put it in S3 buckets. You can put it in Oracle, Postgress, whatever you want, right? But what a graph database is like Neo forj, it's written from scratch. Again, we've been at this for 20 years now, right? And we've taken every single layer in the stack of the database and we've optimized it for exactly what you said, which is traverse this deeply connected web of information at lightning speed. So we are frequently because it's perfect. >> It's called the knowledge graph basically these days, right? Exactly. And it is frequently not even a thousand times, but a million times faster than if you put it in a classic relational database, which is great at many things. It's just not great at traversing this deeply connected data that an agent requires in order to answer >> and they and they're bounded by latency, too, because their accuracy is only as good as the most best data possible. >> That's exactly right. >> All right, let's talk about your momentum in the company. Obviously, great success. Love the tech angles. I think every companies wants their own data mode. They want a palunteer-like environment. All the smart money and smart people are using graphs with other databases. Talk about the momentum in the company. You mentioned the re some of the revenue figures. Can you quote the numbers? Can you share some of the momentum where you guys are at and what's your focus now? >> Yeah, so we're we're here at the New York Stock Exchange. We're not yet a public company, so we don't we don't disclose our our numbers publicly. Suffice to say, we're hundreds of millions of dollars of revenue. And just to give you a sense of the momentum, like 2026 is off to a flying start in Q2 of this year. So, we're we're recording this in midepptember. So, our last quarter was was Q2 of of of this year. We generated about as much revenue in just that single quarter than all of 2025 combined. Right. So, that just gives you just a flavor. it is really taking off and there's this widespread recognition that AI and graph is this again match made in heaven. >> Talk about the community that's developing around graphs. I think this is super fascinating um because the people who are doing graphs were early adopters because they saw the value and it's almost like they are discovering the superpowers and it's spreading. Talk about the how that's spreading in the community and what you guys are doing about it. >> Yeah, one of the things that I love about this company, of course I'm extremely biased being the founder the founder, right? But but one of the things is that we've always had this practitionerled adoption where the people who do the real work. They find us. We're open source. We have now in the cloud world, we have a free tier of our cloud offering. They self on board. They fall in love, right? And they can build it themselves, right? And so we only ever sell into people that are just champions. We never sell. just start top down, push it down into into the org. Now, we of course at this scale, we also engage with the real technical leadership of the global 2000, >> but it's based on this foundation of the people who actually sit there doing the real work wanting to choose to work with Neoforj >> and making it easy. >> Making it easy, I know, is always hard. Take talk about the ease of use feature. How are you guys making it easier? Because, you know, we want to get our graphs going in our company. Every company I talk to is trying to figure out the brain for their company. They're all come to the realization that okay, we need a company brain. We need a Google page rank. We need to have a palenteer. We have data that's valuable. How do we protect it? >> Yeah. Yeah. And a brain like even the human brain is physically a neuron connected through another neuron through signapses. It's physically a graph, right? And associatively, we we think associatively which is also a graph, right? Yeah. It comes back to what we talked about before. the fact that we now have AI. I spend most of my time thinking about how will this how can Neo Forj help my customers AI applications become better. But it's also a massive superpower for us internally as architects of our own product because all of a sudden we have a way of making it so much easier to get data into the database and then query in pure English, right? And those two things is the main building blocks >> and no one really has to give up anything to use Neo4j. They can still use their data links. They can still use everything else. You just connect into it. >> Well, and the other thing is we've also done a lot of investment in we we actually are tomorrow at at our event we're talking about a new product offering called virtual graph which is super super exciting. What virtual graph does is it's the entire Neo Forj product platform. So all of the tools, all of the solutions running on top of it. We can talk about Graphware later if you want to as an example of the solutions running on top of Neo Forj, all of this, all the AI capabilities, but it sits right on top of your snowflake or your data bricks. So you don't have to move your data. It uses in the weeds technically it's called predicate pushdown queries, right? So it runs all of this directly on your pabyte data lake with minutes to get started, right? And so that's really exciting. >> That's really on boarding fast. All right. Well, great to see you and I know you got to go. I really appreciate your valuable time coming on the cube. Uh it's been a couple years. Um we have a lot of your team members on and you got a great team and again graphs are just the beginning. People starting it'll be standard fastest ISO standard on the database piece. Congratulations. Um and we'll we'll we'll talk more later. >> Thank you. >> Pleasure. >> All right. Founder and CEO of Neo4j. Again, the graph database is turning out to be the the heart and soul, the connective tissue, the brain of organizations. And there's benefits. AI is highly compatible. You don't really have to get rid of your other data and databases to really make it happen. Of course, the results are fantastic with AI. I'm John Furry, your host of the Cube. Thanks for watching.