Build neuro symbolic AI agents with OriginTrail Decentralized Knowledge Graph - Branimir Rakic

Our next speaker is Brana Rakich from origin trail. So origin is uh decentralized knowledge graph and what happens if you toss in uh decentralized AI on top of that. So brown up.

Yeah,

thank you. Thanks. Thanks for the intro. Hey, early applause. Maybe it's too early.

Don't don't give me such an early applause. uh maybe it's not uh worthwhile. But anyways, thank you for sticking around. I know it's 4 p.m.

so probably everybody's a bit tired. Um I'll try to make it interesting and I'll also try to um give enough CTAs so you can somehow pitch into this whole idea of origin trail. What I will be speaking about is slightly different from the previous topic. Uh but it's quite related. Uh we we won't be talking about the usual digital assets we are used to but rather what we call knowledge assets.

um and why they are important for AI. So we'll also be touching something called neuros symbolic AI of course agents as well. So a little bit of everything in this title. I know it might be a little bit hard to understand. So by I hope but that by the end of this presentation it will be much more understandable.

I'm one of the founders and CTO behind Origin Trail. Maybe I'm a bit overweight if you can hear that sound. So I'll just move to the other part of the stage which might be a little bit more solid. Um anyways um I'll be telling you all about what this neurosyolic decentralized AI is about and how how you can launch an agent using something called the decentralized knowledge graph. By a show of hands, has anybody ever heard or even used the decentralized knowledge graph here?

Couple of hands. Nice. Which means most of you haven't. So then I get to explain it fresh, which is great. All right.

So um I'll start with some some problems uh just so we set the stage so we understand what we're trying to solve for and um obviously as AI is becoming kind of the dominant technology or we like to say AI is the new UI technically all of these issues are very important and we need to somehow tackle them. So very quickly because I know you probably know about pretty much all of them. Hallucination number one everybody ran into it. It's not just an issue it's actually a feature of neural AI. So the current AI that is dominant Chad GPT like actually let's do a quick exercise.

Um when I point a finger at you I would like you to say a word that fits into the previous sentence that the previous person said. So let's start with you. Start with a word. Basically, what we've seen is an LM in action. Everybody took the previous sequence of words and guessed the next word that makes sense.

So, this sentence kind of makes sense, right? So, you literally did what you literally hallucinated something that nobody had in mind just like 3 seconds ago. Um, that's a problem, but it's also a feature. So, you know, depending on how you use it. Um, thanks for participating by the way.

Um, privacy obviously huge issue. You don't want to feed everything to JGPT. Um, data ownership and intellectual property another one. I'm sure I'm speaking to the choir. We're at a crypto conference.

So, I think ownership is part of our ethos here and there's no need to go deeper. Centralization of course as well. So obvious clear issue uh do we want to have open AAI or any other of the big big companies in being in control of all of the info and the the UI and the knowledge we are consuming essentially and driving our systems with that comes with bias naturally and even one more thing which is called model collapse which is very interesting. Turns out the more AI generated data there is in the world the more other AI trains on that and because that is hallucinated sometimes junk it actually makes these AIs collapse or go worse. So there those are just some of the problems or pathologies I would like to call them and what I'll be speaking about today is some ideas and like directions that we took and and tackling some of these problems obviously in an open source decentralized way.

So, um, I'll try to skip through like another 10 slides. Hopefully, I have enough time for it. And, um, hopefully that sparks, uh, a little bit of a discussion, and I'm happy to chat more about it, uh, outside as well if, uh, we have a booth, by the way, just outside of the the the hall. Um, some people even go as far as to say that LLMs are brilliantly stupid. This is the the one of the the leading critics of AI um of current AI, Gary Marcus, who's an AI uh scientist and who actually had a bunch of AI products before.

And actually there there is sort of a a feeling of slowing down of that that big LLMs are no longer as progressing as much as they were. So that it's not an exponential thing anymore, but rather it's kind of flattening out. uh which makes sense because they're running out of data, but also we're kind of, you know, getting to the point of limits of what this approach of guessing the next word really well can do. So, there's a new thing that that that is in town. Obviously, there's agents, but there's something called neurosymbolic agents, and I'll explain that in a bit.

Um agent market goes burr. This is just kind of a very quick illustration where you can see what what the agent um market is projected to be. So, a lot of growth. Everybody's betting on them. You probably even start getting sick from the word of agents.

Everybody talks about agents. Um, but they're here to stay. There will be plenty of agents all over. We'll be interacting with them and they will be interacting with each other. So, it's an exciting opportunity.

Uh, but the question here is what kind of agents do we want? How can we build some really really good agents um that ideally don't have those problems that we've seen? Uh, and there's a hint from history. history uh kind of looks a little bit like this. This is a a picture from a paper by Google DeepMind called Welcome to the era of experience.

And what they believe is that we are exiting the era of human data where we would feed all this data that we produced somehow over the years like you know stuff it into that LLM and then you know sort of push something out. um and rather go into era of experience where agents go and experience the world and therefore generate data like they make a mistake um you know they learn from it they store it in their memory they some next thing they try you know succeeds okay they have that experience as well and the assumption here is that we're going to go again ballistic once we approach the world of agents actually learning from their experience and storing it in memory actually there's some hint of that that bump before that you see Alph Go and Alpha Zero. That's um actually what happened before LLM broke out. So neural AI took all over to all our attention previous back in the days. Maybe some of you remember there was like the first AI that beat I think GIC as part of the chess master happened like 20 years ago or something like that.

That AI was actually symbolic AI and what is symbolic AI is just kind of computes what are the options. Chess turns out to be a simple game uh that it you know has a finite number of options. So the that AI I believe I believe it was IBM's AI was basically computing all of the options but like it wasn't hallucinating. It was just like creating genuine plausible paths in this game and whenever you were playing playing against it you were just filtering out the remaining paths for it and of course it will win. Um, but turns out that if you take the game called Go, it's actually such a weird game that it has so many possible states that there's more states than the atoms in the universe, something like that.

Um, so it's not possible to premputee all the states because then you don't have an infinite hard drive, right? So they had to do something more clever. They had to do some predictions. So it's kind of a mix and it's an early neurosy symbolic AI. It didn't just compute all the states but it did compute the states close to where you were and then it would predict just like you guys predicted the word what is sort of the most common possible next step.

So it's a combination of this neural guessing hallucination and actually computing something which is plausible something that you know we can see a few steps ahead and the assumption is if we take that approach and we apply it to for example physics or math this alpha proof that you see with the red that's actually recently won um I think third place the international physics um uh uh uh Olympics so it's still not be better than the best human but it's proving math. So it's not hallucinating. It's taking some system and it understands the system and it actually goes and and make proves a theorem which is a huge thing. That's something LLMs cannot really do. They just guess something.

So this this is the thesis here is when we have agents that can do both symbolic learning, you know, doing something with actual logic and with this creative guessing, we're going to have all the pieces we need for some really awesome AI. So the only thing that remains is we make it decentralized and trusted, right? So I'll talk about how we do that today. Um there's three main components of agents. Number one is systems.

They either interact with other agents or with other systems. So that's their domain of agency. Then there there's the AI models. They do rag. They do chain of thought.

All kinds of things. They're the u the sort of natural language interface. But what I like to focus most today is memory. Memory is the the kind of the key point. That's where the agents record these experiences.

and basically all the outcomes, you know, mistakes or not mistakes. So when they need next time to make some decision like in chess or somewhere, they can look up that memory. They don't need to guess if they already saw that situation. And um so it turns out that if we have some really really powerful memory, we can make some really powerful neuro symbolic agents. Uh but if we feed that that memory with junk, it's another garbage in garbage out system.

So it won't be reliable. So we we want a way to tame that junk. So the kind of agents we're looking for are the ones that reasonable that have some reliable reasoning. They don't hallucinate. They have information providence.

So that means we know if some junk is coming here or not. We want to know who's who basically fed it some information. Where did it learn something? Was it the agent itself or was it some random guy from the internet claiming the earth is flat? Um we also want fair value exchange and incentives.

Obviously that's why we're here. uh these agents naturally will be connected. We live in a connected world. So there will be multiple data sources. We want to have a respect for data ownership.

So we want both us as humans but also as agents agents to have their own data that they own and ultimately privacy preserving. I think this also goes without saying saying uh for this crowd there's no need to explain that. So ultimately the vision is of something called trusted collective neurosyic where we combine this creativity and these facts. So the facts basically having some deterministic system that can do logical reasoning that can do rule-based inference and that has trusted execution. Well, I'm pretty sure when I say trusted execution, you guys already kind of infer that I'm going in the direction of blockchain obviously um and not just blockchain but uh also these these elements above actually fit to something called knowledge graphs.

So this domain on this this left side of the brain the rational one is actually in in the the vision that we present what we're building as the decentralized knowledge graph on the right side we have creativity so the basically the LMS they are great for some things statistical predictions natural language things like that questionable execution but you're not necessarily going to give it all the power you want you want to pair it up with some with with a with a left brain right so left brain and right brain when they work together you get something very powerful Um, what are graphs? If you haven't had a chance to touch them, this is a screenshot of one of them from the origin trail decentralized knowledge graph. And basically what they present is a semantically rich data representation, something that AI loves. It basically you give it very very concrete data, something like a typed object in u in like let's say an object-oriented programming paradigm. Um, and it's able to actually do logical reasoning.

It's able to do inference even without like an LM. So because if there's a rule encoded let's say a rule says um DNA is only comprised on AT and GC molecules right and they only combine like that so then you can use that rule to basically you know understand that something is wrong for example that that somebody connected something that shouldn't be um and other things it has flexible data model so that means you have things not strings so these things can be connected uh as in this subject predicate object pattern and it's a very mature technology. That's one of the key things. It's uh it's been developed under the opices of W3C. So there's plenty of standards used by Google, Amazon, Netflix, Uber, NASA, you name it, everybody.

Uh and they're using them as AI for a very long time. So way be way before CH GPT and um they've basically perfected a way how to create value from data uh but in silos. So what we tried to do with the decentralized knowledge graph is bring that in an open source decentralized way to to the world and that anybody can actually take that power and do it for themselves a little bit just more about knowledge graphs. So think of it as two entities. You connect them with a relation and you can with this you can model the entire world.

You can model a relationship between people between things between agents whatever you want. For example you can have Toy Story who has a director which is his name is John I believe I don't see it from here and then that could be another thing and then you have some literal for example and then you can make as many connections as you want. you have to change your data model because uh you're building an app and you figured out you need a new feature or your agent learns about a new concept just adds a new connection description of that thing and it can go wild. So essentially we get the decentralized knowledge graph which is this middle layer of the memory and it's a decentralized network of nodes that share public or private knowledge assets. We have the blockchains that are underpinning this knowledge which provide data provenence, identity, ownership.

Actually, no data in this paradigm sits on the blockchain. We don't want to bloat the chain. The chain doesn't have the query engine that we need. It has some query engine, but it's not as powerful as as as knowledge graphs require. And then on top of that, you can build agents or any type of AI with protocols like MCP and so forth that leverage the two layers below.

So that each of these pieces of information has structure and it has an own known publisher. What does a publisher do? A publisher creates knowledge assets. Each record is a knowledge asset or a memory for your agent. You can consider it as a knowledge asset.

What is a knowledge asset? Well, it's an NFT of knowledge. Technically, it contains an NFT, cryptographic proofs of the state of that knowledge of that knowledge, pardon me, has some graph data and some vector data. You can combine it. You can use one or the other.

But ultimately, the point here would be that you have an ability to create a record. The record has an attached NFT to it and then whoever has the owner, whoever is the owner of the NFT gets to manage that asset. So, I can update it kind of like a git branch. I can like do more commits, but only if I'm allowed. And then this piece of knowledge fits with other knowledge in the graph.

So it's consider it as one entity and I can connect it with other entities. This can be knowledge about any object in the world or any NFT or any anything um or any agent or it can be a single memory and I think I have a slide that shows that memory later. Um I know I'm close to the end of time so I'll try to speed it up a bit. Ultimately what happens is you have this network of nodes which again the decentralized knowledge graph is not a blockchain. So there's two groups of nodes the co core nodes and edge nodes.

Uh edge nodes designed to run on edge devices phones laptops and so forth. So you can keep your data private on your device. So you don't have to publish it to the decentralized knowledge graph and therefore make it not private. But if you do want to publish something that you want to be public, you will publish it and that will be hosted by these core nodes. And these core nodes are basically running with high up time.

They're incentivized to do so. They're incentivized by those fees that you spend to publish uh using the track token, the native token of the decentralized knowledge graph. Um so technically what you can do then is you can run a um a decentralized agent on your phone that has certain knowledge on its own and it can also interact with knowledge of others and therefore they can generate memories together through interacting with the world sorting some of them to remain private some of them to remain public and those memories um as I spoke previously in the in the workshop um are actually something that they can cross share in the same graph. So for example, we had a situation where one agent on Twitter um we told it like my favorite food is pizza and then you went to know go you can go and ask another agent hey I told this other agent my favorite food is pizza like can you my what is my favorite food can you go find it and actually it did it went and and it said yeah it's pizza so technically what we're creating is a collective decentralized trusted memory where each agent has its own records uh and is able to decide how to share them with others while these records Records are used for rag as it says here. So decentralized rag.

So every time you read a record or you every time an agent makes a decision based on some memory, it can tell you exactly which memory it used, who was the publisher of that memory at what time and what it contains. So it's not like hallucination, but it's actually a piece of knowledge that was packaged in in in such a way that it can be used uh and reused. Um and the network has been growing quite quite steadily. This is just a screenshot from I think this morning of the the staking UI. So this is the number of knowledge assets on the right the the steep hockey and the little less steep hockey is the total fees in track tokens.

So um it's it's it's going quite rapidly. So we're reaching close to 1 billion knowledge assets. By the way this is used by plenty of um enterprises as well. So origin trail actually started 2013 2013 without any blockchain or whatsoever and we started building solutions with knowledge graphs and very quickly realized that if we want to build something which tracks bigger bigger systems bigger supply chains which is where we started um it couldn't do so without decentralization. Everybody wants to keep their own data.

Nobody wants to you know put data in one place and you know we call that one place whatever. So actually we we married knowledge graphs with blockchain ever since the first idea was born in 2016 2018 we launched the version one the mainet of origin trail so it's been around for quite some time launched on Ethereum back then uh expanded to polygon nosis base polar dot meanwhile um and from 2022 we've actually started doing these integrations with bunch of uh enterprises including Swiss railways uh major US retailers like Walmart, Walmart, Target, Home Depot, um a bunch of other companies. And just recently, we launched another venture called Humanitech together with uh with some of the biggest u names in the adult industry. Um you can see a video about that outside. Um anyways, we cannot build this whole thing alone.

It's quite a big thing. It it's very ambitious. So there's a cure code which is a call call for papers. If you guys are building or thinking or researching in this any of these areas, decentralized AI, knowledge graphs, neurosymbolic AI, we'd love to to chat and we'd love to see um what kind of research you can contribute to the ecosystem. There's also um uh some some pool of rewards for the best papers.

So, if you scan this uh it will tell you a bit more about it. It's not open yet, so you're kind of very early. uh you can consider this a bit of alpha and um and yeah this this is this is a joint thing this is an ecosystem thing so it's it's definitely something that we'd love to to see um more folks join and and join this this vision of collective neuros symbolic AI there's one more thing you can do um at this uh conference we're also participating in a hackathon so you can build your first neurosymbolic agent with memory uh stored on the origin decentralized knowledge graph so if you scan this This is your classic hackathon page. Um, and you could build something like this. For example, this is a screenshot of the edge node of origin trail which showcases uh one small um uh set of knowledge assets connected and including its owners and publishers and so forth.

So this is actually the equivalent of a blockchain explorer in the origin real world. You explore graphs. Um, so yeah, you can you can take out agents, edge nodes, and even just the classic SDKs and build something exciting. As long as it has to do something with neuros symbolic AI, we're happy to to see your submissions. And if you're building, we're just outside.

So, uh, do come to our booth. We're happy to help you out. And with that, um, I'm just going to do one one more CTA. If you want to join the community or submit your paper or build agents, this is the best place to start. This is the official docs and um yeah, happy to to have had a chance to present this idea in front of you folks.

I'm hope I hope it was interesting and looking forward to seeing you at the rest of the conference. Thank you. I'm not sure if we have any time for questions or if we even have questions. We do, right? All right.

I'm going to try to not hug the microphone. I have a lot of questions. So, but so one that strike me is uh so these graphs are basically structured data that is linked together uh with edges and uh are all the graphs uh are are there like separate graphs or or is there like one big graph that where everything is actually connected?

That's a great question. So you can launch your own we call it paranet paranet graph. So basically consider it as a boundary. Um and then you decide what you what kind of boundary should it be. So for example we can create 10 paranets today that are all public.

Let's say one is about sports, the other one is about cars whatever and we can make them public so anybody can publish knowledge into them which means also somebody can like publish junk inside which also means that you can make it permissioned. So you can say okay only these people who let's say have this NFT or who have this or that thing they would be able to participate and add knowledge to this paret. There's also a way to make private parinets so among multiple nodes. So for example let's say you and me we want to create two agents that work together and we can have our own paret that only our nodes get to share and synchronize these knowledge assets in between each other. um yet still the chain is used so all of the the fingerprints and the NFTs they stay there so we we can maintain this trust uh between the two of us the two of our engines so there's a lot of flexibility this feature is called parets I didn't get to speak about it much but happy to show show more if you want

was there someone else or can I ask another follow-up question

I'm good if you guys are good

uh just a quick one is um so essentially there's a this is a peer-to-peer network

yes

where anyone can join and be uh either a core node and with bigger hardware or edge node by just like synchronizing some part of the state and serving it to other peers in the network.

Yes, very good point. So if you anybody can run a node and to run a core node you basically need some solid server but it's not like a you know a bitcoin mining rig. It doesn't like eat up all the because there's no mining involved. uh but it does require uptime and it does require a certain amount of stake as with many of the other protocols. So stake tokens on on the node and basically from that moment on your node is part of this core network which means it shares the fees with other nodes uh that are accumulated from publishers.

By the way, there's no inflation. The token has been preminted 2018. There's like everything is circulating. There's no, you know, incentives of that sort like you know where does the yield come from? it comes from publishing from usage from that spike you've seen.

Uh and and actually that then these nodes compete on uptime. So the better the uptime, the more rewards you'll get, the the more stake you track, the more rewards. The lower your fees because that helps adoption, the better the rewards. So and finally, the more knowledge you publish, the better the rewards. So there's a combination of those four factors which are all positive for the ecosystem that evolved over the years that would help your node.

So you can for example you don't even need to create your knowledge. You can just open up your node as an RPC to the rest of the world. So it's not like in Fura but it's like your node and you take let's say 10% of the world publishes through you. You get a bunch of benefits from that because you helped uh the ecosystem grow. Um edge nodes they don't get any rewards because they're not hosting the the entire network but they can plug in they can host your private data.

the the idea is that they're a bit more like your app on the phone, while the core node is more like a server, but it's all open source and and pretty much anybody can can participate. So there's no there's no central government that would say whatever. Um and you can actually integrate different chains. So there's there's um currently like I said three chains supported but uh we are looking to support many more chains so that you can create these knowledge assets on like for example Sana or something like that. No not Solana.

Not Solana. I'm just kidding. Um any chain we're really not u we're not religious in that sense. We think that if people want to build with this and it brings value to them whatever chain they want that that it's fine with us. Um thanks for the question.

There's one more over there. Yeah. Hello, my name is Valentina. Thanks for amazing presentation. Uh probably my question related to the previous questions.

Um as I understood we have this knowledge graph containing like proved facts and then we have uh agent user communication history. Um is there any let's say protocol for syncing this data and maybe like factchecking of the information that user inputs and because probably we want to use this information provided by user just like somehow automatically or semi-automatically extend our knowledge graph.

That's a very good point. I spoke about it briefly at the workshop like two hours ago. Uh so I'll try to summarize two points. One is there is no uh embedded algorithm that is going to verify if something is true or not because something like that doesn't exist in the most general sense and I'm not sure if it will ever exist. Um so the protocol is very neutral.

It lets anybody publish anything even garbage. So like I could come and feed your agent garbage. It's up to your agent to then decide like what what would I do with this you know and the best way for it to decide is to see how it fits it existing world model which the world model is ideally a graph in this sense but also can be um some form of system that you built on top that let's say either filters by identity or filters by some form of reputation um and that is something that needs to be built on top so it's not part of the the dick per se but it's something that we for example people have experimented a lot with um in DI for example there's quite a lot of interest about origin trail in in DI space and um there's a project that has been basically creating knowledge graphs from scientific papers so what they did is they would um let everybody publish PDFs literally convert PDFs to knowledge graphs but then they would have a system that would go on their own check like is this a real paper where is it published you know and it would go up a couple places on the web and basically publish its own knowledge assets saying hey I I've checked it you know and so now the question is you know is this good enough for your case for them it was you know so it's um but yeah it's ultimately you need to build let's say your let's say truth algorithm for your domain of knowledge uh and it can be done it's um it's very interesting topic to to look into though like reputation scoring yes yes when you think about it social media does that a lot you know but that it's questionable what reput reputation is like why does Elon Musk keep popping up in my Twitter feed? I don't even follow him, you know. So, he has a huge score and essentially they have a graph under the hood.

They it's a social graph and they have points to each one of those those uh accounts. Think about you building your type of algorithm that is under your control and you know then you get to alter it over the way and then you assign reputation to a source therefore trust the source the whole uh the the tagline of origin trail. Uh but yeah, happy to chat more about it if you if you want outside. Uh I see there was one more right.

Hello, my name is Dennis. Uh could you briefly run us over the tokconomics?

Sorry, what?

Tokconomics.

The tokconomics. Ah okay. Yeah. So um again whenever somebody adds knowledge to the graph they would pay a certain amount of fees in the track token. track is this limited supply 500 million token that was preminted 2018 um actually launched with an ICO and this was um there was all the tokens that were will exist the assumption is the more people use the better thing we build the more people will use it the more tokens will get locked the less tokens in supply then you can kind of think of what comes next uh but that's kind of the key assumption from the beginning and then um what does the token then do so when I publish some knowledge to DKG the the tokens get locked into contracts and then these core nodes they compete to provide the service for that the service is not just storing it's also serving so querying um and essentially uh you can as a as a node as a node runner get pieces of of that big reward pool that gets accumulated over what we call epochs which is one month is around one epoch um and compete for those rewards.

How do you compete? Four different things like I mentioned uptime. If your node has higher up time, it has basically it's going to be proving that over the course of the period through something we call a random sampling proof system, which we're just releasing right now actually. Um, and it's it's proving that over a period of time. So the the more proofs it submits to the chain, the the higher the uptime.

The proofs are actually Merkel proofs of the graph. So they prove that they actually host the knowledge, the right knowledge. Um, and then on top of that, three factors. So the amount of stake you attracted to your node. delegators would delegate stake because they see your node is good.

Um the amount of publishing that your node is responsible for like I mentioned previously. So somebody can like I can run a node and open up my node as an RPC into the network so that somebody can publish through my node. That's a positive behavior for the network and therefore it's incentivized. And then finally the let's say ask price the vote of the node towards the network. So if my node is asking for less, if it's less greedy, let's say, that's positive for the network as well.

So that means over time the fees go down and uh the adoption is more um uh lubricated if you will. So those are the four the four elements of tokconomics. There's a whole RFC on that. We have a very uh comprehensive set of RF RFC's that are always vetted with the community. So if you guys are interested in learning more details, I can share that as well.

But I I hear the on the side. So, I think it's my time to hop off this stage and um thank you again. Do I leave this