Podcast: Play in new window | Download
Subscribe: Apple Podcasts | Spotify | Amazon Music | Android | Youtube Music | RSS
As the Cyc project winds down, one of its main contributors, Benjamin Rode, is attempting to migrate as much as possible of its upper ontology to RDF. It’s an ambitious project, much like Cyc itself.
Cyc arose from the early AI research community, and Ben has a fascinating perspective on that milieu as well.
We talked about:
- his work to port as much as possible of the Cyc upper ontology into RDF
- the surprisingly long history of neuro-symbolic AI
- some interesting details and anecdotes about the origins of the field of AI
- the deep intertwining of the histories of symbolic AI and computing itself
- Doug Lenat and the origins of the Cyc project and the Cycorp company
- Lenat’s concept of “white space knowledge,” the framing for unstructured natural language text
- how an attempt to port the Cyc ontology to RDF is “not as insane an exercise as it might sound at first”
- some parallels and distinction between the RDF and Cyc worlds
- the unique characteristics and capabilities of the CycL language, in particular its homoiconicity
- the central question for the current era: “can these two frameworks (symbolic AI and probabIlistic AI) work together in a synergistic way”
Ben’s bio
Ben Rode came to formal domain modeling and ontology engineering by way of Douglas Hofstadter’s Gödel, Escher, Bach: The Eternal Golden Braid, which he read in high school. His interest has since developed into study of machine learning, causal inference and induction, temporal reasoning, ontology evolution, and neurosymbolics. He holds a graduate degree in philosophy with philosophy of mind and analytic philosophy as areas-of-focus; the subject of his dissertation was the use of formalized contexts in common sense reasoning. He joined the technical staff at Cycorp in 1997, where he’s played an active role in developing the Cyc ontology for a number of contracts, including extensive work on database schema integration, ontology extension and mapping, inference development, and domain knowledge acquisition from subject matter experts, in addition to assisting with research on using the Cyc ontology for LLM-assisted formal knowledge capture. His current research interests include translating a subset of the upper Cyc ontology into RDF, large language model-assisted knowledge graph extension, and the use of knowledge bases for validation and verification of large language model output.
Connect with Ben online
- email: benjamin dot paul dot rode at gmail.com
Resources mentioned in this interview
- Cyc
- Automated Mathematician
- Heuretics: Theoretical and Experimental Study of Heuristic Rules
- Eurisko
Video
Here’s the video version of our conversation:
Podcast intro transcript
This is the Knowledge Graph Insights podcast, episode number 53. It turns out that contemporary explorations of hybrid AI are actually re-opening conversations that started almost 70 years ago. In the early days of AI, neural network “connectionists” and symbolic AI researchers saw their work as naturally complementary. Out of that primordial AI ecosystem emerged Doug Lenat’s Cyc project, an ambitious effort to account for all of humanity’s common-sense knowledge. Ben Rode is now trying to bring that work to the RDF world.
Interview transcript
Larry:
Hi, everyone. Welcome to episode number 53 of the Knowledge Graph Insights Podcast. I am really delighted today to welcome to the show, Ben Rode. Ben is a longtime ontologist, best known for his association with the Cycorp project, and we’ll talk a little bit about that and the Cyc Project in general. He’s also working on, and the reason I wanted to have him on the podcast today, he’s working on a project to port the Cyc attempt to account for common sense knowledge in the world into the RDF world that most of us are familiar with. So welcome, Ben. Tell the folks a little bit more about what you’re up to these days.
Ben:
Okay. So in the course of finding my way to a new position, I’m working on several projects of my own. One of them, as you mentioned, involves trying to port as much as can be ported from the upper Cyc ontology, particularly the 2014 version of that that is in the public domain. It’s available under public licensing on GitHub, to try to port as much of that as possible into the Resource Description Framework and the semantic web. And this has actually become an experimental interest of mine and I think it may be we can learn quite a lot from doing that.
Larry:
Yeah, that’s a good opportunity to remind me to ask you about the history of all this. And it goes back, you did a presentation a while back that I saw, where you set out the overall history of the development of the field of AI and how Cyc and other things fit into it. Can you talk a little bit about that?
Ben:
Yeah. Well, I mean, I think the framing context there is partly the question of, well, why would you want to translate the upper ontology of Cyc into RDF and what would that be good for? And perhaps there’s even a larger question there of what would Cyc or the semantic web be good for in the current AI context? Now, I mean, I think the story here really is maybe much older than a lot of people realize. I mean, the roots of AI generally go back quite far. And here I think it’s very important to distinguish between what we sometimes hear referred to as symbolic AI versus sub-symbolic. I think both the semantic web technology stack and Cyc can fairly be described as symbolic AI. Large language models are a good example of sub-symbolic AI, sometimes referred to as generative or probabilistic AI. Now, I mean, it’s really interesting that this really starts… For symbolic AI, it’s really quite early. I mean, it starts back at the end of the 19th century really with what we might term as math validation.
Ben:
What you have here is a set of concerns coming up in the mathematical community that the foundations of mathematics might not in fact be sound, that there might be hidden contradictions. And certain developments, spearheaded, for example, by Bertrand Russell, it led people to believe that the worst fears might be realized. So what you have developing in the late 19th and early 20th centuries is an effort to put mathematics on a sound foundation. And one of the things that comes out of this, not only is it symbolic AI, computing itself and digital computers and programming really have their origins within this thread. And then a little later than that, really starting up maybe in the early 1940s, you have information theory, which is coming with Claude Shannon and then associated with that is a movement that has been almost completely forgotten now and I think regrettably forgotten.
Ben:
It was called Cybernetics. If I remember right, the name originated with the Greek, hopefully I’m pronouncing this right, kybernētes, meaning steersman. I mean, the interest was in homeostatic systems, our systems where there was re-entrancy of information and energetic cycling to maintain homeostatic conditions. And there was a belief that this could tell us a lot of things, physics, biology, information, science. Some of the key names there are Norbert Wiener, Heinz von Foerster. Let’s see, I think it was, I believe Humberto Maturana and Francisco Varela and also a guy named Hans Jonas, who I’ve only recently found out about, and I’m starting to believe, beat everyone to the punch on that. But this is the thread that in some ways gives rise to generative AI. I mean, it comes through things that were called perceptrons that were pioneered by Warren McCullough, his student Walter Pitts.
Ben:
But an interesting aspect of it is that when you go back and look at some of these writings, particularly Norbert Wiener and Claude Shannon, one of the things that really comes up is they were talking in terms of complementarity. They were talking in terms of what we would call symbolic and sub-symbolic AI complimenting each other. Shannon was talking, for example, about auto-tuning versus code optimization. And if you think of auto-tuning roughly as being the machine learning, generative AI end and code optimization being the symbolic end, those were seen as being mutually supportive. And I think that is a feature we really want to… We’re at a point where we need to be reexamining that. I mean, it is being reexamined under the heading of very… We hear about hybrid systems or neurosymbolic AI and this really is it. We can talk some about what are the comparative strengths and weaknesses of symbolic and sub-symbolic AI in a little more detail if you like, but I think we-
Larry:
I love that too, that everything you just said goes back, because virtually everybody else I’ve talked to kind of marks that 1956 Dartmouth Conference as the dawn of AI. Well, that’s where McCarthy coined the term AI.
Ben:
That’s where McCarthy coined the name, yes.
Larry:
Yeah. But also it just occurred to me like, duh, that didn’t just come out of nowhere. And in fact, I talked to our mutual acquaintance, Pat Hayes, about the origin of that and he said, “Yeah, the whole notion, that coining of the term artificial intelligence was to distinguish it from cybernetics, to distinguish-”
Ben:
Well, no, exactly. And I mean, McCarthy at the time, 1956, he was sort of the young Turk and Norbert Wiener was the gray beard. Again, there may be people in the audience who know much more about this than I do, but based on the accounts I have read and what I have heard, McCarthy was kind of concerned that if Wiener was in that conference, Cybernetics were going to dominate and he didn’t want that to happen. He wanted it to be about symbolic AI. So folks, I mean, not only Wiener but McCullough and Pitts were kind of cut out of that discussion. And in some ways that may have been a little unfortunate. I think the history might’ve been very different if there had been more interaction and more communication between those communities.
Larry:
Yeah, because now it’s like what, 50 years later we’re circling back and I don’t know, probably, I don’t know, 10 or 20% of my podcasts over the last 51 episodes have in some way addressed, or probably most of them have talked about the neurosymbolic AI and treating it as a new thing, a relatively new thing, at least in the current architectures, is the need to rectify LLM shortcomings and to accelerate knowledge graph building.
Ben:
Yes, exactly. And you just identified, I think, what in the current context are the really important points. And let’s be clear, large language models are one of the most important things to happen in human cognitive studies I think in the last century at least. They are very, very important. And it’s not fair to say that… I mean, people talk about this as being on an auto-complete algorithm, which at some fundamental level, it is recursive auto complete with probabilistic next token extrapolation, but it turns out you can accomplish a heck of a lot more than anybody ever thought, than a lot of people thought you could. I mean, some visionaries I think did see the potential. In fact, if you go back to Claude Shannon’s where even in 1947, he was publishing what amounted to a version in his work on information theory, he was publishing what amounted to a version of that algorithm.
Ben:
But it is true that what you are getting, even when you are using it to get insightful and novel results is an output that in some sense it has a high probability of being plausible in the context of the training data and plausibility doesn’t guarantee truth, that’s one of the fundamental problems. The other one being that these systems are, in a sense, they’re holographic. The information that is driving the output is stored in this matrix of neural network weights, which makes it very difficult to interpret why the output is what it is. So recovering justifications that are actually actionable is a challenge. This is the explainabiliy problem.
Larry:
Yeah, sorry. That brings up the need for knowledge representation.
Ben:
Exactly.
Larry:
And that’s where I want to jump ahead and get a little bit into, so out of that milieu you just described, and you also just made the case or set up the situation for the modern neurosymbolic architectures. But before that, there was this whole development of the field of knowledge representation and semantic technologies. And the thing that you know most, I mean, you’ve been embedded in the Cyc community forever. Can you talk a little bit about how Cyc emerged from that milieu?
Ben:
Yeah, no, I mean, this really starts… So coming out of that Dartmouth conference you mentioned, there were a number of different ideas about knowledge representation that were starting to gel. I mean, four really essential names, Marvin Minsky, John McCarthy, Alan Newell and Herbert Simon were all in different ways thinking about what I think Newell and Simon called physical symbol systems, where essentially the idea was to create what amounted to, I mean, something almost like a high level programming language for representing knowledge so that inferences, including common sense inferences, could be treated almost as the execution of programs. I mean, this is very much particularly through McCarthy, this is entangled again, with the origins of programming itself, because one of the things that came out of this was the LISP programming language and the whole concept of S expressions on which that is built. And that was just an absolutely crucial development in the history of programming.
Ben:
So now we come to the figure of Doug Lenat, who was actually a student of both Minsky’s and McCarthy’s, and mentored by both men. And his idea was to take it one step further. And it’s actually really interesting to see how this happened. This started back in, I mean, I think it was 1976 when Doug published his PhD thesis. So that was something called the Automated Mathematician. And the idea was actually to build a symbolic program. It was based on heuristics for theorem discovery. This was intended to discover new mathematical theorems. I mean, we hear about this a lot about generative AI doing this these days. This was much earlier than that and it was symbolic, and it was using almost a kind of set of business rules for theorem discovery. And then in 1982, there was another publication from Douglas called I think Heuretics: Theoretical and Experimental Study of Heuristic Rules that tried to expand the automated mathematician to more open-ended heuristics.
Ben:
The upshot of that was a program called Eurisko. And again, this was not about theorem proving per se, but theorem discovery. But then Doug had an epiphany of sorts, I think. In 1984, he published a paper with J.C. Lee Brown that conceded that both AM and Eurisko were what Doug would later call ignorance-based systems. I mean, they were dependent on bespoke categories and rules that were kind of gaming discovery. In other words, in effect the system was being primed to follow paths that humans knew had been successful, but in a constrained or ad hoc sort of way. And Doug wanted more than that. He wanted general knowledge to be involved and he wanted a kind of representational evolution to be happening there, ontology evolution, if you like. And he came to believe that the only way to achieve this was to try to formally encode as much of human common sense knowledge as possible. And the upshot of that was Cyc and The Cyc Project, and what eventually became Cycorp. And that started up around 1984.
Ben:
A really important thing though, I mean to bring out about it that I think is often lost in discussions of the history of Cyc and that is, it’s clear to me at least the original intent of the system was to serve as a kind of set of guardrails and a development framework for machine learning. This isn’t completely unlike the way ontology-based knowledge graphs are now seen as providing guardrails for large language models, except in Doug’s mind, the guardrails needed to be there for the learning process itself, which was seen as operating continuously. And whether or not large language models are actually doing continuous learning as a whole is a topic for another day. In any case, in 1984, Bobby Inman, Admiral Bobby Inman recruited Doug to join the microelectronics and computer corporation in Austin, Texas. And that was a very early AI effort that the US government was supporting mainly through DARPA.
Ben:
I think it was intended to be a counterweight to Japan’s fifth generation project, which was another early AI effort. And the US government was somewhat worried the Japanese were going to get the drop on us in AI. So hence MCC, neither of those projects actually resulted in much, certainly not AGI, but The Cyc Project spun out of MCC in 1994 and incorporated as Cycorp three years before I joined.
Larry:
Nice. And so ’84, I mean the internet existed then, but that was about it.
Ben:
The dark ages. Yeah, no, I think it was still sort of in the, again, the chronology. I’m a little hazy on the chronology, I confess, but I think that was the ARPANET era. It was certainly not in widespread use.
Larry:
That sounds about right. Yeah. Well, and that’s so interesting. Well, I want to go a little bit into, so Doug has this vision of capturing human common sense knowledge, what big things don’t fit inside of small things. It’s just mind-bogglingly vast to try to even predict what that…
Ben:
Well, yes, it was a vast vision and this actually in some sense, it raises a specter that is maybe haunting knowledge graphs to this day I think, because you touched on what I think is the key point here and that is you can’t predict. The idea on the one hand was to capture as much of common sense knowledge as humans could capture really. I mean, Doug often talked of it in terms of white space knowledge. And what he meant by that was in reading any unstructured natural language text, what is the framing background knowledge that you need to bring to bear in order to correctly interpret that text? And the aim, and it was an incredibly ambitious one, was to try and capture all of that or as much as could reasonably be captured in Cyc. Now, whether or not that is a feasible enterprise is maybe the $64,000 question.
Ben:
But I mean, you can certainly do quite a lot with formal representations, especially if you confine yourself to delimited reasoning contexts or domains, but the problem of recognizing that you’re in a particular domain that requires a certain model to be applied and also critically recognizing that your situation is approaching, shall we say, the boundaries of one of those domains, because reality is always more inventive than human beings can be, is a challenge. And how much of that problem can be dealt with formally I think is still in some ways an open research issue. I will say right here, I am skeptical that it can be dealt with formally in its entirety, but this now maybe gets us into the whole question of what is the appropriate role for generative AI in connection with this and how can these two very different conceptions be used in a way that is synergistic?
Larry:
Yeah, I’d love to go a little bit more into that, but one thing I want to just quickly ask about is you mentioned that in ontology design and engineering now, there’s just this conventional wisdom that you constrain your ontology to a domain. You’re very clear on what you’re scoping it to. So that just seems like sort of a pragmatic practice way of like, okay, this is crazy. The only way we can manage this is by constraining it to a domain. Is that what’s going on there in terms of modern-
Ben:
Yeah, no, I think that is absolutely right. With the proviso that this, I mean, this opens up the whole question of, I guess you would call it cross-domain interoperability, which is it’s almost like the whole issue of trying to bring semantic layers to bear in doing federated querying across data silos, only now we’re not dealing with data silos, we’re dealing with knowledge silos. And here I think there is a role for upper ontologies, things like the Basic Formal Ontology that Barry Smith and Pierre Grenon created within the semantic web framework. I am firm in my belief and I think it’s a warranted belief that the upper ontology of Cyc can also help us here. And I think there are a lot of ideas embodied there.
Larry:
So is that the starting point for your attempt to get the Cyc lore into RDF?
Ben:
Yeah. Yes, it is. Absolutely. And it’s not as insane an exercise as it might sound at first. I mean, again, as I think the BFO, the basic formal ontology indicates. I mean, that actually says another critical piece of intellectual genealogy here, Pierre Grenon, who was one of the individuals instrumental in creating the BFO, also worked at Cycorp for a time. And I mean, I would say that looking at that ontology, particularly in the occurrent-continuant distinction there, I can see the intellectual heritage. It’s very clear.
Larry:
And in terms of timeline, this overlaps, it’s kind of like it starts way before the semantic web, but the people who contribute to the semantic web, some of them are involved in the Cyc Project.
Ben:
Yeah. No, it is really very interesting. I mean, probably the one name that is of most signal importance here might be RV Guha, who was Doug Lenat’s graduate student, if I’m not mistaken, and also his teaching assistant for a time at Stanford, where Doug was before he came to Texas. And he would go on to, he was instrumental in the creation of the Resource Description Framework. I mean, Pat Hayes also had affiliations with the project and Cyc itself was involved with some of the contracts that were connected with some of the DARPA efforts that fed into the creation of the semantic web. I mean, we were involved, as I recall, with DAML and OIL in some of these projects.
Larry:
Yeah, that’s right. DAML, that was it, Defense Agency Markup Language?
Ben:
Yes. Yeah.
Larry:
Yeah, and OIL, that was something… Yeah.
Ben:
I’ve forgotten what that acronym stands for, frankly.
Larry:
Right, but those were kind of predecessors for OWL, right?
Ben:
Yes. All of that I think was in the development light cone of OWL, if you like, the past light cone.
Larry:
Yeah. And that’s a whole other thing is getting… I mean, it’s one thing to have the knowledge in there, but then to build logic into the ways you work with that knowledge, that’s a whole other conversation. Hey, I want to go just a little bit. So those parallels with RDF and Cyc, as you embark on this project, is it just how knowledge representation works that leads to some commonalities? Or like you mentioned RV Guha and other connections like that, I’m just curious if you can stitch together a little bit more of the parallels between the RDF world and the Cyc world.
Ben:
Well, this is really interesting. So parallels and distinction, so it might be more fruitful to focus on distinctions first because I mean, they’re important, but they may not be as important as they might appear at first blush. So one of the critical thing with Cyc, as you might imagine for an undertaking that was intended to capture common sense, the representation language CycL, is extremely expressive. It’s important to distinguish there are two representation levels there. There’s the epistemic level, which is the level of ontological representation and the heuristic level, which is something like the implementation layer. And the epistemic level is very expressive indeed. I mean, in some sense, the expressivity I think is that of LISP itself, because the whole system was built on LISP. And so, among other things, two very important features that it has. Number one, and this is kind of I think sort of a heritage of S expressions. The notation itself was prefixed notation, which means that your operator, predicate, logical, connective, whatever is in the argv[0] position.
Ben:
That seems like a minor technical detail, but it’s very important, because it means that you can have predicates that have arity higher than two. So that’s a critical distinction from object properties and data properties in the semantic web, where the whole conception is predicated on graphs, and the object properties and data properties are essentially the labeled edges within the graph. So that’s one key feature. Beyond that though, there is the whole question of what can be an argument to what. And CycL is very permissive on this in some ways. It allows for quantification not only over individuals but over classes and predicates, and indeed expressions generally. So the term of art here is homoiconicity. It has the homoiconicity of LISP. Expressions can serve as arguments to other expressions, which means that you can have assertions about other assertions and you can write rules that are quantifying over assertions, which means at least in principle, you can have self-modifying representations with all of the complications that that implies.
Ben:
And in some sense, I get the feeling, although I’m not the person to really go into historical detail here, I think RDF was trying in some ways to stop that madness, to constrain the expressivity-
Larry:
I was just going to ask about that, because the way you describe it-
Ben:
I didn’t jack the semantics as badly. Yeah.
Larry:
Yeah, you can imagine the power that that entails, but I wonder if, like you just alluded to the simplicity of RDF, which is just subject predicate object, and that’s the units you have to work with. And then now with RDF 1.2 and RDF-star, you can look at triples as things-
Ben:
Well, see, here’s the thing, because I think as folks have found occasion for saying more and more complex things using RDF graphs, what’s happened is the system is getting extended. In some ways, one gets the impression it’s kind of grown like Topsy, but you now have extensions like RDF 1.1, which lets you deal with named graphs. So now you’ve got abstract individuals in the system that are effectively reifying entire graphs associated with that. And then I think RDF-star, if I’m not mistaken, now we’re getting into areas that I confess I’m only just learning about, but I know that there are extensions like RDF-star that allow for actual quoting of triples. So now you do have what are effectively assertions about assertions. And I’m not sure with some of these, I mean, with named graphs at the very least, the triple has to be somewhere in the default graph structure of your repository as I understand it.
Ben:
But with the quoting, I’m not sure even that is respected. You can talk about hypothetical or counterfactual or entailed triples. So now you really are starting to verge closer and closer, I think, to something that is almost like the lambda calculus, to something where you can potentially recreate both Russell’s paradox and Gödel’s Proof.
Larry:
And I’ll just mention that a few episodes, I think two or three episodes back, I had Ora Lassila and Adrian Gschwend on to talk about RDF 1.2 that talks about some of that. But hey, I can’t believe it, Ben, we’re already coming up on time.
Ben:
Okay. Yeah…
Larry:
But before we wrap up, I just want to make sure, is there anything you want, any loose ends you want to tie up from the conversation or anything you want to make sure we share?
Ben:
Oh, the loose ends. I mean, again, I guess one thing I would circle back to is just the importance of complementarity. And I mean, we’ve got these two very, very powerful frameworks, paradigms. I don’t know even what’s the best name to use, but I mean, on the one hand, you have symbolic AI, which is really formal logic. I mean, essentially what the reasoning is very much like some would say actually is program execution. I mean, justifications are audible, I’m sorry, auditable and debuggable and provided the reasoning domain is circumscribed, truth of premises guarantees truth of conclusions, but the more complex it gets and particularly the more self-referential it gets, the more complicated the semantics become and the more expensive the reasoning becomes, frankly. And reasoning and you’ve got this problem we mentioned that reality is always more inventive than we are, which means you often have to revisit the underlying ontology.
Ben:
I mean, development is not an entirely front-loaded process and adjusting the ontology over time is hard. And then NL understanding and generation is a whole other problem we haven’t talked about, but it involves a lexical translation layer, so that’s an additional complication. Sub-symbolic AI understands NL in a very real sense. I mean, this is something that effectively is passing some approximation of The Turing Test. It’s flexible. It can deal with probabilities. It’s in some degree adaptive, questionable how adaptive, but that’s another, again, that’s a topic for another conversation, but it’s very good at generating useful hypotheses and extrapolations if you know what you’re doing. But again, as I said before, you’re getting plausibility, you’re not getting truth.
Ben:
The central question is, can these two frameworks work together in a synergistic way where, I mean, in some sense it’s almost a kind of generative… what I’m envisaging at least as a kind of generative and test loop. On the one hand, LLMs are playing, it plays to the strength of LLMs, which is in some sense hypothesis generation and to the strength of technical formal symbolic AI, which is analysis and maybe writing herd on empirical testing programs. So the question of what the architecture for supporting this would look like is an open one and it’s a research project that I’m really hoping to support and be a part of if I can.
Larry:
Yeah, I think you won’t be alone if that’s any consolation. I know a lot of folks involved in that. Hey, one very last thing, Ben. If folks want to connect with you or follow you online, what’s the best place to find you?
Ben:
Well, you can find me on LinkedIn. I mean, just look for Benjamin Rode, last name spelled R-O-D-E. So I’m there. You can also reach me at benjamin.paul, P-A-U-L, .rode, R-O-D-E, @gmail.com.
Larry:
Cool. I’ll put that in the show notes as well, of course. Well, thank you so much, Ben. This is really eye-opening and I really appreciate your insights.
Ben:
Well, thanks. I’m really glad to have this opportunity to talk about these issues.
