transcribe

Daniel Davis: Grounding Generative AI with Context Graphs | Episode 49

Knowledge Graph Insights · 38m · transcribed May 2026
More from Knowledge Graph Insights Business
𝕏 Share ▶ YouTube 📥 PDF 🤖 .md

Transcript

0:01 This is the Knowledger Graph Insights podcast, episode number 49. When Foundation Capital published their article about the trillion dollar opportunity presented by Context Graphs, many people were hearing about the concept for the first time. Not Daniel Davis. He's been developing an open-source context graph platform since 2023. His work in complex domains like aircraft safety and autonomous vehicles as well as his study of quantum mechanics have led him to explore ways to ground probabilistic AI systems in logic and knowledge.

0:35 >> Welcome to the knowledge graph insights podcast where experts on knowledge graphs and the semantic web share their wisdom with the digital and business communities. Our mission is to demystify and democratize knowledge graph technology and practices to make the benefits of knowledge graphs and semantic technology accessible to all. And now here's your host, Larry Swanson. >> Hi everyone, welcome to episode number 49 of the Knowledger Graph Insights podcast. I am really delighted today to welcome to the show Daniel Davis. Uh Daniel's the co-founder and co-creator uh at Trustgraph, which is an open source software project that builds graph stuff that we'll talk about today.

1:13 uh based in San Francisco. And welcome to the show, Daniel. Tell the folks a little bit more about what you're up to these days. >> Oh wow, Larry, that's a that's a lot to unpack there. I mean, how much time do you have? Uh yes, I am the co-creator of Trust Graph with Mark Adams, who is a bit more well known in the graph community than me, but uh uh he likes doing he he's not he likes building graphs. He doesn't like talking about them so much. And uh I'm confident that he would agree with me on that. Although I am trying to make his cat Fred famous because I'm actually working on a um new video on um our guide to understanding RDF, which is something that a lot of people have asked us about and how Mark taught me RDF so many years ago by uh with three simple sentences about his cat Fred. But um trust graph uh is what we've been working on for the past few years now. And we've had a couple of different ways of trying to explain it to people whether it's a context operating system, context development platform. Um some might even think of it like a context science uh platform which uh I think is kind of an interesting analogy as well. But u I myself have quite a diverse background. Spent a lot of time in um DoD aerospace. Came out to Silicon Valley almost 10 years ago to work in the autonomous vehicle industry.

2:33 um focusing on cyber security and safety and that's why I write articles about things like determinism and information risk and trying to attribute value to information. But in that world, I also was doing complex knowledge work where you read one document that's 800 pages long and then you have to read a statement that references another document or maybe it references 12 other documents and you just keep you know tracing down this chain of references and you have to understand which one of these documents actually takes precedence. Why do these statements conflict with each other? Do they conflict with each other? how do I try to come to some sort of opinion about this and and and in the safety critical world opinions aren't allowed. It's not like um auditing for enterprises that you can have opinions. Uh they take a much grimmer view on that and that's where that word determinism comes in and whether determinism means what people think it means. And how is that for an introduction?

3:35 >> Well, that's it's perfect because it sets up all the things we want to talk about. The first thing I want to talk about I think well it's it's so hard to to choose but of one of the reason you came to my attention is I forget somebody oh my friend Yen in in um in Munich brought you to my attention and I was like whoa this guy's been talking about context well before December of 2025 which is when apparently the rest of the world started thinking about context and context grass. How did you tell me a little bit about maybe the story of your connecting with Vicki or however I mean that combination we were talking before we were on the air about your experience with autonomous vehicles discovering Vicki and his interest in in um in uh context graphs and then ever a lot of what you just said is like a reason to need context graphs to do the stuff you want to do. So maybe talk a little bit about your journey into this the context realm. Well, so much of this comes from the problem I was trying to solve in the autonomous vehicle world.

4:31 You know, this is work that I've been doing for years in DoD aerospace with risk management and cyber security and safety and just running complex programs is it's so much about the paperwork and how you make decisions, how you justify those decisions, how you comply with regulations, understanding the regulations. And for autonomous vehicles, the scope was just unprecedented when you look at the number of things that could go wrong. And I we could literally talk for the next like few days just me rattling off scenarios and you'll go like wow I never thought of that. I never thought of that. Wow. Wow. And you you just start going like how do you manage this?

5:15 And well that was kind of what I was um sought out. That's what I was having to solve and um looking at all the different ways of doing this and trying to combine a basian approach with risk management and realizing the data sets were going to be huge and how do you manage that and it kind of turned out to be an unsolvable problem at the time and around that time because I was I was working at Lyft I got brought up to manage a lot of the issues that were going on with theft actual IPO which again more regulatory stuff with the SEC and how processes are applied across the entire enterprise, how these comply with SEC regulations and expectations and how this is audited. Uh and just even how we were measuring our cyber security performance as a company, how that was getting reported to the board. Again, very similar problem just, you know, slightly different problem space, slightly smaller scope. And around that time um Mark's company Trust Networks was actually acquired by Lyft and um I met him and uh I kind of got introduced more to um graphs and knowledge graphs.

6:27 I actually hadn't even worked with knowledge graphs you know prior to that you know I was much more in deterministic structures and you know dod aerospace. I was the one always saying like why are we writing in this Python? I should we should write it all in ADA. and all the people would just look at me and go, "What is ADA?" And I just kind of I would do that just as a joke, but also partially believing it. I still advocate ADA. I like ADA even if it makes developers cry. It was designed to make developers cry because it always works. But that's another story. Um, and that was that was back in what 2018, 2019. And you know, Mark and I have worked together off and on since we both left on various projects. And you know around I guess it would have been 2023 or so. I guess it was when Chad Gh when Chad GPT first came out in GPT3 and we started realizing hey this technology is actually going to be useful and I I saw it as a way to solve these problems that I had had throughout my entire career.

7:31 you know, having information you're trying to glean across thousands upon thousands upon thousands of pages, you know, the ability to read all this and digest it down to exactly what you need to know. But we also knew that how would then you explain the decision making process, how do you know where you got the data from? And it was pretty obvious that LLMs and being neuron nets and again my background and having to deal with determinism and provability for systems. Well, you just know that these these are inherently probabilistic systems and you're not going to be able to prove in a court of law, which we know is why this stuff really matters is when the lawsuits fly, is somebody's not going to be able to testify on why X occurred because neural network went through all these pathways and there were weights and I don't know how it ended up over there and ultimately it's just not going to scale to that. So it needs these structures, these graph structures, which of course is more of the neuros symbolic AI approach, which kind of brings me back to um kind of today. So let's let's fast forward late 2025 and um it's actually pretty interesting how it all got started is um I guess um Foundation Capital wrote the trillion dollar opportunity article, you know, you know about context graphs.

8:56 I totally missed that. Didn't know anything about that. It was my friend uh Kirk Marple and >> dropped on Christmas Eve pretty much. So you might have been I don't know. Yeah. >> Well, no. Kirk Well, no, because Kirk wrote his because Kirk dropped his article on Christmas Eve. So that's how I found out about it is Kirk had wrote an article in response to it and dropped it on Christmas Eve and he sent me a message going, "What is going on here?"

9:24 and his article had already gotten over a 100,000 views and I checked back like an hour or so later and now it's over 200,000 views and we're like what is going on? Why? Like so we kind of it was during the holidays and I I would have thought is this really a good time to launch an article? So I thought well let's try something. Let's try an experiment. So I decided to write what I called the context graph manifesto and I published it on Christmas. No, no, no.

9:54 New Year's Eve, actually. And it did really well. Did really well, too. And you know, that's when there was just so many other people dog piling on the topic. And it kind of just for a couple weeks, you know, people were now complaining about um I'm tired of seeing about, you know, context graphs. I actually had the the CEO and like co-founder or founder of Chromma DB like whining in my comments and I was like, "Oh, gee, I wonder why you're unhappy about this. you you you're the CEO of a vector database company. Um so I'm like gee come on can it be more obvious? I mean like and like don't come like we actually Trustcroft deploys quadrant. We use that for databases. I don't know why you're upset. Um and it was quite shocking you know that this term all of a sudden came out of nowhere and and then a lot of people started to try to take credit for who created it. And I know a lot of people were sort of like, wait a minute, we've been working on this 3 years. No, we've been working on this 5 years. And that's when I saw that, you know, Vicky Froen had given that talk in 2019 about context graphs.

11:01 And I actually watched his talk and I went, "Wait, he's actually talking about what we've been doing." Wait, wait, wait. He deserves the credit for this, although he doesn't agree with that. So that's why, you know, I then, you know, had a conversation with him. And actually um Vicki and I chat all the time uh now and which there may be more to talk about that a little bit later as well depending on how some of those conversations go. So um but but we think here's what happened because we also have friends at Accenture McKenzie um the people that are really know what's going on with enterprises because enterprises pay these people to be told what to think. Let's be honest. And we've heard from multiples within these organizations that around November to December, all of a sudden that's all they were hearing from enterprises. They were talking about context, context, context. And I'm relatively certain what happened was 2025 was the year that all the AI PC's failed. And what are they doing in November or December? They're doing their year- end reviews and they're going, "Why did this fail? Why didn't it work?" And they're obviously coming to the same conclusion that this AI technology LLM need very specialized specific context to be able to do these very specialized task and workflows. This very specialized knowledge work like that. I used to do understanding mill handbook 516 516 for military airworthiness for fixedwing aircraft and all the hundreds of standards that come or uh come under its umbrella. very very very very specialized, you know, uh, terms. Like I even did a an article and video just about one term within that called the modified rhyme test and what that means and getting down to like the semantics of that because that was my job was understanding what the modified rhyme test was. And um we also now know that 2026 is the year that all major enterprises are doing AI readiness. So, we're kind of preparing for what's coming in 2027 when context systems will actually start to be rolled out in these enterprises and what that's actually going to look like because uh we we really really really strongly believe they're going to look very different than what a lot of people are calling the data 3.0 lakehouse era of the modern enterprise data estate.

13:45 >> Interesting. So that so much of what you're talking about there's um I know well I know several people who've been doing like neuros symbolic stuff for many years like Yan Osman at uh Algoraph and and Fronzen and and and people like that but but it's been sort of like very particular industries and and use cases I think but now it sounds like every enterprise in the world is is concerned about this and knows that they need to integrate this now. Um so and and you you mentioned so many things earlier about all the um like regulatory compliance um u crazy corpuses of manuals and stuff to work through. Um tell me how like how so I guess maybe disambiguate for me like what are like what not a regular but a knowledge graph from a context graph. maybe start there like how a context graph differs or or is it just a kind of knowledge graph or how would you describe that?

14:41 >> I've tried to define it with a really simple definition in that it's a graph structure that is optimized for AI usage. I do think there are really three different layers to it which is very very very closely aligned with Vickiy's philosophy going all the way back to 2019. I think the semantics of how we describe it though I think is evolving a little bit. So to my mind there's three parts to a context graph which if you look back at like the foundation article cap foundation capital article they were focusing purely on what they were calling decision traces which I would call kind of that middle layer of explanability and provenence that sits on top of I know some people called it a system of intelligence a system of grounding a system of knowledge. So I I would think of that that's your kind of initial graph structure. You know you've taken or you think about it from a basian perspective because we love you know the basian approach is that that's your prior that's your knowledge prior so that you you've you've initiated your system with knowledge. Okay. Then as the knowledge gets used or context uh I guess context is the better word at this point as a context gets consumed and turned into actions and workflows you know that's that system of records again which is you know how that initial context is being used what are the results of that which again basian approach this is how the basian network evolves is that you made some priors and then you have that feedback loop. I think the top layer though is the one that's most important and I I haven't decided how to really describe that. Um it's kind of a a feedback layer, but it's more like it's kind of like a supervisory layer in that it's it's kind of what we as humans do every day. It's that comparison of how does this compare to something I've known in the past or I've observed this before. Is it now the same? And you're constantly updating your knowledge, your internal context to this new information. And according in our opinion all of these things working together or a context graph system because we're also big proponents of this should be done in a single graph system. That's the power of graphs is when you have all this information in a single graph so that you can find those relationships because a lot of people have argued no no you want to have lots of small graphs and um we really don't understand that perspective and we think that just comes from a lack of experience with really scalable graph systems.

17:38 >> Yeah. Quick question on that you mentioned because I think I can't remember if it was an article or but you talked about the lay the layers. Are you talking about each all of those layers represented in a single graph? Like you talked about the generated intelligence. You were just talking about the system of the records um and the system of initial intelligence. Is that the initial intelligence that's like incoming data and and knowledge or tell me I guess maybe maybe walk us through each of the layers because that seems and are they all included in the in the one graph because that's really interesting architecture to me.

18:12 >> Yes. And the way that we do it. Yes, all in a single graph. That's the trust graph way of doing this. Now, we do have a lot of c we do have a lot of graph management capabilities that you can create logical separations in the graph. Uh we have even where you can create these modular units called context cores that you can just take and pull out and put over here somewhere and load it back in. But those are also that's not just a graph unit that's also has the map the vector embeddings which is needed for the retrieval process. So, and some of this is actually coming down to kind of our implementation of some ontologies too because um we now use the W3C prob uh ontology for the basis of that system of records. So, for instance, creating that initial context prior. I kind of like that.

19:06 Maybe I'll start going with that. The context prior. Let's get really basic. >> Okay. Yeah. Nice. >> The your context prior. Yeah. because we are big proponents of of basian um actually that's one of the things I've been talking a lot with Vicki lately about basian belief networks which is some really interesting topics um is take my example of military airworthiness um despite being an expert in it I once upon a time was I would never be able I would never say I know all this stuff in my head it's impossible It's hundreds of thousands of pages of expertise. Matter of fact, when um that particular standard is kind of is kind of uh broken down into about a dozen or so different domains, we actually had subject matter experts for each domain and I was just kind of managing all of those because it was so one of the reasons I talk about the modified rhyme test is that was one of on one of the programs I was running.

20:07 That was one of the things we were actually testing and you know I'm looking at these 25page test procedures for just how talking about dip thongs and how different kinds of speakers say different kind of things and I'm going wow we're testing at this level of granularity for this like of course we need somebody that's just a subject matter expert in this one topic. So there's no way one person can have all this context to be able to perform these jobs. So creating those contexts prior to be finding all those knowledge sources and ingesting them in the system to get that structure so that you have all those knowledge priors necessary to understand a topic like how do you demonstrate an airplane is airworthy safe to fly and by the way that was only for fixedwing aircraft rotary craft helicopters etc had their own process because it's that different from just fixed wing to which you're probably wondering why my head didn't explode and I you know it came close a couple of times.

21:15 >> Well, my hope is that your context graph saved your brain because you offloaded some of that cognitive labor to it. U >> well I didn't have it back then though so I just had to suffer. >> Yeah. Well, okay. So suffering was >> what what that meant is what that meant is you know you have like 12 different PDFs open and you're copying pasting into spreadsheets and you're trying to remember what came from what document and what can't and then you're trying to figure out how do I explain all this to somebody and then how do I do I trust all the people that are giving me their opinions on on their expertise and it's um it's a very much a very difficult you know human problem to to coales all of this into ultimately making a single decision. It was coming down to a single decision by a single individual. You know, will I sign that this airplane is safe to fly?

22:11 >> Yeah. Now, I'm going back to the the emergence of this like your collaboration with Mark because your background, you've got all this really interesting background, but but Mark was really the the I guess the catalyst for for RDF graph thinking. Um was that Oh yeah. >> Yeah. And then can you because that's something that's really of interest to me is the collaboration between you know the there's so many you've mentioned most of them already all the the subject matter experts but also like the I think the closest analogy to the to UTR's relationship is like the relationship between like a data engineer and a knowledge engineer. Um I don't know if that's exactly right because you're more of like a software engineer it sounds like. Um >> well my background is actually quantum mechanics. So I mean my back my background is actually in Yeah. Um, which is interesting because um, one of the things I've talked about is a lot of the math underpending, this actually is an interesting discussion I don't think I've talked about is one of the reasons why I I've from day one not had a lot of confidence in LLM and it goes back to something that I saw with quantum photonic systems 25 years ago and that we were coming up. We're trying to model this quantum photonic system and we were using what's called an overdetermined system. Are you familiar with overdetermined systems?

23:35 >> I'm not. No. >> Okay. Most people, it's one of those weird things, you know, I'm, you know, I'm a I have master's degrees and engineering and you would think like you'd think this would be something I'd run into, but it wasn't. You You always are taught like with, you know, linear algebra that, you know, if you have five unknowns, you need five equations. Well, they don't really ever go into, well, what happens if you have five unknowns and 50 equations? Your initial thought would be like, well, isn't that better?

24:01 Turns out it's way, way, way worse. And in our case, we had a situation sort of like this. We had an overdetermined system. And it's a basian approach, too. The way you solve this system is you kind of have to make an assumption. You know, what are these variables that we're trying to solve for? And the math was just not working. The system was just breaking. We could not understand it. We had meetings with so many mathematicians and they were getting into all these absurdly complex mathematical structures and saying well it's this kind of structure plus this kind of structure and you and we're just going what are we doing here? And none of it was working and then all of a sudden somebody said I think you have a regularization problem. We went what?

24:50 And it was this very niche category of mathematics that looked into kind of the randomness of these systems. And if you use these regularization techniques, you would find that uniformity in the system actually caused it to be unsolvable. So that the act of regulariz I don't even remember how they used to say it but that process of interjecting some noise some randomness into the system actually is what made it solvable.

25:22 So when I look at these LLMs I immediately my want my mind went well this is an overdetermined system. LLMs are overdetermined systems. you have the same data with millions upon millions of data points and I know how the math works out on this. The math doesn't work in that case and that's one of the things we've actually seen with you know people now re researching synthetic data. They talk about like synthetic data collapse in the LLMs and there and you because if you give this uniform data it just doesn't seem to behave well and you know again you know I had that experience with overdetermined systems for you know quantum systems and I I just kind of made that connection going all the way back and I saw you know the way the transformer architecture of LLM is designed this is just kind of inherent in them. So there are limitations to what this can do which is why you know I was convinced that those symbolic structures those grass structures would be necessary to kind of guide these kind of structures in a way that we can understand and give some control.

26:36 >> That's so interesting that you were when was this because that's >> all the way back in my academic life. We were doing research on photonic crystals which are >> No, I'm really curious. You have this you have this you're you're you're studying quantum theory back in college and then whether and then you this insights takes hold and then LLM's come along and you go like boy this is familiar. Was this the rise to that article you did about determinism and then and even the genesis of trust graph like the the um the seeing like the problem was immediate apparent to you and then was it was all the subsequent activity like not all of it but was it largely about sol addressing that the the issues you saw with LLMs >> to some degree sure. Um, I mean, again, I was very confident going all the way back to when we saw the emergence of LMS that that the deep learning approach on its own wasn't going to solve these problems because again, you know, I had dealt with determinism so much and how you have to rigorously prove how a system performs. And I know that with a probabilistic system, you can't do that.

27:45 And maybe that's also because of my background with quantum and understanding that we do live in this probabilistic universe. And um I used to I remember you know sitting in advanced electromagnetics you know classes where you're forier transforming Maxwell's equations going what are we actually even doing here and then you start to see like wait this kind of starts to look like the same stuff that we're talking about in a quantum like wait are we are we talking about the same stuff here and you start you start and then you just your mind starts racing and you go like stop I can't handle this anymore just focus no I don't want to go there And um but that experience I think did give me a different perspective because I've always I guess it is because of my background dealing with quantum theory and quantum systems. I've you know for 25 years I've just accepted we live in a probabilistic universe and that there is no certainty but also seeing pragmatically and practically when the stakes are high when people's lives are at stake when massive amounts of property value is at stake that that's just not a suitable answer. And you have to be able to prove with a high degree of certainty or you have to be able to demonstrate that the risk is not high or the risk is acceptable in a way that people are comfortable signing their name on a document that then attributes legal liability to them. So there's that understanding of the probabilistic nature of the universe that we live in, but then understanding the practical uh applications of How do we actually use these systems in an everyday way where somebody ultimately ends up being the person that's liable for the operation of this system? Which, oh, by the way, not to derail this conversation, there's actually a lawsuit right now between Amazon and Perplexity on this exact issue on who where the liability lies between an AI agent and who deploys it, which it it came down to uh perplexity having agents that would make recommendations on purchasing things in Amazon and Amazon sued them.

29:59 And this court case is going in all sorts of weird directions. It using the Computer Hacking Act as its basis for determining someone. It's it's a wild case that's going to set some really really major legal precedents for the the liability of AI agents. And it's it's not being talked about much at all. Matter of fact, my my friend Chip Block on LinkedIn is one of the only persons I see uh talking about it at all.

30:24 >> Yeah. know it's it seems like such an obvious thing like where where's the ch like the the the governance you like where's the accountability in these things but hey you just as you were just talking now I think I think this was the last line in your article on determinism you said trust is not about eliminating uncertainty it's about managing it well and is that sort of what we're talking about here the the the under developing these mechanisms for understanding the context in which you're operating you have a chance of managing this like um like do you think Perplexity andor Amazon will be knocking on your door you know for for help with their issues or >> probably not um uh I think well definitely not Amazon they have uh um yeah um they have their own way of doing things um but to answer your question yes um no absolutely up. I don't believe there's anything that's I don't believe there's anything that's identically 0% and I don't believe there's an identically anything that's 100%. Everything exists in between. And that's how you actually make these decisions when there are consequences. And I think a lot of a lot of people building an AI are doing it in a consequencefree way. Now, if you put together a demo or you're just playing around building something, you know, what are the consequences? You you might say, well, you might waste your own time, but perhaps they don't see that as a waste. They see that as a learning experience or perhaps they just see it as a form of entertainment to be I mean that there's a lot of developers that are absolutely doing that right now.

32:07 They're bored to tears in their day jobs and they're seeing this as an opportunity to actually get to build something that they haven't been able to do. So, there's a lot of that going on. But um and and that's another thing too is that most developers um a lot of time they're so disconnected from the outcomes from the software that it it is a little bit consequencefree in terms of like really severe consequences. there might be like some personal consequences to your own career trajectory, but you're so far disconnected from, you know, is this going to kill somebody? Um, that I think people have a really hard time even making that connection. Whereas, you know, I was running, you know, programs that were as a result of direct loss of life of some of our international uh partners, um, you know, back in my, you know, DoD days. Um, and you know that was a a a really sad story because here you have a total, you know, loss of life because of something really really really kind of trivial in a way and just a a trivial misunderstanding.

33:17 And if you try to distill it down to like one cause, you actually can't. It's a lot of like little things which you kind of cobble together leads to this scenario. And so it's a very unsatisfying exercise. And you know there's not one thing that you can point to that was the cause of this bad outcome. It was a lot of little things and a lot of you know thinking about it as a graph. You have lots of relationships and lots of data points that all have to come together in this very convoluted long traversal path to even make sense of what happened.

33:53 >> Yeah, this goes right back to where we start in the start of the autonomous vehicle stuff and the insane amount of information. But hey, I can't believe it, Daniel. We're coming up close to time. Um, but before we wrap up, is there any I love the We could literally go for hours on this, but is there anything last? Anything you want to revisit from the conversation or just make sure we share before we wrap up?

34:14 >> Well, I think that um I I guess I did want to talk about um I I know you don't like you don't want people promoting stuff, but you know, as I said, you know, Trust Graph is open source and it always will be that way. And you know I this is something we developed for people to be able to try to understand this themselves. It was practically designed for something that anybody can use in their own applications. Um you know we're big proponents that one of the reasons um people aren't quick to adopt graph technologies that's hard to work with. We tried to make it easy to work with and easy to manage. It's um a complete system that allows you to do these things and test these things and see for yourself. And um you know ontologies even themselves can be tricky to work with but you know we designed it so now that you can you have an ontology you load it in the system it can now structure data you know with that ontology you can explore that you can begin to understand it. So um it's out there for anybody to use. Uh we would love you know people's feedback on if there are use cases ontologies. Um you one of the things that we're really exploring are temporal relationships which is something we're be talking about a lot in the future and you know anybody that you know has ways that they'd like to manage this in the past.

35:31 Um you know we would love to talk about them. We'd love to understand their perspective. Um, and you know, we are even talking about, you know, I'm going to be doing this video probably recording either today or tomorrow, just going through basically Mark's view of how to structure RDF and why all that matters. And I think u um going to make Fred's I'm going to make Mark's cat Fred famous no matter not going to give up until it happens.

35:57 >> I love that. In the um in the conventional, you know, the ontology world I'm familiar with, there's like the the pizza ontology and things like that. I'm I'm I'm enjoying the cat um on boarding uh which worked for you and and I'm I'm hoping Mark has brings many more folks along with that. >> Yeah. Well, and I think it's I think it was eye opening for people to see because the three statements are, you know, Fred is a cat, Fred lives with hope, and Fred has four legs. And you can structure that just as three triples. You can. But then you walk through, well, why would you add these things like classes and types? And then why would you add all these other things? Because then when you need to add more information, you know, it's gonna have to fit into this way for, you know, disambiguation because um I I love the example because, you know, Fred, you know, most people would assume that's probably not a cat, but Fred is a cat.

36:50 Or most people might not realize that Hope is also a cat, but hope can be a person. Or it also it could be the concept hope that Fred is a hopeful cat. maybe, you know, he thinks he's about to get fed and that's why he's full of hope. And these things semantically matter. And these are also things that LLM don't do very well with. But the thing that LLM struggle the most with, and this would like this is going to be a teaser for a lot of other stuff coming in the years to come. Looms do not understand time at all.

37:22 >> Okay. Well, cliffhanger. I love this. So, we'll have to have you back on in a year or two. But, um, hey, one very last thing, Daniel, if folks want to connect, uh, what's the best place to connect, uh, with you online? >> Well, I'm pretty easy to find on Well, I shouldn't say that because I have a common name, but, uh, LinkedIn, uh, Twitter, trustgraph.ai has all of our information. Um, trustgraph on LinkedIn, Twitter, we're in all those places. Or YouTube, uh, our YouTube channel, trustgraph.ai, at trustgraphai. That's where we release um a context graph and action video, my video on what is a context graph, which is kind of some of the things I said here. And also um this this basically kind of explainer on on RDF uh coming very very soon. Hopefully I'll have it up in just a couple days.

38:09 >> Great. Well, yeah, sure that because it'll be a little bit before this episode goes up. So anyhow, I'll get all those as many of those links as possible in the show notes as well. Um well, thanks so much, Daniel. I really enjoyed this conversation. >> All right, thank you for having me. >> For show notes and to sign up for our newsletter, visit our website at knowledgergraphinsights.com and please rate and review us on your favorite podcast platform. Thank you for listening.

Summary

Daniel Davis, co-founder of Trustgraph, discusses his development of an open-source context graph platform aimed at enhancing AI systems by grounding them in logic and knowledge. He emphasizes the importance of context graphs in managing complex decision-making processes, particularly in safety-critical domains like autonomous vehicles and aerospace.

- Context graphs are optimized graph structures designed for AI usage, integrating knowledge, decision traces, and feedback layers.
- Davis's background in aerospace and quantum mechanics informs his approach to managing uncertainty and determinism in AI systems.
- The rise of context graphs has been accelerated by the limitations of LLMs (large language models) in handling complex, specialized knowledge.
- Trustgraph aims to simplify the use of graph technologies, making them accessible for various applications.
- Davis highlights the need for a single graph system to manage relationships and data effectively, contrasting it with the trend of using multiple small graphs.
- The podcast discusses the legal implications of AI decision-making and the importance of accountability in AI systems.
- Davis plans to release educational content on RDF (Resource Description Framework) and context graphs to further demystify these technologies.
© transcribe · For agents Built with care and craft by Gokul Rajaram