Name: Jim Webber, Neo4j
Uploaded: 2025-06-20T15:00:00.000Z
Duration: 17 min 54 s

Clips
More from theCUBE + NYSE Wired: Physical AI & Robotics Leaders

Speakers

Jim Webber

Chief Scientist

Neo4j

play_circle_outline Compute benefits and lightweight nature of graph databases

play_circle_outline Deployment options for implementing graph databases in existing ecosystems

play_circle_outline Novel approaches to building databases using compiler-inspired architecture

play_circle_outline Infrastructure advancements in AI hardware and the need for good compilers

Info
Transcript

Jim Webber, Neo4j

Jim Webber

Chief Scientist Neo4j

Exploring the Role of Graph Databases in AI Advancement

Jim Webber, the chief scientist at Neo4j, engages with John Furrier of theCUBE during the NYSE Wired Robotics and AI Media Week. This insightful discussion delves into the innovative application of graph databases in the realm of artificial intelligence and highlights key perspectives from experts within the field.

Webber shares their extensive experience with graph databases, emphasizing their pivotal role in AI application development. According to Webber, graph databases provide the ... Read more

explore Keep Exploring

What are the advantages of using graphs over other types of databases in terms of compute? add

What are some examples of organizations that have built knowledge graphs of lessons learned and institutional knowledge? add

What approach is being taken by the team at NEO4j to build a completely novel database architecture? add

What are you looking forward to in terms of AI infrastructure advancements and what do you think the infrastructure hardware guys need to work on? add

bolt Powered by CUBE AI

Jim Webber, Neo4j

>> Hello, welcome to theCUBE here in our New York City NYSE

>> New York Stock Exchange Studio. I'm John Furrier, host of theCUBE. It's our East Coast location. This is where our access point is on the East Coast. Of course, we've got Silicon Valley and Palo Alto. We've got East and West connected, tech and Wall Street together. We've got a great week this week. The NYSE Wired community and theCUBE together doing a robotics and AI Media Week all the leaders in AI. A lot of great action happening. Jim Webber is here, chief scientist at Neo4j, a company we've covered extensively on SiliconANGLE. com as well as here on theCUBE. Jim, thanks for coming on remotely from the UK.

>> Absolutely. My pleasure, John. Thank you for inviting me.

>> So, obviously, people know SiliconANGLE and theCUBE. They know we cover extensively the Neo4j and, of course, the graph databases, which really is powering this next generation of AI applications. Of course, the underlying infrastructure behind it, data layers are super important. A big part of the innovation cycle is the data and how data is structured and graphs are a big part of it. You guys are doing that and leading that. And so, as the chief scientist, I got to first ask you, are you not surprised? Of course. I mean, you got to be pretty excited right now about what's going on with graph and all the innovation.

>> I think, not to be hyperbolic, I'm not surprised. I've been plugging away at graphs for a long time. And I think, like yourself, I've realized the utility of graphs both for kind of traditional what we think of as kind of database workloads and analytic workloads, but more recently, with the emergence of particularly generative AI using graphs as an underlay to provide extremely high-quality context. And then so that sort of accuracy hasn't come as a surprise, and that's not... definitely not meant to be egotistical or anything. I just think graphs are amazing. When people find them, they just get put to such good powerful uses that all of this seems just much more pleasing than surprising in a way. I think where I will be surprised, yeah, I think where subsequent generations take this because I think I've reached as just a computer scientist, not an AI wizard, I think I've reached the limits of my current knowledge, and I'm really excited to see what other people do with this stuff now.

>> Yeah, and what's exciting is, too, we were chronicalizing this on theCUBE about a decade ago. We saw kind of the shift around multiple databases. We've heard the phrase, "No database rules the world. " Obviously, now diversity of databases because the environments changed. You have different architectures rolling out. But specifically generative AI is very much graph central because things are generative. Things have to be structured differently to get the latency, to get the security and also the context because graphs provide that great pathways. And most people think of neural pathways, neural nets, yeah, they could vector embeds on one side, but at the end of the day, you're starting to see graphs dominate. Could you share your vision on how people are using that today? Because the hype is high on GenAI. OpenAI just closed a 40 billion dollar financing. It's pretty massive. I mean, just every day, you're seeing more and more activity around that. So that's a big funding round. So that means there's more work going on, there's more investments. How are people... And there's enthusiasm is high and hype as well. But where's the confidence levels now? Could you share your vision on where graphs are being used and kind of where they extend to?

>> Yeah, of course. I mean, look, if you... you mentioned about the diversity of data. And this has been, I realize now, looking at the screen here, that I'm an old man. I've got bits of gray in my beard and my hair, which is shocking. When I started this, it felt like we were on the cusp of something really interesting in data. A decade and a half, two decades ago, data started to become really prominent with this idea of NoSQL and then big data. And that's kind of rolled in, I think, to today's wave in terms of AI. But I really think there's a couple of things going on broadly in the use of AI. And it's really a tale of kind of two cities. There are a bunch of folks out there who are really pushing the envelope with AI and they are very keen to take leading technology and push that into production to get value from it. And according to IBM, that's about 30% of folks working with AI. And then there's the other 70% of people in AI who are probably just as keen to experiment and get value from this, but who currently aren't getting that value. They're not getting dependable systems that they can push into production and get value from. And that's a huge difference, right, the kind of 70/30 split. And by no means all, but certainly, I think some of the reason that folks are getting value from their AI is because they're able to improve its accuracy. I converse with various AIs on a daily basis in my role as a computer scientist. And a lot of the times, those interactions are very valuable. I kind of get very good rubber duck kind of response from a lot of AIs, but sometimes they absolutely tell fibs. So I asked... a few weeks ago, I asked one of the famous AIs to summarize a paper that I know well and it actually gave me a 180-degree wrong answer. So it sort of makes you think, right. I mean, if I wasn't, you know, sort of very familiar with that literature, I would have believed it as well. I mean, I coined the term last year, the Boris Johnson problem, which is something that's expensively trained, speaks very confidently, and doesn't know when it's lying which is in reference, of course, to the former British Prime Minister Boris Johnson, who was caught out, shall we say.

>> Yeah. - Now, I think one of the ways that we can try

>> to solve that problem, and if you look at organizations, for example, like Microsoft, who are doing well at this, is to provide better up-to-date context so that our LLMs can produce better answers. And right now, it seems that the best way of doing that is by providing a context of a knowledge graph, which is chock- full of semantically rich information, business semantically rich information, your inventory, your customers, your products, your contracts, your invoices, all of that kind of stuff might be thrown in there. But importantly, all of that stuff is connected together. So compared to the previous attempts at injecting context into LLMs, which was done by vectorization, you'd say, "Are these two records near to each other," which is kind of high school level trigonometry. Now we can actually do something a bit richer, and we say, "Look, how are these things connected? Are they in the same neighborhood? How far away are they from each other in terms of hops in the graph? " And it turns out by using that structure, using the connections or the associativity in the data as well as the data itself, you get much, much better context and much, much higher accuracy from this. So it seems to me, and I probably wouldn't have been brave enough to say this a year ago, but it seems to be now that all RAG is tending towards Graph RAG and that Graph RAG is kind of eating this whole ecosystem.

>> Yeah, I mean, that's a great point. I mean, in the computer science side of it, vector embeds gets everyone's attention because it does really good search. Again, to your point, not 100% accurate, but it's better than keyword search. It's math-based. So as you see graphs emerge, can you share specifically what graphs are bringing to the table on that next step? Because RAG is certainly, people can see the benefits immediately better search, great for data.

>> Yeah. - It makes more compute available and more storage.

>> That's good for everyone in the industry. But where do the graphs shine? Can you... Are there use cases now, and what's the... where does it go?

>> Yeah, I mean, if you think about vectors, not to at all put the boot into vectors because they're extremely useful, but vectors kind of give you points in a vector space, and that's useful. And then we can kind of noodle on that a little bit and think, "Well, what can I... how can I semantically reason about these vectors? " And I hear people say, "Well, for example, if you've got the vector for king and you minus the vector for man, and you add the vector for woman, you get the vector for queen. " And I think that's not quite convincing to me. I think that depends an awful lot on the kind of embedding function that you're using. It really probably isn't the case. Really in a vector space, all we've got is that two points are near to each other. So if I ran approximate nearest neighbors, the kind of famous ANN algorithm to find my context tokens for a RAG arrangement, I get nearby tokens. And not to labor the example too much, but how do I know when I look for Apple that I'm not getting apple, the fruit, Apple, the tech company.

>> Yep. - I'm here in London, England, Apple,

>> the music company, if you're old enough to remember them because vectors... there's no guarantee that vectors will give you that in that kind of sub-symbolic relationship. Meanwhile, over here in graph world, that stuff's very clear, right. Because apple, the fruit will be connected into a sub-graph that's very much about farming and agriculture and fruits and all that stuff. Maybe its neighbors are lemons or oranges and they're connected in the graph explicitly in a way. Meanwhile, over here, Apple, the tech company, is going to be connected into place like Palo Alto and connected with its founders and it's going to be connected with its products. And Apple, the music company, again, will be connected explicitly with perhaps the Beatles and the albums they produced. All of this stuff becomes a lot clearer when you stop messing around with big long numbers and start thinking about symbolic representations of your data, which are both transparent for me, the engineer, the database person, as well as my... the people I have to do compliance and audit with. And very straightforward for an AI to work with.

>> Yeah, the computer science behind that's great. You're really talking about context, right. Apple's a fruit, and so if there's words around it, cues, if you will, this is kind of what ontologies couldn't crack the code on back in the 90s, if you remember. And this as you look at now, okay, you got scale, and you got compute, what's the impact? Because, obviously, that makes total sense because the graphs will have all this extra information around Apple or whatever the word is or whatever the context is. What's the... How does that factor into some of the compute? Because now we're hearing things like first token out on inference. Okay, I can reason vectors. I get that. Check that's really elementary. But when you start getting into really narrow path accuracy, context is important, but computes factor. The underlying infrastructure has to support it. Is there advantages on graphs besides the context? Does it have a compute benefit?

>> I think it has a... I think, and this is slightly kind of informed opinion rather than established facts, but I think because graphs tend to be very quick to query and very accurate, again, because of that symbolic representation, they tend to be quite lightweight in terms of their compute. So in a given interaction with an LLM, you can actually explore an awful lot of your knowledge graph and really do a good job of picking out the best connected context to then inject into the interaction with your LLM. If you're going to compare that to other kinds of databases in terms of compute, they tend to be more heavyweight. For example, relational databases have to do quite heavyweight joins. Whereas in graph world, we're just following those relationships, which in our implementation is just pointers. It turns out that pointers are the one thing that modern computers, that all digital computers are good at. They're terrible at everything else, but in terms of following pointers around the network, they're really fast at that. So, we can do a really fast, really high-quality traversal through a graph because it's just brilliant context to inject into the LLM, and it raises the quality of the overall system.

>> That's a great point on the pointers there. Pun intended. I got to ask you, let's go hypothetical for a second. Let's just say, for instance, I'm a customer. I want to implement a graph database on theCUBE. I got all these interviews. I want to put them in a graph. I got security videos, I got all kinds of context, or I'm a big corporation. What does the deployment look like? How do you guys engage? How do people deploy and manage graphs if they already got all the SQL-NoSQL, unstructured data laying around?

>> Well, you'd be following a reasonably well-trodden path, John. Thankfully, it turns out that you're not alone in thinking about building knowledge graphs of the lessons learned and the institutional knowledge that you have there. I've seen other companies do this. I've seen NASA do this in the US. I've seen militaries do this around the world because people want to learn and improve. I've even seen universities do this in terms of a citation graph and people who worked with people who wrote papers in subjects and so on. So the data model is naturally graphic. But of course if you've got a bunch of data that's already existing and it's spread around your existing ecosystem, then you've got a couple of choices. One choice which is very non-invasive, would be to use a graph kind of as a fancy index. So you leave the data where it is, but you create a curation layer on top of that data and you use the graph and graph queries to be able to explore into it very quickly. And then the leaf nodes of those graphs point to existing records in your document database or a row in your relational database or so on. And, of course, the flip side of that would be a kind of rip and replace. You decide that this knowledge graph is so valuable that you want to host some or all of that data directly within the graph itself because you intend for it to be both a transaction and analytical workhorse. And you're going to, of course, in this day and age, use the knowledge in that graph to feed the interactions with your LLMs.

>> Awesome. Jim, the five minutes we have left, I want to talk about what you're working on. What are the cool things you're working on from a computer scientist perspective? What's getting you excited? Where's the cutting edge? What's next on your plate? What are you optimizing for?

>> Well, I'm a full-tolerance person by training, but actually the thing that I'm most excited about in my team here at NEO4j is we're looking at completely novel ways of building databases. So, for example, if you look at Neo4j today and you look at, say, other good databases like MongoDB or Oracle or those kind of things, and you squint it a little bit, you'll find they're basically built from the same kind of patterns. It's because we've all read the same academic papers on how to build databases. So we have a query planner optimizer, we move tables around, would you believe, internally and that kind of stuff. But at Neo4j research, we're looking at taking programming languages as our inspiration because programming languages sort of have similar things going on. You take strings in at the top, and you produce executables out of the bottom. But instead of using planners and optimizers and all that kind of stuff, we have compilers and runtimes. So one of the things I'm very excited about is that we're looking at how to build a completely novel database architecture based on effectively leaning on compilers to do the equivalent of planning- replanning query optimization. And we're in the early stages of that. It's a UK government-funded project. They gave us the equivalent of a couple of million bucks to figure out this problem. And our early work is indicating that we might be able to build something that's an order of magnitude or maybe even a few orders of magnitude faster than the traditional way of building things. Still a few years to go in the actual research work, but we hope to make a significant impact down the line.

>> Final question for you on the AI infrastructure advancements. We're seeing a lot of semiconductor work. Nvidia just had GTC, AI factors being talked about, so there's a lot of work going on below you guys in terms of performance. What are you excited about? What's ready for prime time, as they say? What's the goodness going on? And where's the work areas that the infrastructure hardware guys have to work on? What would you say to that?

>> That's a very fair question,

>> and it actually ties into the thing I'm most excited about because in our world now that we lead on compilers to do all of the hard work for us, what I want from the infrastructure guys is simply good compilers for all of their new fancy electronics. And then the system that we're researching will be immediately able to take advantage of any of the novel exciting silicon that comes out, becomes a very straightforward proposition to dispatch to a fancy GPU or a fancy FPGA, or even if the cost is right, a plain old CPU. And we're looking forward to being able to have that multi- silicon capability in NEO4j.

>> Well, being a fault-tolerant guy that you are, you got the coherency challenges. I got a compiler for the compiler, I got AI for AI. I mean, is there a complexity loop there that you see, or is that not an issue at all?

>> I'm hoping that's not an issue at all, I think. In the later stages, I hope to see that the AIs will actually help us to develop particularly some of our efficient intermediate formats to help feed these compilers. So there's a lot of cool kind of nerdy stuff to be done. I'm very much looking forward to solving it.

>> Great to have you on. Good point there. I love it. I mean, I think AI will do that. You're seeing already code assistants being adopted, agents with authority and agency to do things. I mean, you just learn the compiler and speak compiler to each other, right.

>> Absolutely. - I mean, that's what's happening.

>> Jim, thank you so much for coming in.

>> I know it's a different time zone over there. Appreciate you taking the time to come in to theCUBE here in our new NYSE studio here, our build- out on the East Coast, our access point where we're kind of building a network here. I've obviously got a network effect going on, and we might have to put some graphs there around those date... around those nodes. So, thanks for sharing.

>> Sure thing. Thanks for having me, John.

>> All right. I'm John Furrier here at theCUBE.

>> We are at the NYSE... theCUBE at the NYSE Wired having ongoing programs around the key topics. This week, it's robotics. It's AI leaders who are doing the work, who are plowing the fields, blazing the trails for this next generation, infrastructure data, data platforms, databases, all kinds of activity happening that will flip the script on AI and hopefully usher in robotics and all the goodness that AI will enable. Thanks for watching.