Moving AI Beyond Black Box Answers With Neo4j

[00:00:00] Are you still dealing with agentic AI semantic confusion? Well, Denodo keeps your business harmonised. And by positioning your data platform into a universal semantic layer, everybody can be on the same page. So get back to scaling your business and trusting your AI agents to create proper business decisions. And you can do that by visiting denodo.com to see how.

[00:00:26] What if the difference between an AI answer and useful intelligence actually lies in the breadcrumbs that it remembers? The context. Well, returning to Tech Talks Daily today is Jim Weber, Chief Scientist at Neo4j. I'm excited to get him back on today, especially following its recent acquisition of

[00:00:53] GraphAware and the Hume platform. But today, Jim is going to explain why agents need the story behind the good and bad decisions and why graphs can connect knowledge, conversations and decision traces that feel almost like layers of SimCity. Anyone remember that series? And we will also discuss Neo4j's move from graph database company to a graph intelligence platform,

[00:01:21] explainable AI, and why governments are reconsidering where sensitive data lives and who controls it. And Jim will also address where human oversight must remain involved, especially when automated conclusions could affect somebody's liberty or even safety. So today you can expect graph theory, geopolitics, agent swarms, and a warning that even convincing

[00:01:48] AI can drop an absolute clangor. And I suspect we've all seen that at some point. But enough for me, let's get Jim back onto the podcast now. So a massive warm welcome back to the show, Jim. Can you tell everyone listening a little about who you are and what you do? Sure thing, Neil. Thank you for inviting me back on. I had a lovely chat last time, so I'm quite happy to be on again. So hi, everyone. My name's Jim Weber. I am a fellow Midlander to Neil,

[00:02:17] so I'm a black country boy. But for the last 16, nearly 17 years, I've been the chief scientist at Neo4j. We're a graph database company, or a graph intelligence platform company now. And my job is to make sure that things keep working even when the computers underneath them get a bit crotchety, let's say. Oh, so it's a pleasure to have you back on, especially because you've been incredibly busy. I mean, Neo4j has always been known as a graph database company. But in recent weeks, you've also

[00:02:47] had an acquisition that feels much like a, or feels like a much bigger statement. So tell me, tell everyone listening a little bit more about this acquisition and why now is the right time to reposition Neo4j as an intelligence platform rather than simply an infrastructure provider. Yeah, I mean, there's nothing simple about infrastructure, Neil. You know that. But I think there's a confluence of things that have come together. So in the last few weeks,

[00:03:13] we announced that we've acquired GraphAware, which is a company that has been a partner of Neo4j forever. In fact, I encouraged their CEO when he finished his master's degree at Imperial to get into graphs because I promised him there was a lucrative future there. Now, the lucrative future has taken a bit longer than maybe you would have hoped, but he's had his company acquired now. So that's lovely for him. And what GraphAware have been doing from their early beginnings as a consulting company,

[00:03:38] they've kind of pivoted into this kind of specialized domain of intelligence analysis. And it turns out that they're really good at understanding what the kind of things that police forces, governments, intelligence agencies, and so on need. And they've always built that on top of Neo4j because Neo4j is a really good database when you want to connect facts together and then explore those facts on mass or zoom into individual, you know, kind of interesting

[00:04:06] snippets of that data. And we've been friends with these folks like forever. As I mentioned, I've been friends with their CEO. I used to sublet office space from him when Neo4j was too poor to afford its own offices back in the day. So we've known them for ages and we've worked together for so long on so many of these kind of joint go-to-market efforts. And with the way that Neo4j has developed, with the way that GraphAware's market has developed, so I think with the way the market has developed generally, it just felt like the right time to press the button and bring them

[00:04:35] from being super close friends into being actual family. So that's all happened now. It's all done and dusted. And Neo4j and GraphAware are now one company and we're able to sell an intelligence platform called Hume directly on top and all we are wonderfully tightly integrated with Neo4j nowadays. Awesome. Exciting times indeed. And there's so many big buzzwords this year around AI and agentic AI, but one in particular that stands out, I think, is context. It's something I hear in

[00:05:05] now. And I was reading before you joined me again today that you've said that context is becoming almost that missing ingredient in AI and we have no shortage of data today. So what do organisations still lack that prevents AI from producing intelligence rather than just simply generating answers? Because it feels like we're almost at a tipping point moment and we're taking the next step. It doesn't it. So I think you're right. There's no shortage of data, but that data tends to be answers.

[00:05:36] So, you know, I can go look for a particular HR policy or I can go look for a set of quarterly results or that kind of stuff. And those facts are there. And of course, I wouldn't be true to myself if I didn't say those facts are definitely best stored as a knowledge graph, right? I think you'd expect me to say that, you know, other knowledge graphs are available in true BBC parlance, but I think they're a really good way of storing those facts. But I think what we're learning now as we move into an agentic world and we want our agents to be

[00:06:03] dependable is that facts, the endpoints are really useful, but they're not sufficient. And it turns out that we've known about this for a long time. Back in the day, I don't know if you've ever read a book called Bad Science by a medic called Ben Goldacre. Yeah. And it's a lovely book, right? And the short version of it is when you do science, if you have a good result, you should publish it. And people do. When you do science and you have a bad result, you should also publish it, which people don't. Particularly in his world, in pharma,

[00:06:32] a bad piece of science kind of dings your share price, right? So that was the kind of thesis behind bad science. But actually, if you kind of imbibe that, agents want the full science. They want the good and the bad. They want to know when they've made a good decision. They want to be able to record that in the context of the facts they have available and the conversations that they're having. And they want to also record the bad outcomes, right? Which is something that we're not used to

[00:06:58] doing, especially as humans. We're not used to recording why the good outcomes. We simply record the outcomes. But now we want all of the breadcrumbs that got us to a good decision and all of the breadcrumbs that got us to a bad decision. So we can do reinforcement learning and get more of the good decisions and fewer of the bad decisions. And unsurprisingly, because I'm a graph head, you might guess that each of these layers, the knowledge graph, the contextual layers around the decision traces and the conversational data,

[00:07:27] they're all really well represented as graphs. And they converge really nicely. They compose really nicely. I think about them, you know, we're both old geezers now, but I think about them as like layers in SimCity. You can pick a particular layer and work on it and use that layer in isolation, or you can bring all of the layers into playing and query and understand and reason across all of those layers. And it's a really good way of feeding your AI the data it needs,

[00:07:54] sorry, your data it needs to process and make good decisions. I think certainly it would be foolish of me to make a long-term prediction on this, but for the medium term, it feels like that's the pattern that all of the kind of smart AI folks are gravitating towards. And for as well, I think it has an incredible effect on us as simple infrastructure providers. Now I'll use that pejorative that you labelled me with, you know, because it gives us an opportunity to bring our skills into the mix and help

[00:08:19] the AI people who aren't necessarily useful, valuable systems. I must admit you had me at SimCities there. It takes me way back. I was 50-50 whether I said overhead transparency, but then you just lost half of your audience at that point. And if we zoom out for a moment, I think the acquisition of GraphAware positions Neo4j, it feels like almost an alternative to platforms like Palantir Gotham. That's what immediately sprung to

[00:08:48] mind for me there. But without turning this into a feature comparison, what is it that you believe governments and enterprises are asking for today based on your conversations you're having there? And especially things that weren't a priority even five years ago, because so much has changed over the last few years. Yeah. So look, I think there's utmost respect for the Palantir platform from our side. We see

[00:09:13] the kind of impact that those folks can have, but any competition is healthy. I think Palantir has been so dominant and there really hasn't been any competition there. And if we ask, how would you compete against such a successful behemoth? Where would you even begin? Whereas for us, we begin by saying, well, look, we've got a different engine, so we can offer different capabilities compared to Palantir. In fact, because we're a Graph, we're often a lot faster. We're also smaller and nimbler.

[00:09:41] We spend time with our customers and we want to build the features in the platform that they want to use. I think that in a general sense is timeless, right? That's just about being a good competitor in a kind of naturally functioning ecosystem. But you also can't ignore the geopolitics here, right? People are now, particularly those of us outside of the US, we're thinking more broadly

[00:10:06] about how to keep our systems dependable. And that is a wider question than it used to be five years ago, right? Yeah, our cables to the US could be cut by an enemy nation. And then suddenly your data is on one side of the planet, you're on the other. America itself could have local policies that make it, for example, the American cloud providers aren't going to be compatible with your own security and privacy requirements. So you might want to actually weirdly bring some of that stuff on shore

[00:10:34] or even in-house. And I think more broadly that, you know, even if we disregard the ick with the Palantir's kind of quite insane pamphlet that they publish, which is a terrible misstep, I think, in a commercial sense to project your politics into that space. Even if we ignore that, you know, we're giving people an opportunity to say, look, it's your data. You put your data where you want it. And that's quite a different motion from Palantir, which is, oh, thank you for the data.

[00:11:02] It's ours now. And we shall do with it as we see fit. But I think people are wise to that now. People are like, well, actually, some of this data, I think, is value in kind of giving to my platform. But also, I want to be able to take it back when I want. And I think that's something that we offer, the kind of flexibility of deployment. The fact that your data just sits in a database and you can export it, delete it, do what you want with it. And we have nothing to say about that. It gives people a lot of confidence. And then when you layer the tooling on top,

[00:11:29] I think it's a really compelling offering for those bodies, governments, police forces, intelligence forces, and so on, that are thinking smartly about the governance of their data and the use of their data. So I think it's going to be fun from a kind of competitive level. But these other geopolitics and so on is going to play a certainly going to play a part as well. Words I never would have expected to say three to five years ago. But we are where we are, as the proverb goes.

[00:11:54] Yeah, 100%. It's so refreshing to hear you talk of this approach and this alternative too. So AI, though, if we look at it from afar, often criticised for being just a black box, particularly in regulated industries and government. So how important is explainability now becoming? And should organisations trust an AI recommendation if they can't understand how it reached its conclusion in the first place?

[00:12:21] I mean, yeah, the strict answer to that is no, right? So, you know, let's think about the models that we're all using and to a greater or lesser extent getting value from. They're stochastic models, right? But they don't know about your circumstance. They don't know about my circumstance. They've been trained on basically scraping the internet. So they are savants. They've got incredible powers for producing excellent natural language and all that kind of stuff. But they're left to their own

[00:12:49] devices. They don't have any idea about your business. I think that's the important thing. I think of a model as being a lossy compression of the internet, right? And what you want to do for the bits of data that are important to you, you want to restore that lossy and make it lossless. And that's where the rag pattern came along and then later the graph rag pattern, right? So you're introducing your own data as part of the flow when you're interacting with a model.

[00:13:13] And what we're now seeing is that that kind of pattern can significantly increase accuracy. And actually, there was a scholarly article from the National Innovation Centre for Data here in the UK up in Newcastle. And they demonstrated that the graph rag actually significantly increases accuracy and reduces cost because you use fewer but better quality tokens as part of this flow.

[00:13:39] And so what we're doing now is we're approaching kind of an asymptote where we're able to get good answers in individual interactions with systems. But systemically, we're going to grow past that. Systemically, as we talked about earlier, we're going to end up having agents that live for a long time. And those agents are going to reflect, or at least from a human point of view, I don't want to anthropomorphize them too much, but they're kind of going to reflect on what's gone well and what

[00:14:05] hasn't. And they're going to reflect on that in terms of recency. Oh, these things went well recently. I'll try and do more of them. These things went badly recently. I'll try and do less of them. And that's the way that we're going to incrementally improve the quality of the agents. Now, do they have to be 100%? I think that depends on your situation. Please don't run your nuclear power plants on agents that aren't safety critical. I think we all know that as systems

[00:14:30] people. But for a lot of work that we're doing, we're starting to see agentic systems approach and surpass the quality of human systems. Because remember, we're not fallible either. I'd like to think of myself as more than a stochastic producer of words. But there's an element of that locked in here. You prompt me with a question, and I flood you with the answers that I've been trained on. And I'd like to think I'm smarter than that. But ultimately, what counts is the outcomes.

[00:14:56] And if we're now starting to see these more dependable agents that have their context with them, the three layers of context memory, so conversation, decisions, and knowledge, they're able to work for the long haul and improve over time. Now, what I particularly like about the approach that we advocate for is under the covers, we can still use deterministic algorithms to refine the underlying knowledge graph. We can page rank it or cluster it or find neighborhoods

[00:15:25] and weak links and all that kind of graph theory stuff, that mathematics that's been around for 300 years. So we get like a top down and bottom up approach where the agents are recording what's going well and not going well. And we're also improving the quality of the data underneath them. I think that's what's going to move us forward to a system of agents where for a given level of quality, we can start to match and surpass human systems.

[00:15:49] And we are hearing more and more about digital sovereignty, particularly as governments and large enterprises alike begin reassessing their dependence on overseas technology providers. It's a topic I hear frequently across Europe and press briefings, etc. It's something that keeps coming up. But how have you seen that conversation change? And why has sovereignty become such a strategic issue rather than just an IT decision?

[00:16:16] Yeah, again, this is I think partially geopolitics, right? The world has changed in weird ways and old stabilities that we thought were there for decades are now changing. You know, we have war in Europe again. So is my data safer in Europe? Or should my data be in Japan or America, which is out of range of that kind of conflict? So this is a complex thing. I think partially it's geopolitics. Partially, as I mentioned again, you know, there are missteps, you know, the ick that we got from the

[00:16:44] Palantir pamphlet. That wasn't just nerds like me listening to that. There were policymakers who read that and thought, I'm not convinced that that is something that I want to be part of. And then, you know, by association, you think, well, okay, what other stuff am I utterly dependent on, but over which I have no control? And in a world with so many moving pieces, I think you want to start to control, like, like, just like an enterprise architect would do, but kind of, you know,

[00:17:10] prime ministerial level, your policy to policymakers start to think about what do I depend on that I don't have control over? Who can pull the rug from under my feet? How likely is it that this, you know, this company or this country will want to pull the rug? And then you start to make some rational prioritized decisions about what, if anything, you re-onshore or you indeed bring back in house. I think that's all being prompted by the, as the Chinese were alleged to have said, but actually didn't, the interesting times that we live in.

[00:17:40] Love it. And if we look at everything we've talked about today, from policing and cyber defense to tax authorities and defense organizations, I think one thing that they do all have in common is they involve incredibly high stakes decisions. So for everything that you've seen here, what have these real world deployments could talk you about where AI is genuinely helping analysts and where human expertise must remain part of the decision-making process? It's probably a question you get asked a lot,

[00:18:10] but from everything you've seen and heard and experienced there, what have you seen? It's a spectrum, right? Yeah. So look, if I'm going to order your arrest because some orchestra of agents has decided that you're a bad person and you've broken laws or whatever, I think I still want human oversight of that. Even if the human's the dumbest link, it just feels that's a moral imperative, right? You don't take actions straight through processing with agents that could be wrong because sometimes humans can

[00:18:38] spot things where the agent's dropped an absolute clangor and the agent can't spot it. I have the same thing in my own work. I work on fault-tolerant computing. Today, the agent told me, oh, this committed minority will be a problem for you. And in fact, in the protocol I work on, the only thing that can commit is a majority. And it was just absolutely wrong, but it was convincing if you were a layperson, you would have read that and said, oh, absolutely, we must therefore delete the code. So I do think in certain things, you always want human in the loop. I think humans want that as

[00:19:07] well, right? I can't imagine Prime Minister Burnham devolving that kind of level of decision-making to the cleverest computer in the world. But equally, I've seen examples, even in historically conservative or cautious sectors like banking, where certain parts of their business are now doing straight through processing through agents. I was at the Yao conference in Australia last year. My friend and former colleague, Scott Shaw, who's at Commonwealth Bank in Australia, was talking about

[00:19:36] how they're actually doing straight through processing now with agents for certain parts of their banking operations, because they feel they've got the right safeguards in place and the infrastructure in place to do it. So they're actually delighting their customers more than they would with kind of your old-fashioned rule-based engines or even with human-based stuff. So I think most of us sit somewhere on that spectrum between humans have to be involved because it's super important and people will get hurt, right the way

[00:20:01] through to, for this part of my business, this can be automated with good monitoring and governance and all that. But this can be automated with agent or agents. And after the infamous ROI struggles a few years ago, it does feel now that enterprise technology companies are moving beyond just providing infrastructure and instead taking responsibility for business outcomes, which feels like a big step forward too. So is this what you're seeing though? Is this where you're seeing the industry heading?

[00:20:31] And how does graph intelligence help organisations move from just reporting what happened to understanding what should happen next and improving business outcomes, adding extra value, etc.? Yeah, I mean, you're right. It's certainly the way we're heading, right? And as graph people, we're quite used to this. It turns out that graph theory, this 300-year-old branch of mathematics, has already been pretty good at making predictions. So if I can spot a pattern in a graph and I can,

[00:20:57] for example, create another edge, that might bring you into a terrorist cell, for example, and now you're on my hit list. So we're quite used to this. But the notion that we can actually get agents to do that for us, so we can not only bring the best of things like machine learning, like graph neural networks and so on, but we can get the agents to kind of agent speed, not human speed, agent speed, to go over bits of our graph infrastructure and make decisions. It would have seemed like science fiction three years ago, but I think it's becoming remarkably normal now.

[00:21:26] And I think the only thing that is much more up for debate is the thing we just discussed. Those results, are they immediately executed or are they surfaced for human intelligence, for a kind of human-in-the-loop style model? I think that's the way that we're definitely heading. And if we were to look ahead, if we have this conversation in, what, three years' time, don't worry, I'm not going to make you wait three years before we talk again, but if we did fast forward three years into the future, will analysts spend their days searching for

[00:21:54] information or AI and graph intelligence fundamentally change how investigations and decisions and collaboration happen across government and enterprise? Because it feels like there is so much opportunity here, but I'm curious how you see it all evolving. And I know we can't predict the future, but I'm curious where you see it heading. I'd say what I think. I think there's going to be a real confusion about the word agent, particularly in intelligence. Because you think about it, what are my agents working on?

[00:22:22] It's like, well, do you mean the human ones or the robots? Because I think there are going to be a vast deployment of automated agents of the style that we've discussed, right? The ones that are dependable in production for the long haul, the ones that improve over time, because they have an ability to tirelessly just pour over data, right? And I think we're going to see a rapid expansion of that. You see this a little bit already, the kind of more advanced folks are already quite happy with this pattern. And like everything, right? It's going to, you know, the kind of, the folks

[00:22:51] that are kind of, the alphas, they're already out there doing this, and it's going to trickle through to the rest of us. So if I was an intelligence agent, a human one, I would be guiding my swarm of robotic agents over the data that's available for me, feeding me with pertinent stuff, deep diving here, blah, blah, blah, making and connecting facts here, so that I, as a human, can then act in the best possible way. And I think that synergy for this kind of work anyway, is where we're going to be in three years, I hope. I think, you know, there are probably, we don't know, of course, because

[00:23:20] they operate in secret, there are probably already intelligence agencies around the planet that are moving in this direction, because it feels like, as a systems person, that's the best kind of architecture for this kind of system. As for the rest of us, you know, those of us in banking and finance and retail and all this other good stuff, I don't think we're far behind. I think a lot of the things where today, we are bound up in rules, static rules that come from policy documents, or static rules that are enforced by a Python program, or what have you, that's going to become

[00:23:50] accessible to the agents. And over time, the agents are going to find when it's actually pertinent to bend or break those rules, document them to the humans, but then get on with the rest of their life under the new set of assumptions, transient, of course, making better decisions and giving better business outcomes for the folks that are users of the systems. So I think this is the way that software architecture or systems architecture as a whole is going to go. I suspect it will be the users of things like, you know, GraphAware, Hume and so on that are going

[00:24:20] to be at the forefront of this, because it just falls really nicely in their niche, but the rest of us aren't going to be far behind. Well, it's so refreshing to hear you talk about the end of this era of black box intelligence systems, and governments beginning to ask how they can responsibly deploy AI without understanding how conclusions are reached and understanding that statement there. So that's a

[00:24:45] big talking point. And yourselves at Neo4j, it's repositioning from a database company to an intelligence platform company. It's a big trend right across there. And I suspect there'll be a lot of people listening wanting to carry on this conversation we started today. So for those people listening, where would you like me to point them? For sure. Look, if you are interested at all about Neo4j and the kind of the data technology and the intelligent graph intelligence platform tech,

[00:25:10] please come to our website, Neo4j.com. You could always hit me up via the socials. I am at jimweber.org on blue sky. I don't use the Elon site anymore. It got a bit weird, gave me the ick, but please hit me up. I'm more than happy to have a conversation with you. And you can always drop me an email. I'm just jim at Neo4j.com. Privilege of being the first Jim at the company all those thousands of years ago. Love it. And I love this next generation of intelligence analysis,

[00:25:39] how it's finally moving beyond dashboards. Nobody wants just another dashboard. And one of the most important things we've discussed as well, of course, is how conclusions are reached, why context is becoming that missing ingredient in AI and how connected data is helping governments and organizations are like in highly regulated sectors. Finally, move from information retrieval to intelligence. Great points there. I will include links to everything that you mentioned. I encourage

[00:26:07] people listening to carry on this conversation. Exciting times ahead. And it's just a pleasure as always to speak with you. And regardless of what I said earlier, I'm not going to leave you waiting three years. We'll get you back on early next year and see how things are evolving. We have to test these predictions, Neil, right? You have to see how accurate my predictive capability is versus some electronic agent that can do it better. I think Jim left us with a useful distinction there between finding facts and understanding how

[00:26:35] they connect. AI agents can process information, remember decisions and improve. But dependable intelligence that requires context, explainability and control over data. And this is such an important point to make. And the closer decisions move to someone's safety, liberty or livelihood, that is a much

[00:27:00] stronger case for human review. So I enjoy Jim's image of an analyst guiding a swarm of agents. And although it might feel like science fiction, so much of today's technology did just a few years ago. So it's clear where we're heading here. And touching on Neo4j's acquisition of GraphAware, also showing infrastructure providers moving closer to the decisions that their technology support.

[00:27:28] So a massive thank you to Jim Weber for returning. Remember, you can find Neo4j at Neo4j.com. You can find Jim on BlueSky. I will include links to everything there. But over to you. Would you trust an AI conclusion if you couldn't inspect the path behind it and it didn't understand the context? Fairly obvious answer there. But let me know your thoughts. TechTalksNetwork.com. But that's it for today. Thanks for listening. See you tomorrow. Bye for now.