Podcast: Play in new window | Download
Subscribe: Apple Podcasts | Spotify | Amazon Music | Android | Youtube Music | RSS
Much of the conversation around AI architectures lately is about neuro-symbolic systems that combine neural-network learning tech like LLMs and symbolic AI like knowledge graphs.
Frank van Harmelen’s research has followed this path, but he puts all of his AI research in the larger context of how these technical systems can best support people.
While some in the AI world seek to replace humans with machines, Frank focuses on AI systems that collaborate effectively with people.
We talked about:
- his role as a professor of AI at the Vrije Universiteit in Amsterdam
- how rapid change in the AI world has affected the 10-year, €20-million Hybrid Intelligence Centre research he oversees
- the focus of his research on the hybrid combination of human and machine intelligence
- how the introduction of conversational interfaces has advance AI-human collaboration
- a few of the benefits of hybrid human-AI collaboration
- the importance of a shared worldview in any collaborative effort
- the role of the psychological concept of “theory of mind” in hybrid human-AI systems
- the emergence of neuro-symbolic solutions
- how he helps his students see the differences between systems 1 and 2 thinking and its relevance in AI systems
- his role in establishing the foundations of the semantic web
- the challenges of running a program that spans seven universities and employs dozens of faculty and PhD students
- some examples of use cases for hybrid AI-human systems
- his take on agentic AI, and the importance of humans in agent systems
- some classic research on multi-agent computer systems
- the four research challenges – collaboration, adaptation, responsibility, and explainability – they are tackling in their hybrid intelligence research
- his take on the different approaches to AI in Europe, the US, and China
- the matrix structure he uses to allocate people and resources to three key research areas: problems, solutions, and evaluation
- his belief that “AI is there to collaborate with people and not to replace us”
Frank’s bio
Since 2000 Frank van Harmelen has played a leading role in the development of the Semantic Web. He is a co-designer of the Web Ontology Language OWL, which has become a worldwide standard. He co-authored the first academic textbook of the field, and was one of the architects of Sesame, an RDF storage and retrieval engine, which is in wide academic and industrial use. This work received the 10-year impact award at the International Semantic Web Conference. Linked Open Data and Knowledge Graphs are important spin-offs from this work.
Since 2020, Frank is is scientific director of the Hybrid Intelligence Centre, where 50 PhD students and as many faculty members from 7 Dutch universities investigate AI systems that collaborate with people instead of replacing them.
The large scale of modern knowledge graphs that contain hundreds of millions of entities and relationships (made possible partly by the work of Van Harmelen and his team) opened the door to combine these symbolic knowledge representations with machine learning. Since 2018, Frank has pivoted his research group from purely symbolic Knowledge Representation to Neuro-Symbolic forms of AI.
Connect with Frank online
Video
Here’s the video version of our conversation:
Podcast intro transcript
This is the Knowledge Graph Insights podcast, episode number 33. As the AI landscape has evolved over the past few years, hybrid architectures that combine LLMs, knowledge graphs, and other AI technology have become the norm. Frank van Harmelen argues that the ultimate hybrid system must also include humans. He’s running a 10-year, €20 million research program in the Netherlands to explore exactly this. His Hybrid Intelligence Centre investigates AI systems that collaborate with people instead of replacing them.
Interview transcript
Larry:
Hi, everyone. Welcome to episode number 33 of the Knowledge Graph Insights podcast. I am really delighted today to welcome to the show Frank van Harmelen. Frank is a professor of AI at the Vrije Universiteit in Amsterdam, that’s the Free University in Amsterdam. He’s also the PI of this big program called the Hybrid Intelligence Center, which spans seven Dutch universities, multimillion euro grant over 10 years. Welcome, Frank. Tell the folks a little bit more about what you’re up to these days?
Frank:
All right. This Hybrid Intelligence Center occupies me most of the time, and that’s been a very exciting ride over the past five years. We’re just at the midpoint and we have five more years to go.
Larry:
Nice. How is it going? Are you satisfied? Are the expectations of the grantors being met and are you happy with the progress you’re making?
Frank:
Yes. It’s obvious to say that the world of AI is super dynamic now. All kinds of things have happened in the past few years in AI that nobody had predicted when we started, the rise of large language models of conversational AI. That has also really affected the notion of hybrid intelligence. It’s been an even more exciting ride than we had expected.
Larry:
Yeah. That’s right. Yeah. I think excitement is the word of the day. Hey, one thing I have to observe, earlier today before we recorded this, I was doing a presentation with some information architects, and the subject I was talking about hybrid AI architectures and neuro-symbolic loops and all this stuff. One of the people in the presentation asked, “Hey, what about human AI? Shouldn’t that be the architecture?” Then I said, “You’re going to love my next podcast guest,” because that’s the whole point of this hybrid intelligence idea, right?
Frank:
Yeah. The core idea of hybrid intelligence, hybrid standing for hybrid combination of human and machine intelligence. Think of hybrid teams, where a hybrid team is made up a bunch of people and a bunch of AIs who collaborate to get a task done. That, if you want, the tagline of the Hybrid Intelligence Center is that we’re working on AI that collaborates with people instead of replacing them. If you work on AI systems that collaborate with people, then you certainly need to solve all kinds of different problems and answer all kinds of different questions than where you are thinking about AI in the replacement mode.
Larry:
Yeah. That seems to be, like in a lot of circles, there’s this assumption that AI is just here to replace people, but you’ve been… Long before that was a meme and people talking about it, you were working on this hybrid concept. Has that heightened the urgency around your work, the current state of AI expectations?
Frank:
It has heightened the urgency, and it has also opened all kinds of doors. One of the big hurdles in AI-human collaboration, say five years ago, was really the conversational interface. It was hard to talk to AI systems, and they certainly wouldn’t talk back to you in a coherent way. Well, we all know that’s now a solved problem. But what happens in the middle is the real challenge. We don’t think that the large language models are going to solve all of the collaboration between humans and AI systems. We want our AI systems to do things that the language models are not very good at, but we’re using that technique in a kind of sandwich model. Now, the language model does the conversation on the front end, it does the conversation on the back end, and we’re working on the AI agents, the smart that’s in the middle, to create these hybrid teams.
Larry:
As you say that I’m thinking about that’s just one aspect of the hybridization of this. That that’s one way that humans… When you think about hybrid architectures, where LLMs can help build knowledge graphs and they can also fill in knowledge gaps in LLM architectures. What other obvious complimentary things are there between… What do humans need help with and what do machines need help with?
Frank:
Right. There are some obvious things like the perfect memory that machines have and the imperfect memory that we have. Okay? That’s a nice example of where members in the team can really compensate for each other’s strengths and weaknesses. Humans suffer from a whole host of these cognitive biases. For example, we suffer from the recency effect. We believe information more if we’ve heard it recently rather than when we’ve heard it in the past. We believe information more when we’ve heard it more frequently rather than… There’s no reason to believe something more if you hear it more often.
Frank:
That doesn’t make it more true, but it’s how our brain works. Not always such a good idea. Computers can help us to compensate for all of these cognitive limitations. Conversely, we are very much aware of the context in which we operate. We are aware why we are doing something. We are aware of the implicit norms and values that govern the task that we’re doing, that we’re expected to obey in a particular group to perform a particular task. Computers don’t have any sense of why they are doing something, the context in which they’re doing it, the social and ethical norms under which they should operate. That’s something where the human component can compensate for the machine limitations. These are just a few examples of that complementarity.
Larry:
Yeah. That’s one of the things I think about a lot is that what we call in my world stakeholder alignment or stakeholder discovery or working with subject matter experts to make explicit their tacit knowledge in their head and things like that. It seems like that’s probably always or mostly going to be a human capability. Is that… You probably have research that backs this up, right?
Frank:
Well, and if you want to collaborate with a computer, then you better make sure that there is some alignment between you and the computer. We can only collaborate because we share some of the way we look at the world. You need this shared worldview in order to collaborate with somebody. You don’t need to have a completely fully aligned worldview, but one of our use cases is in the operating theater in a hospital. All the people who are standing around the patient, they have to some extent aligned their view of what’s going on. That allows them to collaborate.
Frank:
We then also need the worldview of the computer to be aligned with how the surgeon looks at the world. The computer has to be aware of the goals and the expectations and the intentions of the surgeon in order to respond properly. These are the things that we take for granted when humans collaborate, that we have an aligned worldview, that we are aware of each other’s goals and intentions. That I can speculate about what you are thinking, so I can anticipate. I can explain to you what I’m doing. All of these things we take for granted between people, and they suddenly become really challenging if the team is a hybrid team where some of the agents are AIs.
Larry:
That’s so interesting to think about the onboarding your machine colleagues to operating room culture. How does that manifest in the work you’re doing?
Frank:
A concrete example is… It’s what psychologists call a theory of mind. It’s what people have a theory about somebody else’s mind. When we have a conversation, I’m assuming you know all kinds of things. I’m not going to explain some things to you because I know you already know them, so I know that you know them. You are not surprised that I don’t explain them to you because you know that I know that you know this stuff. Okay? We have this mutual knowledge about each other. That’s what psychologists call the theory of mind.
Frank:
My colleagues in the Hybrid Intelligence Center, one’s in Groningen and in Leiden, they are working on computational theory of mind where we can equip a computer with that ability to reason about different layers of knowledge that agents have about each other. They perform experiments where they can show that even in different settings, in a competitive setting or in a collaborative setting, computers with a theory of mind perform better than computers without a theory of mind. That’s just a very concrete example of a new research question that you suddenly have to start asking because you’re thinking about collaborative AI rather than replacement AI.
Larry:
Yeah. I just today was listening to a podcast with a developmental psychologist talking about how children develop theory of mind and how that it’s… Is that a learning, like a neural network kind of thing? Or is that more like knowledge that you can impart with a knowledge representation mode?
Frank:
Right. You ask is it one or the other? The answer is probably yes. Right?
Larry:
I get it. Yeah.
Frank:
I think in many of these problems, and I think this theory of mind is one, we are actually gravitating towards neuro-symbolic solutions where part of the system is doing neural-based learning and part of the system is doing explicit logical reasoning, knowledge graphs, ontologies, they all play a role there. But there is also the reasoning by analogy, which we could learn by neural learning algorithms. These neuro-symbolic architectures are a recurring theme in the center.
Larry:
Yeah. I think it was… I’ve watched too many videos, but I think it was you who was talking about, and a number of people have mentioned Kahneman’s Thinking, Fast and Slow, the systems one and two thinking.
Frank:
Yes.
Larry:
Is that a sound analogy or is that just Kahneman’s so well-known and all these dynamics.
Frank:
No.
Larry:
But it seems to hold up pretty well.
Frank:
I think that’s a… It’s contested and people are now doing this sort of, “But that’s not really what Kahneman meant,” right? But as a first order approximation, I think it works really well. You can almost feel it in your head. With my students, I give them a picture of the American president. I don’t need to know the answer. I just want to know how long it takes you to recognize him. Okay. Click. Okay. Okay. That’s Donald Trump. Now it takes him a 10th of a second.
Frank:
Okay. Then you think, okay, who was his predecessor? That takes a little bit longer. Right? Then you think, who was his predecessor’s predecessor’s vice president? Okay? Then you just feel another part of your brain clicking in. It’s no longer the neural activity that instantly recognizes a face, but you need to start reasoning and remembering and the predecessor of the predecessor and then their vice president. It’s this logical thinking and fact retrieval and combining it. That’s the neuro-symbolic architecture, the system one, system two at work in your head. I think that, as a first order approximation, I think that’s a very plausible architecture also for AI systems.
Larry:
The way you just described it, it sounds extremely plausible. That’s a perfect example of that, the time to recognize and associate concepts like that, and the way it’s so fun to think about. Your background is in computer science and then in AI, right?
Frank:
Yes. Actually, for me, AI is simply the most exciting part of computer science. Right? That’s where all the fun problems are. I was lucky to be a very early member of the Semantic Web movement. Everything we now call knowledge graphs was running together with Dieter Fensel recently passed away, sadly, running the first European Semantic web project around 2000, helping to define the web ontology language, the OWL standard, working on these knowledge graphs, working on technology to make them scalable. That’s my background, really core work on knowledge graphs. That technology has now matured so much that all kinds of companies, companies big and small, are absorbing it, are marketing it. Then it’s time for the academics to take up their tents and move somewhere else to new problems and embedding these knowledge graphs inside these neuro-symbolic systems to build these collaborative agents. I think that’s a very natural flow of the content.
Larry:
Yeah. Part of the reason I ask that question is it’s probably because you’ve been exposed to so many use cases around it, but you use so many actual brain anatomy examples in presentations you’ve done, and you were talking earlier about the operating theater. I just was not assuming, but I’m wondering, did you also study medicine somewhere along the way or something?
Frank:
No. In my next life I’ll be a neuroscientist, but not in this one.
Larry:
Okay. Nice. Hey, so I love just this concept of the notion of building these hybrid teams, but at a meta level, you’re building the teams that make these teams the creation of this whole new infrastructure possible. Tell me a little bit about that? How you run this program?
Frank:
It’s a challenge. There are seven universities across the country. We’ve now employed 50 PhD students across these seven universities, and an equal amount of faculty as their supervisor. That’s a crowd of 100 people. We really work hard at keeping the thing coherent. Every PhD student has two supervisors from different universities, often from different disciplines, supervised by a linguist and a computer scientist, or by a psychologist and an AI researcher. Together we also work on a number of we call them case studies, so really applications of this hybrid intelligence teaming. The operating theater that I mentioned was not just a random example. We are actually working on an operating theater application of hybrid intelligence.
Larry:
Are there certain applications that are better test beds for what you… An operating room seems like really interesting place to do this, but are there others that are logical and obvious?
Frank:
Yes. Another example that we’re running is actually the one that we started with is very close to home, build an assistant for scientists. Build a colleague. We have a team of a bunch of PhD students and post-docs and senior scientists. Why not inject a few agents into the team? I would like to come into my office on Monday morning and hear a voice say, “Hi, Frank. How was the weekend? By the way, I read another 1500 papers over the weekend, and we really have to have a conversation.” I cannot read 1500 papers, but the AI can. That’s one of the other use cases is build a scientific assistant. Another use case is for, let’s say, ordinary citizens, a lifestyle support for diabetes patients. The diabetes patient, there is a whole host of these apps on your phone and on your watch that try to nudge diabetes patients into a more healthy lifestyle.
Frank:
Regular body motion, regular eating, et cetera, but the traditional apps are always focused only on the patient themselves. Where actually the best way to influence the patient is to influence their environment. Make sure that the family doctor is part of the conversation. Make sure that the neighborhood nurse is part of the conversation. Make sure that the family members are part of the conversation. Have these conversations, these different nudges of behavior, this influencing the lifestyle, have that coordinated by an AI agent. Then you inject an AI agent into the social environment of a diabetes patient as a way to help them improve their lifestyle and avoid the worst of the disease that they’re suffering.
Larry:
Nice. That first example you gave reminded me of the famous semantic web paper in Scientific American. The opening anecdote is talking about an agent, I think it was more like a travel thing. But you said, “Hey, can you just read the 1500 papers that came out overnight and summarize them for me?” The articulation of agents as what we call these things that we’re collaborating with, is that… Because there’s all this talk of now… I don’t even know exactly what it means in the current discourse around what agents actually are and do, but it sounds like you have a really well-developed and articulated idea of what an agent is.
Frank:
Yes. Sometimes I get a little bit annoyed by this term, agentic AI now as if it is something new. Come on, guys. Read the literature. Right? This idea goes back to the late ’80s, early ’90s. Like you already mentioned, Jim Handler, Ora Lassila, and Tim Berners-Lee’s paper on Semantic Web services with agents. That idea has been around, and I think by now we have accumulated so much technology that we can actually make this technically work. Much of the classical literature of multi agents has been about communities of agents that talk to each other.
Frank:
If you think about, if you hear the language model, large language model crowd talk about agentic AI, it’s about agents talking to each other. There’s no human in the loop there. We are really taking the stance that, no, it should be a hybrid team where it’s not just a bunch of agents going off and do their thing. It’s a hybrid team of humans and computers that compensates for each other’s strengths and weaknesses and solve it through collaboration, which I think is an essentially different take from both the classical multi-agent view and also the modern agentic view.
Larry:
That’s it, because the modern agentic view seems to presume that AGI is right around the corner and we just have to corral it, which to me is absurd. But tell me a little bit more about the older concept of agents? Because like so many things, there’s probably a lot of really good insight that we can derive from that and should probably be applying.
Frank:
Yeah. One of the classical models of multi-agent research was called the BDI model. BDI stands for belief, desires, intentions. There were programming languages and architectures based on equipping artificial agents with beliefs about the world, desires that they were equipped with, goals that they had, and then combining the beliefs about the world with the goals that they had, combining that into intentions. Intentions to do actions. This BDI architecture, I think that’s still a very valid way of thinking about how to program agents that collaborate with … that’s also about how we think about each other. We think, okay, yeah, my collaborator has certain beliefs about the world. I know we have certain goals, the desires, and therefore we have certain intentions to do stuff. Therefore, I can predict what her needs are, and I can predict how to help her best. This BDI architecture is something from the past that is still very valid in the modern research.
Larry:
It sounds like it has some foundations in that theory of mind stuff we talked about earlier that-
Frank:
Absolutely.
Larry:
… operationalizing those insights that come out of that.
Frank:
Absolutely.
Larry:
Yeah. Hey, one thing I want to make sure we get to is you… I don’t know. One of the papers you shared with me, it’s five years old. I remember as I read it thinking… I was surprised when I looked down and saw the year because I thought it was contemporary. But so I’m wondering how well the four research challenges that you anticipated in that paper, are they the actual challenges you’re facing or have different things come up? Can you talk a little bit about that?
Frank:
Yes. The four research challenges, we summarize them in the acronym CARE, right? The C for collaboration. The A for adaptation. The R for responsible behavior. The E for explainability. Right? Together CARE. Those are all things that you would expect in a collaboration. You would expect it to be collaborative, adaptive to the circumstances, responsible behavior in a team, and you should be able to explain your behavior to the team members. I think maybe the hardest one has been the one about responsibility because this notion of ethical behavior, of responsible behavior, of having a shared set of norms that governs the behavior in the team, that has proved quite elusive to make that computationally concrete.
Frank:
I think if we’re stuck on any of them, I think that’s the one that we find hardest to tackle. The world is making a lot of progress on explainable AI. Adaptive AI there, reinforcement learning is a big ingredient. Learning from interactions in the world. If the world changes, you learn from the changes in the world and you learn to adapt. Collaboration is at the heart. I think we’re making progress on all of those, but the responsible and ethical and shared-norm behavior, I think that’s been maybe the hardest one.
Larry:
That’s really… Now, as I’m going to ask you to put on your PI hat for a minute, and is that in your responsibility of administering this 10-year grant? Are you like, “Well, maybe we need to shift a little bit more to the responsible part”? I’m just imagining how you might operate.
Frank:
Yeah. No. Certainly I’m convinced that this is an important part of AI. I’m also very happy that this is at the center of the AI debate in Europe. I think that’s what sets European AI apart from the debate in America, which is mostly dominated by big tech, and the debate in China, which is mostly dominated by a very centralist government. I think this notion of responsibility and responsible behavior of AI systems is… Or to European AI, actually, it’s a strength of European AI. Just because we found it hard, it’s certainly not a reason to drop it. Absolutely not.
Larry:
No. I’m sure the hardest stuff must be the most probably rewarding. Yeah. Like the, I don’t know, tempered steel or something that’ll come out of that. Yeah. Now I’m wondering too about just in terms of the allocation thing of this research, I saw you have that really interesting cross… These concerns crossing a grid about how you… You mentioned earlier how you match up your PhD students with different advisors. Can you talk a little bit about that intersection of concerns and how you manage that?
Frank:
Yeah. We call this our matrix structure, and it’s an intersection of three different areas. One area is what we call our problem space. Right? Those are the four challenges that I just mentioned. The problems are how to collaborate, how to adapt, how to behave responsibly, how to explain. That’s our problem space. Right? Then we have our solution space. That’s where all the AI techniques live. Right? That’s where reinforcement learning lives. That’s where neuro-symbolic architecture is, natural language understanding, computational theory of mind. Those are the techniques. Then we have the problem space. We have the solution space. Then a third, we have the evaluation space where we measure how well our solutions, our proposals live up to practical circumstances. That’s where we work with a surgeon in the hospital, where we work with family doctors and specialists on diabetes, where we work with legal experts on the collaboration and law and where we work with scientists to build a scientific assistant. Those three.
Larry:
Interesting. Now I’m going to have to circle back in five years and see how all this culminates. But hey, Frank, I can’t believe we’re coming up close to time already. But before we wrap up, is there anything last… Anything you want to revisit from the conversation or just make sure that we share before we close?
Frank:
Yeah. I would want to emphasize that this is… This collaborative AI, as far as I’m concerned, that’s not just a small niche topic that some people do in a corner somewhere. Right? I think this collaborative AI, these hybrid teams, I think that should be the way we think about AI in general. It’s how we should think about large language models. It’s not to build automated teams of agentic AI that zoom off and then do their own thing and come back with a solution we no longer understand. I think this idea of hybrid intelligence, which is a combination of neural learning, symbolic representations, knowledge graphs, ontologies, the whole lot, that combination in order to collaborate. I think that should really be the way all of AI should be done to keep in mind that AI is there to collaborate with people and not to replace us.
Larry:
Well, I would vote to put you on the board at OpenAI and Anthropic and all those places. Yeah. No.
Frank:
I’ll take it. I’ll take it.
Larry:
Exactly. No. But seriously, I think no, that’s great. It’s not like… My main mission with these podcasts is to share practices and principles, but if we can do a little evangelism about the importance of humans in AI architectures, I will do that till I die. Yeah. Well, hey, thanks so much, Frank. One very last thing, Frank. If folks want to connect with you or follow you online, what’s the best place to stay in touch?
Frank:
Best way to stay in touch is go to our website, Hybrid Intelligence… Just Google hybridintelligencecenter.nl, .com, .org, whatever. All of them work. Go to the website of Hybrid Intelligence Center, and that’s where you’ll find us, what we’re doing. If you want to know more, you’ll also find ways to get in touch with us.
Larry:
Fantastic. Well, thank you so much, Frank. I really enjoyed this conversation.
Frank:
Right. It was a joy to talk. Thanks for the opportunity.
