Solving AI Alignment by Modeling Beliefs in Code

But how did we assume this problem is going to develop? The fact that we focused on designing data models before we even wrote a line of code was really helpful because we could always revisit our assumptions.

I have the desire to be able to quantify a philosophy alignment score of my followers. Like, I'd like to know who are my best followers. Not just, like, how much you've engaged with my app, but how much do you actually believe in the philosophy?

How do I model my belief as a tree? Right. So, for example, sometimes, sometimes we have an initial belief that is then supported by multiple other beliefs. We want to plant it, and we want that belief, that new belief, to grow, and we wanted to kick out older beliefs in the past.

What are our thoughts around quantifying subjectivity in beliefs?

Yeah, sure. Why don't we jump over to.

Welcome back to the ABCs for building the Future podcast, where we bring on people working on big things with big stories to tell. We bring on entrepreneurs, founders, leaders, researchers, scientists, and builders of the future. And in today's podcast, this is the Epistemic Me team. We're continuing our saga of Building in the Open. And today's episode, we're going to be talking through the object model and the data model in our architecture and some of the design choices behind our technical implementation. So if you're a developer that wants to leverage a personalization layer in the form of our SDK, which makes it as easy as possible, that's our goal, to quantify subjectivity, model belief systems of people and your users, to be able to personalize value to them, whether it's recommendation or whatever it might be. So to start with, we're going to go over some project updates and just do a little roundtable. We've had some exciting developments recently, and then we'll go into the problem framing and get immediately into the design and the architecture so you could best understand it. So first thing, we'll just do a quick little roundtable in. If you had to pick one thing that you're really excited about in terms of progress we've made recently, we'll do a roundtable as a team. Jonathan, you want to go first, then Dean, and then I'll go. And I think this will set the stage of just like. We're pretty stoked.

Yeah, sure. I think just seeing the SDK kind of come to life and the use case of this chat survey app that mirrors your beliefs is just kind of like a really cool. I mean, once I started building it, it's kind of like I Could see it everywhere. You'd see how I wanted to use it in a lot of different circumstances. And we recently kind of like refactored our data models to create this epistemic context, which is a really flexible way to easily start building a chat app like this and then layer on your own epistemology on top of it, which is just really powerful thing. Hopefully we'll be able to explain how all that works today.

Cool. Awesome. So some refactoring and more progress made on our golden reference application which we're calling the Belief Mirror Chat reference app or some combination of those words. Dean, you want to go next? What progress are you excited about?

Yeah, we've been making pretty good progress on building models of. Haven't quite come up with the name here. I'm just going to call it Models of Influencers. So we've been kind of working slightly with the Don't Die team to see if we could ingest some of Brian Johnson's kind of material about his Don't Die philosophy and see if we can like structure it into a coherent version of that philosophy that you can actually interact with and then we can actually go into, embed into these applications like the application you just mentioned and have users actually interact or even update their beliefs in the context of a coach like Brian Johnson or really anyone. I think we want to design it in a way that it's really scalable, such that we could take anyone's content and then in a matter of minutes like bootstrap a whole model of their. Of them.

Yeah. And this is such a powerful use case and unlocks so much. So I'll just frame it in a user story. So this is as a influencer or coach or somebody that has a following, a community that I want to be able to scale my interactions with them then. But I also really worry about personalization. So an AI version of me, is it really going to be understanding my philosophies, who I am, why I am the way I am, why I think, the things I do? So that's exactly what we're building is a way to bootstrap that with a high degree of personalization so that you could scale your message and your philosophies and your interactions with your audience. And on the flip side, the audience in this example, the user. I as a user, let's use the Brian Johnson use case. I as a user, go to Brian Johnson as my AI health coach and it's embedded with all of his philosophies around being scientific and data driven. Around your health. And I talk to this coach and then underlying that epistemic me updates my belief systems and that interaction becomes more personalized over time as we learn more about your beliefs. And therefore our recommendations from the Brian model of self that we've created to you will be a higher efficacy and efficiency over time. So that's like the really exciting thing where, you know, some people might think that could. In your head, you might think that might take weeks or months to bootstrap that. And what we've done it is made it so it's like hours, days, like on a different timescale to get that up and running. So just providing some business value context on those progress updates. I'm really excited too, to answer the question myself. I'm really excited actually at how and I was messaging Jonathan about this yesterday. I have such an optimizing mind. I love optimizing shit. So streamlining our processes for like content creation, the content engine, how we get the podcast created, the thumbnails, like just all of that stuff that happens pre production production, post production, and then creating a content flywheel that has been going really well to broaden our message and hopefully grow our community followings. We had, in our first podcast launch, we had a YouTube fan, I guess it wasn't originally a fan, kind of a hater at first, put something in the YouTube comments saying like, oh my gosh, this is so, you know, egotistical. So I responded back with a lot of kindness and then they actually responded back and referenced a lot of work that we talk about pretty regularly amongst us, like Michael Levin's work, predictive processing, Fristen's work, and I thought that was so cool. So I, me personally, I'm excited by how quickly we are streamlining our content creation process to broaden our awareness and our funnel of potential contributors and building a community of people that are concerned with building the future in a good way. So that's what I'm excited about. Okay, so the next big feature or milestone we're working on, I'll just frame this really quickly and then we'll get right into the object model because we got a lot to go through and that's a big double, triple, quadruple click. So next big feature of milestone we're working on is creating a model of self of Ryan Johnson, taking in all the content that he's created over time, his surface area and body of work, the books he he's written, his don't Die philosophy, the blueprint protocol, all of these things, and trying to create a model self a Brian. And as Dean said, strap that with our Belief Mirror chat reference application to bootstrap a quick AI health coach that is personalized to Brian Johnson. And that will prove out whether or not we could do this very easily with other people with a community where they have similar problems of wanting to scale influence and philosophies and interactions with their communities. So that's what's really big in front of us. And by the time this podcast is out, we'll have already been live with the GitHub repo open sourcing, which comes out on the 13th of January 2025. So at time of recording this is the 12th, so that'll be tomorrow. We got a lot of work to do. We're all going to meet together at my home and actually just go through a big working session to make sure everything is good for long. And after that we'll be open for business, so to speak. And any contributors that are interested in contributing, you know, human level expertise, like if you're a researcher within like behavioral science, neuroscience, metaphysics, those fields or philosophy, and you want to be part of an advisory council that helps build the future with us with this open source project or if you're on the technical contributor side, as we've mentioned many times in previous podcasts, please reach out. Look at the GitHub, look at our website EpistemicMe got EpistemicMe AI. You'll find a way to join us. So framing the problem now that we've gotten some project updates out of the way, Framing the problem of epistemic me and what we're actually solving, human understanding is really fragmented and unstructured. And you could see this even when you go out in conversations daily, like it's a lot of just unstructured conversations you might have that makes it very difficult to model beliefs, perceptions, behaviors accurately. And we believe without a precise framework to map and quantify that and analyze human belief systems, we really lack the tools to enable us to optimize personal growth or enhance decision making or really optimize for personalization as a problem. And so the fundamental question to me is always, you know, if we don't fully understand ourselves, how can AI understand us? And Epistemic me is really set up to solve that. Well, so to frame it a little bit more from a design standpoint, Jonathan and Dean are joining me, my technical co founders. With that kind of groundwork I laid, how did you two approach designing a system that bridges those abstract kind of philosophical concepts like epistemology Gaining knowledge. What is knowledge? What is truth? With actual concrete technical implementations in the SDK. I'll leave it to one of you two to get us started on that one.

Yeah, yeah. I mean, yeah, well, I was going to say, I think the first thing was just kind of deciding how are we going to model the self. And it might sound simple. We created something called the self model. And so when you start out, when you start building epistemic me, you build a developer model with an API key and then you can create models of users in a familiar way that you could probably use the user ID that's already in your database and hook right in. And then from there you can create a model of self. We call it a self model, which is like a container for the particular philosophy or the epistemology that we're going to model for that user. I think that's was like kind of the starting point for how we, you know, started converting the philosophy, at least from a top down point of view, what it looks like.

Cool.

Yeah, I, I, I, I would echo that. I think like another thing that we kind of focused on more from like even like a philosophical perspective is just like defining our terms like what are our nouns? Right? Like, like as Jonathan mentioned, like it sounds super simple, but just agreeing on like five or six nouns that really matter to the system and like really trying to get clear on what they mean is actually really important in these kind of problems for kind of setting up a framework for thinking. And I think as we've gone about changing this over time, we've realized we didn't get it right the first time. But the fact that we focused on designing data models before we even wrote a line of code was really helpful because we could always go back and revisit our assumptions and think about, okay, how did we assume this problem is going to develop? And then as, as we start to introduce new concepts, we go back and we redefine our terms. So I think this is, this way of working is really something that defines epistemic me in my opinion. And I think if you are a contributor coming into the system for the first time, you could really grok the system by coming into our proto subdirectory or subrevo and just checking out our data models and seeing if they make sense to you and if they don't make sense to you, leaving us a comment and basically saying, hey, I don't understand this. So these ideally should be self documenting. As Jonathan mentioned, you should start off with the self model and kind of Click into the belief system and the dialectic. That's really the. I think it's understanding the terms, which is really, like, what's important here.

Great. So another thing might be, like, I imagine audience might ask, like, what are our thoughts around quantifying subjectivity in beliefs? Could one of you touch on that just at a really high level? And then, you know, as we dive in, I think things will connect more.

Yeah, sure. Why don't we jump over to the belief system? I think actually that might bleed really well into how we think about predictive processing in the context of an epistemology over belief. And Jonathan is actually one of. I think could speak really well to our predictive process implementation. Maybe. Jonathan, you want to jump into there?

Sure, yeah. I think. Well, just philosophically, if. If we're trying to understand what's the belief system of, what's the subjective belief system itself? I think you have these kind of, like, metaphysical questions that hit you right away, and that is, what is the world and what is self? And so there's multiple different ways to model the world and self. I think the way we probably typically think of how science works and, like, is that there is a single world out there, and that as humans, we are kind of continually discovering things and using the scientific method to understand sort of like causal events and reproducible patterns in our environment. And that is one way of looking at metaphysics. You could call that a materialistic point of view. There's different sort of labels for it that are sort of, like, embedded into a lot of our philosophy of science. But some of the other approaches that I see used more in the humanities in an academic sense, but also are coming out of neuroscience. And areas like psychology tend to create a world that is specific onto a individual self model. That means that the self model has its own belief system and has its own world model, and that it's trying to use that belief system to sort of like, generate or create or understand that world as it's sort of like, you know, moving through time and consciousness in order to deal with subjectivity. I think these models on the other side of the spectrum, where we consider the world to be distinct onto the self as a starting point, are really critical to characterizing subjectivity. And those are the types of models that we want to use where subjectivity is really important. So, yeah, there's a. There's a lot there, and my mind is racing a couple different directions, but I think I'll just kind of leave it there for now and we can dive in Deeper, I think I'll also.

Add to that, I think like, I think one way to think about subjectivity and kind of like a philosophical framing that we found really useful is the difference between ontology, which is facts, and epistemology, which is like, how do you know facts to be true? So we really think this concept of the belief is really important. And what the belief is is this ability to represent a fact that may or may not be true. It's a probable fact. Right. And the idea is as you state a belief, we all have beliefs that you know are oftentimes not true. I think all of us know this in our day to day lives. So we want a way to represent something, something that someone believes and then also be able to layer on context that indicates how true the user actually thinks this is or how true the system believes that it is based off the evidence that we have gotten from the user. So I think understanding that content, that idea of belief as not necessarily representing truth but representing probable truth, and this idea of an epistemic context as being something that we can point to the belief and say this is all the data, evidence, probabilities we're collecting so that we can try and determine whether this belief is actually accurate. So that's that difference between ontology, which is facts, and epistemology. We use the epistemic context as kind of the inputs into the epistemology so that we can then produce new beliefs. And once we know those beliefs are likely true, we can produce ontologies, which is we can produce facts about the world from those beliefs. So that's, I think how that philosophical framing might help better understand this data model and some of the choices we made.

Yeah, I'll add in before we dive in super deep. And thank you Jonathan and Dean for the flyby on predictive processing. The difference between ontology versus epistemology as facts versus process of acquiring facts and the process how that shaped our, our thinking. Another addition I would give the audience is one of the most fundamental underlying constraints or I guess, measures of success. I suppose I should say better is we want for a given user, for their given context to be able to accurately predict what is the next best question for you to evolve your worldviews, your belief systems. So that's part of our requirements and that also shapes a lot of our technical direction and implementation and design decisions. So just some of that context. And then before we dive in real quick, I'll just frame again the happy path use case we'll be talking about because a lot of these Concepts intersect with many different domains to different levels of breadth in depth. So it's really complicated. But we have this high level use case that we'll talk about, which is like the belief mirror chat application. So imagine you kind of have the chat application like ChatGPT or Claude, and you're just talking to it, but underlying it, epistemic me under the hood is extracting your belief systems, updating them, and then we could also show you a mirror of those beliefs. Good example is you, hey, we asked you a question on what do you think about sleep? And I say something like, I think I only need three hours of sleep a day. And so there is obviously extra data out there that maybe challenges that belief. And then we'd be able to process, oh, we took that statement from you and we extracted the fundamental belief and now we can quantify the subjectivity of that and we could also reconcile that and give you new information backed by science and research to challenge that. So imagine in your heads a chat application and by the time this podcast is out, I'll try to post a screenshot in the video so you could imagine this and visualize it. But as we talk through these more technical concepts and tie together philosophy, belief systems and technical implementation, just keep that in mind. Like we'll be trying to talk through the lens of that Happy Path chat application where you just talk to it. We extract your belief systems, we mirror that. Cool. So Dean, I know you have your screen up. You've had it up for a little while and you're ready to go. Do you want to explain to the audience how you're planning to structure this? Like, what's on the left, what's on the right?

Yeah, sure. I think what could be helpful here? We were talking about this before we started the podcast was almost think about this like a systems design interview. So if you guys aren't familiar with the systems design interview, it's kind of used in engineering interviews where someone will basically tell a candidate, hey, we have this system with this set of requirements. How would you go? Go ahead. Building it typically from scratch. Luckily for us, we already have a framework which we can use to help us think about this problem. So I think one way that we could structure this is we could start off with this example app that we were just talking about. Maybe Robert, you could play the role of user or someone who's, who needs this application built. You can walk us through some of the requirements and then Jonathan and I can talk to you about how epistemic me using our data model using our interfaces can help you solve that problem from the perspective of you integrating with our SDK.

Cool. So I am the builder of an application where I would like more personalized models of my users to leverage.

Okay.

I'll ask questions about the system design. Yeah, that works. And then I'll also say, Jonathan and Dean, if I'm not asking good enough questions, please feel free to just be like, all right, yo, here's actually what the next system of design interview question should be.

Yeah, sure.

But yeah, I guess for that problem, how might I go integrating this with my existing user data? So let's say I'm building an application. I already have a hundred thousand users, and I would like to be able to personalize value to them more. And I think epistemic me could help me do that. So how exactly would I go about as an engineer in my own code base, you know, in GitHub, how do I actually implement your SDK?

So we said that you have an existing user base of about 100,000 users. You said 100,000 users.

Yeah, let's go with that.

So.

Yeah, and I, I, I assume this, this user base that you have is a set of users who are interacting, let's say, for this, like, Whoa.

Whoa, whoa, whoa, whoa.

What was the name of the application again?

Let's call it the Robert Health Coach application. The Robert. Let's call it that. Yeah.

So the Robert Health Coach app is typically collecting user data about, I would assume is collecting user data for these a hundred thousand users. Users are maybe interacting with it. They're supplying the app their health data, and then the application is maybe making some recommendations to them based on that health data about what's the next best action to take. But historically, you, Robert, have had to hard code all of these rules. Like, I'm going to, going to make some hypothetical assumptions here, right? Like maybe you get access to the user's HRV score, but you actually have to go in and you have to say, you have to provide an opinionated set of habits that you think that that user might pick up. So you have to code in and go in and code an exercise plan that relates to HRV and then essentially tells the user how effectively they are measuring up against some score or metric that you have decided. However, there's no existing system today that just lets you interact with those users and spin up those metrics on the fly based off what they actually care about in the context of maybe some larger Robert Ta philosophy that informs them which paths they should take. So the Idea here would be we have raw user data. How do we start to instantiate these essentially observation contexts which observe some component of their data and then spin up these habits that measure how well they are performing against them all in this very generic way that requires you have to do you to have to do little hard coding in your. In your implementation, setting up the requirements. Right.

Yeah. So it sounds okay. So a hundred thousand users and I would like to get out of the business of hard coding for every HRV threshold and xyz. This is logic, our policy. Some amount of prescriptive plan for new habit changes to the user. I would like that because that, that costs me a lot right now to even maintain that. To go in to that kind of spaghetti and like change it is. It's a lot. It's annoying. I don't really want to deal with that. So you're saying that with epistemic me, how it's set up, I wouldn't have to deal with that problem anymore.

Yeah, exactly. And we're also assuming that you. That it's not completely generic like what makes Robert's health coach app different than any other health coach app. And it's kind of the philosophy that you have yourself. Right. Like Robert, you're a content creator. You've produced a lot of articles. If I look at the set of all of your content, there's a distinct Robert TA philosophy that maybe doesn't prescribe exactive, prescriptive plans, but we want to help inform the type of plan that we're giving to a user. Right. So you have, I think the respond, don't react strategy that might help a user, maybe not in the health context, but maybe in a more mindfulness context, like emotional. Come up with a certain emotional plan to help them deal with their inner turmoil. So this idea of your philosophy is still really important here because if what it's. It's what makes your app unique. Right. You produce this really great philosophy. But we don't want you to have to code up every single way of measuring how a user meets that philosophy. We want it to kind of happen organically.

And as we talk to you, this. Yeah, that is one of my concerns as, as the owner of this application is personalization to me and not to other people for my followers, because I'm really concerned with brand trust. Like if this sounds too AI or if it's too generic or if it's not really me, then I think my followers would not trust me as much and that that is like a core fear I have As a brand is like my brand credibility is lower. So in using these AI tools, like, I really want to know, is this personalized to me. The other thing I want to know is, so I make new content all the time. It sounds like these belief systems, you can feed it content and extract beliefs and then update the lease system behind the Robert model cell. But I imagine then I'd have to get a developer on my team to retrain it every time. That's an operational cost. Is there a way to automate that in the future?

Yeah, exactly. Yeah, I think you hit the nail on the head. So we've structured and I think this kind of relates to the data model that you're seeing on the right. We essentially have this concept of an epistemology, which is going back to the idea of process. That epistemology can take in some events and it can produce new beliefs. So you, Robert, are constantly updating your philosophy with new articles, new books, new pieces of content. We would listen to that stream and we would run our epistemology over it to consistently update your model of self so that we can just kind of listen to you and just sort of update alongside you. And that that idea of like event listening is really powerful because as we'll explore later in this kind of architecture, the exact same model applies to your users as well. So from a, from a philosophical perspective, we're building your model itself exactly how we're building the models of your users. We're just treating the resources kind of differently in order to do that, to do the same thing.

Okay, yeah. And then backing up a little bit. Thank you for that. Dean. One of the questions I had is about like user data. So would this be either like a SaaS service that I hook into and then you have some level of my user's data, or do I like host this and I own all the data? What are the options available to me?

Yeah, Jonathan, also want to make sure you have. If you had anything to add there.

Yeah, Jonathan, if you want to add.

To that, I mean, I think it's built with the idea that you're self hosting and that you're controlling your user data. And I think even more ideally is that as a user you can hold onto your own belief system data. And those are some use cases that I think would be really powerful to empower as well.

Could you talk through, like give some flavor to the audience of one of those use cases that are top of mind to you?

Yeah, I mean, I think one of the places the project started, I just Imagined being able to go into cloud or ChatGPT and the interactions that I have that I can extract my belief system and that, that's like, hey, that's something I can take with me. You know, something that, that was not part of my data set that I can go into any other system and sort of like generate it into a prompt context that then this, the system can understand me better so that any place I go I have this like centralized kind of like belief system that is building on top of all of all of my interactions with these AI agents. And okay, from the user point of view that, that, that kind of vision.

But yeah, yeah, so that's like a, you know, represent me everywhere sort of like value statement to the user themselves. I think that's cool. I'm gonna have to write down that tagline too as our project evolves. I like that. Represent me everywhere. Okay.

I think they can speak.

Yeah, speaking just to add a bit more flavor to that. I think that's a really powerful idea because we can have different users running their own instance of epistemic me with their own like self hosted and self maintained user data sets. But other epistemic me instances can talk to each other. So there is an opportunity to move self models across different deployments of epistemic me. Could the user want to do that? So our interfaces are designed to do that.

And that goes into. I'm taking my hat off the interviewer hat, but that goes into just to highlight one of our core philosophies that we talked in the previous podcast, which is if we're successful with this in ten years, twenty years, a hundred years. This is a core interface and schema with which talk about release systems and transmit that data back and forth across systems.

Right.

So this becomes like a protocol like TCP. So USB 2.0, USB 3.0. So yeah, that just to give some higher level vision as to what Dean was saying. Okay, back to the use case. Robert Health AI health Coach I now understand that epistemic need could help me reduce some of my operational costs of integrating new content to evolve the belief system of the model of self, of me that interacts and scales my interactions with my community and follower base of a hundred thousand users. I guess for completeness, just for completing this systems design interview, let's just say peak load is maybe like 20,000 users submitting transactions at once. We'll just say that's peak load just to be a little complete. And let's also say that my community base is around the globe as Another requirement. It's not all centralized to just one location. So we have some cloud requirements or compute requirements that might inform performance requirements across the globe too. I'm trying to think of what else that might inform the system's design or if you had questions of me, of the interviewer to like get more constraints. Happy to answer your questions.

Yeah, I'm trying to get this. Having some technical troubles here even using the tool. So say peak users, you can see 25,000. Any 25,000. 25,000. And what was the other requirement?

Global. Global user base is globally distributed pretty evenly. No like hot zones. This is the perfect startup. Yeah.

100%.

Yeah.

These are, these are, these are, these are really solid non functional.

Could I could also add some like more functional requirements beyond the performance requirements?

Yeah, sure.

I have the desire to be able to quantify a philosophy alignment score or loyalty score of my followers. Like I'd like to know who are my best followers. Not just like how much you've engaged with my app, but how much do you actually believe in the philosophy. That's something that I would care about is some cool report like that. Because then I could email my true fans. Right. Or like my best followers and I have some level of quantification to drive that and automate that kind of thing.

Yeah, that's a great idea. My computer, like I don't know if you can hear it through the mic, sounds like it's about to explode. I think excalidraw needs some performance optimization. So let's switch to Google Drive over here. Yeah.

Do you have draw IO? That's a decent.

I do have draw. Let's see.

We're doing it live, everyone. Yeah, it's all good. Cool. This is shaping up really well actually. And this is a fun way to set it up. Dean, to do like a systems design interview. I, I think that was a great idea. Thank you for. Yes. Yeah. All right. Let's just wait for Dean to get back up to speed with all the requirements. It was a hundred thousand users, global app, 25,000 peak concurrent users. I guess we could also say then it would be Maybe let's say 4x that so 100,000 concurrent transactions at once.

Sure. So we'll, we'll say that that's your steady state volume, but we want to say that your, your max is about a hundred thousand.

Yeah.

Concurrent transactions.

Cool.

Yeah. And I think, I think we'll start to realize that a lot of these non functional requirements are gonna be more based around how you host epistemic me on your instance and how you horizontally and vertically scale up some of the dockerized containers that we give you. So there's only so much given the current deployment model that we're working with that we can help to support this. But we'll talk through some ideas there about how we support those kind of.

Ideas and then that functional requirement of I'd like to report on a quantified belief score or loyalty score of my community by user. That would be really great.

It's a good one. And I think we talked about before we have this Robert Ta philosophy which we want to embed into our quantified prescriptions. Quantifiable prescriptions, yes, for our users, prescriptions here can be generic, but in this context it's more like health based. So I think we're in a really good place. Let's start off with the functional requirements because I think that maybe sits a bit better with some of these data models decisions and then we can talk more about this towards the end because frankly I think aligned with the kind of building public sort of model, we're still working through how we would work our self hosted kind of solution. So there's still more work to be done here. Reporting. Let's start off with the initial idea I think of you have a hundred thousand existing hundred thousand users. So the first thing that the first resource that you'll want to create, Robert, using the SDK is you'll want to create a developer resource.

Okay.

This developer resource is essentially something that is representative of a developer that you have or you've hired in your system to integrate with epistemic me that will give you a set of API keys that are going to enable you to securely access your epistemic me instance. So or optionally if we host epistemic ME in the future, those API keys will be able to enable you to access those instance. So we're kind of setting this up for that dual model. The API keys are optional for your use case. If you don't need that level of security in your environment, the next thing that that developer resource, once provisioned, is going to enable you to do is it's going to create enable you to create models of your users. So you mentioned that you have a hundred thousand users that you want that you have. You're responsible for kind of instantiating those hundred thousand users through like a batch job of some sort. And those hundred thousand users will all be assigned unique IDs. You'll also probably want to attach some metadata key that enables you to Join them with your unique ID that you have near it. So either you can use our unique ID or you can attach your unique ID such that you can reference it at a later point in time. So those hundred thousand models of users are really just empty shells at those points. So kind of key to our system is the fact that we can support multiple models of self per user. So the model, the model of self to user relationship is a one to end relationship. And this is the idea really that like users have cells that live in different contexts or operate in different contexts. This is actually kind of like a very philosophical idea. We kind of wanted to start off with the idea that the self is impermanent. This is kind of like an Eastern kind of like idea. It's also, it also shows in predictive processing. It shows up in like, you know, Kantian metaphysics as well to some degree. But this idea that users have different selves that are useful for different things and different contexts. So that one to end relationship was really important early on in the system to say user can have multiple selves. So what you'll probably want to do is start off by creating a single self model for your user that's related to your, your specific context. So that, that, that, that one to end relationship maybe isn't super useful. So let's just assume for your use case you're operating under a one to one relationship model.

Okay.

So we essentially have multiple ways of populating the belief systems of a user. Our primary one that we tend to operate with is the dialectic. So we kind of have two use cases here, one where we might want to backflip fill your user.

Like I have existing data, I want to backfill the belief systems with my existing user data.

Yeah, and we also have the case where as we want to build this new feature in your system, kind of like either a chat based system or kind of an existing stream that as the users continue to interact with your app or we onboard new users, we have a way of updating their models over time. And I think the chat based use case is something that we'll dive into first, how we can deliver chat seamlessly into your application. And then we can talk more about the backfill use case because that's actually more related to how we would model Robert. And we can kind of apply the same architecture to do the same thing with your existing users. So we have this chat based system that operates off a dialectic. Maybe I'll hand it over to Jonathan. Jonathan, if you want to start kind of Talking about the dialectic and how that works might work with the system.

Yeah, sure. And this is the real time chat case, so not, not the historical. Great. So I think if say the user's entering a chat session with the AI health coach, the idea is that we have a dialectic and the dialectic is like a thread on a chat, so that might be like a particular session that the user's opening up. This is familiar in ChatGPT or Claude, you know, you kind of like have these different threads. So in the dialectic, you know, is a way of updating the beliefs around a certain topic. So that's kind of how the epistemic space is modeled is that the dialectic is trying to like generate a set of questions and answers. It'll update the belief system for that sort of topic surface area in the model. We are going to like, you know, create, create a dialectic so that, that kicks off the session and then each of the sort of like question answer interactions we call a dialectic interaction and the, so the dialect interaction comes back to that model of self. And it's sort of like if you think about philosophical being in the world, like you know, what does it mean to be and what is each sort of like cycle of perception where I am actively understanding the world around me, testing hypotheses, collecting information and updating my beliefs. That's like one dialectic interaction. So each dialectic interaction is a cycle that is going to update the belief system. In our model we kind of have overloaded dialectic interaction at the moment to where it takes on all of sense making.

Right.

And but like one instantiation of it is a question answer dialectic. And so that's what we're going to be talking about today is that the system is going to ask the user questions they're going to answer and in each of those dialectic interactions it comes with a preloaded like next question. So it's actually going to evaluate the epistemology of where the user's belief system is at. What do we want to learn and try to recommend a question. But as a your health in your code you could sort of like just override it and put a custom question in place. And so that's an option of using the each dial dialectic interaction.

Oh wait, wait, I think that's an important bit. So you're saying every dialectic interaction I'll get the next best question from epistemic me, but I could also have the ability to override that question in my implementation.

Right. We've Called that, you know, there's like kind of a dialectic strategy, you know, or an overriding philosophy by which you want to have this dialectic and it's going to guide the question and answers. Um, and you know, we're, we're trying to like, take steps that we can ask the next fixed question, but also just make it really flexible. So, like, you know, whatever you want to do in the client, you can ask your question and it's going to update the beliefs accordingly.

Okay, cool. Yeah, that makes sense.

And go ahead.

I was just going to say, like, that's flexible for my use case of I want very personalized onboarding because I care a lot about that experience. And so there are specific questions in the initial interaction that I would want to have control over and refine myself over time. So that's good to know. Sorry, Dean, go ahead.

Yeah, and I was just about to say, I think like Jonathan mentioned, we have this next best question that is influenced by philosophy. Robert, that could be your philosophy going back to the existing Robert Tau philosophy that we already have in the system.

Okay.

Or the Robert toss up model.

Really cool. Okay, very cool. Thank you for explaining dialectic to me and the model. Okay. And then for the audience, kind of taking off my interviewer hat, if you're still not sure what dialectic means, just think about questions and answers. Like I, I ground it in the very useful concept of when I ask my. Tell my therapist about something, she asks me a question, I reflect, I give her a response that that's what a dialectic is. It's just like a question, some new piece of information, you reflect on it, you update in your belief systems, whether that's happening programmatically or I'm sitting in my therapist's chair and listening to her, and then I'm updating on the fly in real time, my own brain. So that, that's, that's that coming back to the. The model. Dean, is there anywhere you want to take us going forward with the use case we're talking to?

Yeah, I think I'd want to talk about the belief system and the beliefs and how they all kind of come together and going back to the idea of dialectic as an epistemology. So Jonathan just laid out how we can use the dialectic, which is one framework for updating the belief system. And in fact, we would say a framework for updating the belief system is in epistemology. So he talked, he walked us through one epistemology of how you can actually update the user's beliefs. And I think maybe we should try to dive into what is a belief system and like, how do we kind of tear that apart inside epistemic me? So we would say that a belief system in epistemic me kind of consists of two things. Let's move this over a bit. So belief system in epistemic me is kind of built out of two things. It's built out of a set of beliefs that the user owns, and it's also consisting of a set of epistemic contexts. And you can see, on the right, you can actually see the proto data models that you'd be exposed to in the SDK for those two concepts. So we have a set of epistemic contexts, one to end epistemic contexts and one to set, and 1 to 10, 1 to n beliefs. And all of these are associated with the user. Now it may not be exactly clear why we chose to model it this way, but we can go into it a bit more. So you can see here we have the belief system. It can have a set of many beliefs. It can also have a set of epistemic context associated with it. And a helpful analogy that I think is really useful when thinking about the belief system is actually a database. And I'll use another analogy to describe how you might structure a database. So typically in a database you have, or let's say like a normal SQL database, you have a heap of information. So you have a ton of facts, a ton of rows, a ton of records. However, oftentimes it's really expensive to find the data that you want inside that record, right? So if I'm searching up like a million records in some library database, I would have to look through all the books to find out the exact book I want. Oftentimes the way databases solve this problem is by building a secondary structure called an index that essentially is a much smaller data structure that points to all the, all the, all the records in the database and provides you a way of accessing it, typically in like logarithmic time. That's the fancy, you know, CS term. Basically faster than scanning through every single item through using nice data structures. So the analogy here would be beliefs are the heap of beliefs that we've collected for a user. And epistemic contexts are that secondary data structure we've built on top to enable you to do epistemology over those beliefs, to actually enable you to better update those beliefs. So to go back back to the database analogy and actually use another analogy for like how libraries used to work, like we're all familiar with the concept of indexes and libraries. Right. Like, that's helpful to find you the exact book that you want. But oftentimes it doesn't mean that every single book in the library is true. Right. Oftentimes there's older data, there's historical data that doesn't really make much sense. It's both the role of the librarian to help you find the data that you want, and in olden times, it was oftentimes the role of the librarian to find you the information that was most helpful for what you were doing. Right. You would go to the librarian and talk to them about your specific use case. So in that sense, the librarian kind of plays this role of this epistemic context where it helps us to evidence and find out whether those beliefs are actually true or not that you hold in your system. So going back to what an epistemic context looks like, we have a bunch of different ways of scaling up epistemic context based on what kind of epistemology you're running in our system. So we had talked previously about predictive processing. How do I run predictive processing, which is kind of this framework for thinking about self models and beliefs over my belief system, while I instantiate a bunch of predictive processing contexts that then point to my underlying beliefs and that helps me create these observation context that enable me to instantiate habits and other types of kind of information like state and probabilities, whether those beliefs are falsifiable belief or just belief statements. It helps me contextualize all my beliefs in terms of the predictive processing framework. And then another thing is, we've also talked about these other kinds of context that we want to add over the system. So how do I model my belief as a tree? Right. So, for example, sometimes we have an initial belief that is then supported by multiple other beliefs. So this is actually quite powerful in the concept of, like, Eastern philosophy. Oftentimes we try to model the belief of, like, what is the self? Well, that starts off with like this, like I, like I am this kind of person. And then there's an assumption underneath that that says, you know, the self was true. And if you think about it, a lot of beliefs follow this pattern. There's assumptions that lead to higher beliefs. So how do you model that kind of tree, like data structure over beliefs and how they point to each other? So that's another type of context you can have in the system where you say, let me model my beliefs as a tree, not in the belief, but as a separate epistemic context. That enables me to do things like validate my assumptions, or you could do it at the same time as predictive processing. So that's the flexibility in the framework, this ability to do multiple epistemologies over your beliefs.

So one question there to ground it in the AI health coach use case. So like my goal with them using my AI health coach is that they get healthier. Right. And so the hypothesis here is that whatever recommendation that we make via the Robert AI Health coach built on epistemic me is giving you that recommendation that meets you right where you're at. Like meets the user, where they're at in their context, their belief systems of the world. So one example might be the whole sleeping use case. It might be like, I have a user who has operated for a long time with many foundational beliefs that they built up over time that they could go for four hours of sleep a night and still be really productive because they have a successful career, they're fit, whatever. Let's say they have that. So like, let's walk through that example. Dean of like, they're having an interaction with me. We've teased that out. We've asked them a question, like, how much sleep do you get at night and what do you think about sleep? And we teased that out. How might episode me showed the underlying stuff because that stuff is interesting to me because the underlying limiting beliefs that power, that one in the tree are the ones to tackle. Right? In that example.

Exactly.

Yeah.

That's a really good, that's a really good way to frame it. And I think let's talk about how we have the. You mentioned two things. You mentioned the user's existing belief system and you mentioned this new belief that we kind of want to plant in the user, sort of like Inception. You know, Inception is framework. We want to plant it and we want that belief, that new belief to grow. And we wanted to kick out older beliefs in the past. Yeah, right. So that way that new belief would actually come into being is through your Robert Ta philosophy. Right. Like, Robert, you've probably built up this history of wellness. This, you've written things about the importance of sleep. So as the user interacts in the dialectic interaction with you, suddenly they realize that they're being convinced by you that this belief might actually be true. So at what point in the dialectic interaction we see that in your belief system we end up with this new belief at some point in time that this belief is that better sleep leads to better performance. Now, that belief is still a nascent belief in Their system. Right. It hasn't been evidenced yet by any belief. We want it to grow over time and we want it to invalidate older beliefs that are like hey I can just, I can just grind, grind, grind. And that's the only way to be high performing individual so that better sleep leads to better performance is instantiated as a belief in the belief system. And I think actually the predictive processing context, just to ground this for a minute, Predictive processing is our current framework for kind of creating kind of conditional probabilities and storing evidence over beliefs. And I'll, I'll hand it over to Jonathan in a second to talk more about that. But we would instantiate a predictive processing context over that new belief that we create for a user. Either the user says hey I really want to dig into this belief and evidence it or you Robert, say hey actually this Robert taught philosophy. Every new belief that I create in the user create evidence frameworks around that and that would auto instantiate a predictive processing context over that. And now as the user continues to interact with the system either through events. So for example, maybe one way of submitting events is that their Apple Watch is constantly producing health about their sleep scores and their metrics. Another way is they just report in the dialectical interaction every day hey, this is how I'm doing today. Or it's ideally a combination of both. We have data coming in many different users. We're going to run the predictive processing epistemology over the predictive processing context we've created for this belief. That epistemology is also going to say hey, let me update the evidence for this belief. But also let me scan through all of your other beliefs and determine which ones are no longer valid. And in cases once we've successfully evidenced this belief, we are going to archive your older beliefs. So it's still going to be visible to you. You're going to be able to look back in time and say this is how I progressed. And you Robert, you're going to be able to see what is the rate of change in my users beliefs over time. Time you can actually see how effective your philosophy is doing. You can use that to improve your writings and update the rate of change of users beliefs. But essentially that, that, that archival, just talking about why that archival or temporal aspect is important but that's kind of how that epistemic framework works over this user's beliefs to evidence it and then into the implementation details of how that works I think would be helpful to talk about how predictive processing works. And yeah, I think Jonathan could really speak well to that since he's implemented so much of our predictive processing framework. Jonathan, would you want to talk about that?

Sure.

I think, yeah. Coming from the standpoint of like, let's say we're asking about, you know, beliefs about sleep and performance, we're at a point in the interaction where we're creating dialectical interactions that are creating belief statements. And so if you look at the model for beliefs, there's a type of belief and there's the type of belief that is a belief statement, there's a type of belief that is a falsifiable belief, and there's a type of belief that's a predictive belief. And we're sort of trying to graduate the beliefs through that cycle. But when it starts out, you can think of like, you know, what does this question mean that you said? And it can decompose into sub questions. So you say better sleep leads to better performance. Well, what is better sleep? What is better performance? You know, we can recursively deconstruct, like, what, what the question means and then what beliefs and sub beliefs are sort of like, relevant to the topic space. And, and so we can create, you can think of a tree that is generating belief statements, and we go to a certain level of depth and a certain amount of breadth to just make sure that we've had the user express themselves and get concrete belief statements that are adequately kind of defining that topic space. And then once you get to that point, the system kind of has a threshold where it says, like, well, these beliefs are, we're ready to move on to the falsifiability of it. And like, let's start to structure these beliefs in a way that we can collect evidence about them. And in the predictive processing context, what we do to collect evidence is we, we create an observation context which is like, sort of like, what is the phenomenological state that you're going to collect evidence? How are you, as the user, going to look at information from the world and then verify that this supports your belief? So in the example of sleep, I might say that better performance means that I have a higher or regular HRV in the morning after a good night of sleep? And then once I'm ready to put that in observation context, I can say that when I take a snapshot of this HRV in my app and I see that in the morning, I know that I had a good night of sleep and that I have good energy for my morning, for the rest of the day, and so that can be a piece of evidence. And so that's one thing, like the user's phenomenologically looking at it's part of the world. Another thing they could look at is like a paper on sleep and performance. Right. And there are paper. And so them reading that paper and like, you know, using that content as evidence can also support the belief. And so that's how we kind of graduate belief statements into falsifiable beliefs and make them evidenceable from that standpoint. And then as with you managing the coach, you can take your users through the cycle where it's like, hey, let's, let's work together and collect evidence to support these beliefs that you have. And you could be driving them towards a particular sort of like thing that you've discovered and part of your philosophy, or you could have them on their own exploration to like really just prove these things themselves. Kind of like the Socratic method and had dialect that was used, you know, in the beginning of philosophy.

Yeah, that's interesting. And one thing came up as you were talking through this, which is going back to the whole goal. My whole goal is my user in this case becomes healthier. Right. And all of the foundational pieces and underpinnings that we're talking about with epistemic me as the SDK to build these models themselves to extract belief systems, to quantify subjectivity and beliefs will help me get to how can my AI health coach, representing all my philosophies, my model of self, give you the next best question or data that helps you evolve your worldviews and your beliefs and that activates you to take action and be healthier. So that's just underline to zoom back out for everybody. That is really goal and this is how we seek to achieve it with the predictive processing algorithm implementation on for epistemic me. So in that use case, I'm kind of wondering then, like, my mind's going to, hey, one user is going to have a very different view of sleep than another user. And then between those two users too, another layer of complexity is they might be convinced by different things. So Jonathan, you mentioned a user might be convinced by reading a paper that's like a very rational thinker. There's different influence styles, like, I like teamwork. And that's why we see all these accountability leaderboards like whoop as communities around this. Or it might be like I see a vision of myself like weight loss commercials before, after, oh, I see that more confident me, I want that. Or these other influence Styles. And so I guess one question I had is, is that a possibility that we can enable, which is injecting or understanding and extracting. In understanding and extracting the belief systems, can we also map things like what type of information displayed, what kind of way would best make them healthier or evolve their worldviews in a way where their actions align with more healthy actions? Is that something that epistemic me is set up to do is like be able to ascertain their influence style? And then whatever the AI health coach is like, we get a certain amount of data and then we ascertain, oh, you really like papers, you like rational things, you like logic, you like data, you like citations, you like research. And so then maybe the AI health coach would understand, oh, for this particular user, layer the evidence or the suggestions with links to scientific papers or creditable authority. So curious. I know that's a lot, but just grabbing it in that very functional use case, is that possible basically with epistemes?

I think I would say so. And I think it's really that specific use case you talked about. That idea of given the data that we've collected about a user, how do you interpret that? I actually see there being the role of you, the coach in the system. Right. You've laid out a philosophy of how you want to. Like we, yeah, we call this kind of like a dialectical strategy. And I think strategy is a good term, but it's a. There's a strategy of how you want to interact with your users and how you want to classify them or how you want to guide them into certain scenarios. So you, Robert, can write that philosophy. And that philosophy would be injected into the dialectic, or it'd be injected into the way we update beliefs to push or nudge the users in a certain direction. We, I think, and also as epistemic ME graduates, like the use cases you're talking about, like there would probably be like a scientific method philosophy, or there would be like these set of philosophies which we've cataloged such that you don't need to write it yourself. You could also just pick from our existing catalog of philosophies and then also interact that are get instantiated as self models. You know, we have like the scientist self model who perfectly embodies the idea of science, could start to be used as a substitute for you. Or you could even say, hey, take Robert, merge it with the science philosophy, produce a new self model and have that self model interact with the coach. So I think it's this idea of communication between self models. Again, going back to the idea of philosophy because at the end of the day you have this reference point. You want to interpret users ideas through.

Through.

That's how we instantiate self models that talk to each other and help to contextualize each other's information. And that's how you would build this application, a model of Robert that contextualizes your users and continues to push them in the philosophy that you. You align on.

Got it. Okay, now that's. That's really helpful. So just reflecting back what I'm understanding. Epistemic me is built for the use case of let me analyze and interpret my users and who they are and what might influence them best. Like what might be the best strategy and that could be then personalized downstream for something like let me suggest you a habit or let me suggest you some new data to evolve your worldviews on health. That's really cool to talk through. It might be helpful for the audience because you mentioned plugging in the philosophies and I think that's skipping ahead a little bit to like, I believe what we call it perspective taker. But might be good to just talk about that because I think that is a very useful concept and it is something that we're. Oh, sorry, go ahead.

Can I expand on the last question a bit?

Yeah, go ahead.

What you're asking there, like, relates directly to where I thought that surfacing beliefs would be sort of like a useful thing, especially in health. And like, I'm a health nut and I've been exercising, you know, maybe six to ten hours a week. And while I'm doing that, I'm listening to podcasts, and a lot of them are health podcasts. And so I was hearing a lot of these health influencers get the latest research, make recommendations. And I think even when like, you're like, say a Peter Attia or Rhonda Patrick, somebody's formally trained to like, review health, health papers and make recommendations, I would see that these guys would always look at things differently and come up with like, sort of different, different recommendations. And when it comes down to health, I think the question of like, what is good health? Is equivalent to the philosophical question of what is it to live a good life? It is the foundational question. And so like, what, what you get there is that if you're trying to find a health influencer that meets your worldview, you really have to find somebody that like, sees life the same way as you do. And I think you can, you can exemplify this, say Maybe a Peter Attia versus a Brian Johnson where 1 of 1 Peter Attia is trying to live like, like a hundred good years of health span and he's optimizing for the Centurion Olympics. And so his sort of teleology is that like humans are supposed to live in this world, they're supposed to live a healthy life, they're supposed to have health span up to the end of their life and they should be able to like do like really functional things at age 100, but then die shortly after. You know, whereas as Brian Johnson is saying, health is being basically immoral.

Right?

Like I want to live for many hundreds of years. I want to like put myself in a, in a place where I can evolve my health through AI. And so a new paper comes out on like epigenetics and you know what it means to take Metformin and each of these perspectives that they're going to use, there's a broader philosophy they're going to use to like interpret that data and say, how does this impact my health? And so if, if Robert, as an AI health coach wants his users to be healthy, you know, you sort of first have to align on this whole question of like, what does it mean to be human, what does it mean to live a good life and what is health? Which are not insubstantial questions. So we need, we need to break these down to like the underlying philosophies and epistemologies.

Yeah, yeah, let, let, let me add something else on that, that that is really important and tying back to what you said, which is there's a foundational question which maps to what are the foundational beliefs that you have. And that's actually why we've designed epistemic need to be quite flexible because we foresee a future where as everybody's a special snowflake in the world, right, Everybody's going to be making their own choices. And one of the underlying philosophies of epistemic me is we want to enable people to make their own choices for their own personal lives. Kind of pick your own adventure through life, right? Whether it's health or mental health or emotional house or, or, or whatever vertical they're on. And so, yeah, like just to summarize what you said, there is one end of the spectrum which is, hey, I'm very willing to just index towards Brian Johnson. I want to be immortal, okay? So immediately I'm going vegan. I'm gonna hyper optimize every part of like zone two cardio every day. Like I'M gonna black out curtains in my room. Like, there's, there's, there's a, there's a tension point there between like, let me have normal societal interactions with the world. So like one example here is I've recently started only eating within a six hour window from 10 to 4. And so that means no more dinners for me, social dinners or if I go to social dinners, I just will drink water or something. And some people might find that weird. Some people might not agree with that philosophy. I do. Some others might, they might not. And so epistemic me would allow any user to sort of dial in what is the philosophy I want to follow that's good for me and how I define better or healthier or happier. So that's a really good point of distinction. Jonathan Dean, did you want to add anything to that before?

Yeah, I think underlying that whole conversation is it like really brings up the importance of what we talked about earlier, which is cool. Quantifying subjectivity. Like you could almost argue that, like that, that, that difference interpretation between Peter Attia and maybe I think it was Andrew Huberman or Rhonda Patrick. About a single piece of evidence is almost like a distance metric that we're defining in the system where we can say, oh, how are, how, how, how far apart are these two ideologies with reference to that one resource, that one piece of evidence. So we want that we need to actually calculate that distance metric. And that distance metric is subjectivity. And then coming, coming back to your use case, Robert, like, we can calculate the distance metric between you, Robert, your philosophy and your users. And we can report that back to you so you can understand, are my users becoming more aligned with me? Are they moving farther away? Or you can calculate other distance metrics between other, you know, more foundational ideas that you pull from our catalog and you can help to understand how those users are aligning and how, how close or how much they're clustering in certain ways. And that's kind of the analytics aspect of that.

So quick question there. Back to the design system interview. Put it on my like I want value right away. Hat as an application developer for this AI health coach, I'm building for like 100,000 followers. How quickly using epistemic me could I and maybe this is moving to the other use case now. How quickly could I backfill data to get that distance metric or that philosophy alignment score? Cause I see that as being very useful for, you know, in this use case for me to automate email funnels or marketing automation or that like offer or letters, that kind of stuff. So I'm curious, like how quickly could I bootstrap it? What would the steps be?

I would say, I mean being honest, it's not a feature we've built yet. It's a feature, it's a North Star where we want to go. Right now we're focusing on like how do we build a dynamic model of self that works both for the Robert use case and the user use case and how do we get Robert to interact with users, to guide them in a meaningful direction. That quantifiable, quantifiable overlay I think is something that we're going to want to build in the next milestone for the project and create something that's really exciting, but I just don't have the capability for.

Yeah, and it basically it's on our roadmap is the answer. It's clear that the backfill use case will have to be solved for because there will be previous.

I think we will support the batfill use case. We just won't support that quantifiable distance metric. So we should be able to take your user data and create models yourself for them. We just won't be able to compute the distance metric between you and their philosophy or cluster them together.

Yeah, I have an approach to that in the predictive processing epistemology which is that measurement and disadvantage is called belief coherence. And so in the observation context is a specific way that like we can take like if we are in the same situation and we see the same data, do our police or do our beliefs predict different outcomes? And that is I would say not so far away from being able to implement is like taking two self models and like know having them both generate an ontology in the same observation context and then create a belief coherence score between them. And I think like when it comes to like how well do we understand one another, that belief coherence store is sort of like one of the very North Star, you know, features and that, that we're trying to use this tool to, to, to accomplish.

Okay, cool. Yeah, so it sounds like there's a path and we'll, we'll probably talk about that later once we get together. Yeah, okay, great. So we went through this example and I'm now kind of looking at my notes. I think we walk through this pretty deeply and I have a good understanding of like how this would solve for the, you know, build an AI version of ME use case that's hyper personalized, how this would solve for personalizing recommendations or next best question or a health roadmap or next healthy set of habits to my user base. And then I guess one other question before we leave off of the functional stuff and then have some additional questions to close out. One of the question is around the what do you call it, analytics part of it. So I'm having kind of a hard time imagining what sort of analytics I can get from this. So and I know you know personally we we haven't designed the visuals yet for the analytics, but this part could be a little bit more free form. We'll we'll come back to this and probably update this in future episodes. But how do you guys imagine what analytics might be useful for this use case? Okay, so unfortunately the answer to that we'll have to wait until a future episode because as I was going through post production and editing this, the recordings dropped after that question. So for the future you could look forward to how we think through the analytics, the analytic interactions or analytics and visualization interactions and stuff like that. For now, we'll wrap this video up here and just say thanks for listening or watching this episode of ABCs for Building the Future. And in particular we talked about today our project Epistemic Me, an open source SDK for human belief system modeling and creating models of self, a personalized layer for any of your users and any individual. So we talk through our data model, we talk through technical design considerations and hopefully this gave you a really solid understanding of Epistemic Me. And in future videos we'll talk through how to actually leverage it in application use cases. Right now we really just focus on giving a foundation of our philosophy behind creating this SDK and what people might create with it. And hopefully this invigorated some thinking with that AI health coach use case that we grounded ourselves on. Thanks a lot and hope you have a great day. Take care.

Bye bye.