What's YOUR Next Best Question?

That's such like a kind of like mind bending idea. Am I just something that is experiencing information from my own consciousness? And it brings up these ideas of are we in a simulation?

What predictive processing does is it says that the self is actually using this generative set of beliefs to create its world. Like. So there's no like world model, there's no common world model in predictive processing. There's just the world onto the self that is being created and generated in real time.

How do we come to form new beliefs? How do we know that our beliefs are true? And what are the underlying processes that enable us to actually do that effectively?

To quantify subjectivity in beliefs, to then bridge the gaps of beliefs between two people or parties or models of self. I think that sets up a scalable way to solve the problem of how do we get shared understanding across humanity?

If you reduce your belief system to zero those things.

These more abstract concepts apply to utility in epistemic me. So this is a broad range of topics. I will try to do my best at wrangling us towards a direction. And we do have some questions. So I'll go ahead and just start with the very first one and then dive in. So epistemic knee, that was a name, uh, that it, it really jumps out at you, you know, because epistemic is typically not said in normal conversation. And so when I first heard the project name, I was like, oh, that stands out. So, very first question for both of you, why epistemic me as a name? And what does it mean to you? Dean, do you want to go first?

Yeah, sure, yeah. So the epistemic part of it refers to epistemology, which is the branch of philosophy that deals with knowledge and how we create knowledge. And the me part, I think, refers to the, to the models of self that we're trying to build with the project and put those two things together. And it's, how does epistemology help us create better models of self? So epistemology, the way I view it, and perhaps I think the way that we've kind of arrived at framing what an epistemology is, is some kind of process that is responsible for creating new knowledge. And the way we kind of backed into knowledge was the idea that our beliefs are ultimately are the best representation of our knowledge or a self's knowledge. So how do we come to form new beliefs? How do we know that our beliefs are true? And what are the underlying processes that enable us to actually do that effectively? And if we look at the selves that exist around us? What are the underlying processes that are enabling them to create new beliefs? And then can we model them in a system to create models of self? So that's what epistemic me means to me. I got really into epistemology from the beginning of infinity. He lays out to me. I think he really articulated to me what was epistemology. And he proposes a particular epistemology around generating the best explanation for things in the world. So that influenced me very deeply. And I think, yeah, was it.

Where did you first learn the word epistemology? Dinam. Just like that came up in my mind since you were saying that.

Yeah, I think I. I got into philosophy from a bunch of different. From a def. From a bunch of different angles. I think my first introduction to this to philosophy was in reinforcement learning. So taking some reinforcement learnings in my master's degree, and I found it so fascinating how much of the meta conversation about reinforcement learning was how does an agent make sense of the world when all it can know is the information that it perceives? Right. That's such like a kind of like mind bending idea when you really think about it. And it kind of planted the seed, I think, deep somewhere in my conscious that. Is that true for me as well? You know, am I just something that is experiencing information from my own consciousness? Like. Like what does that really mean? You know? And it brings up these ideas of, are we in a simulation? And it's interesting because when you code one of those agents, you realize, oh, I'm creating a simulation for this thing, you know, and what, what would it mean for this agent to be conscious in an environment where I'm like killing it 5,000 times? So that. That planted that seed. And then I got deeply into meditation a few years later, and an outcome of that was realizing or coming to grips with this idea of no self, or this idea that consciousness is the only thing. It kind of varies from different traditions, from tradition to tradition. In Buddhism, it typically centers around this idea of emptiness. Whereas in some of the more Vedic traditions, they kind of reify this singular consciousness as like all there is. Um, and in a way, maybe call it God. But the idea is that we're kind of reducing all our beliefs to realize that there is one conscious entity that is producing all information. And so this idea kind of sprang alive again. And then I was super interested in figuring out, well, what did the. What Western philosophical tradition have to say about the same idea? And you kind of get into ideas proposed by, you know, Hegel and particularly Kant, who has These really fascinating ideas of like the numina and what is the, the balance between the known world and the not known world. And you realize that in many ways he's saying the same things, just maybe without some of the. Without some of the incense. Yes, but yeah, so there's this unifying idea I saw across many different traditions, practices, sciences, um, and that kind of got me interested into philosophy. And I thought that epistemology was one of these really this, this, this philosophy was this thing that was living and breathing and producing philosophies as a side effect. So that's, that was kind of my path there.

Very cool, thanks for sharing that, Dean. Very interesting. And real quick, what just for the audience, you referenced your master's. What was your master's in?

I was in computer science with the emphasis in machine learning.

Okay, cool. Yeah, so that makes sense with the reinforcement learning. Thanks Dean. Jonathan, Epistemic me. What did that mean to you?

Yeah, well, Dean came up with the name and that's always like great from like a philosophy nerd point of view. Yeah, I think maybe like a bit of a signal to find like minds and people who, you know, really are in this direction. I think Dean did a great job explaining where the name meets the idea. And I think as I thought about the project early on, I thought about, you know, with these AI tools like make a health coach for you, they sort of like have an intuitive idea of what that means. It became a way to explain how learning was happening in the sense of. The sense making for an individual audience.

And distribution to come to my mobile form. And so the thing to do was like be a streamer's best moves because that's basically like how you win at the game is trying to figure culture. If you break that down such a hairy umbrella. Ultimately I think culture ties to individual belief, organizational culture, challenge of how do we work better together, Enable harmony.

Everybody in an organization having a big belief system that is that they can control what people, what beliefs they share with other people. But it might be misunderstanding one another. There's a lot of times you can talk to somebody about something, they might have a deep seated underlying belief that is just. It's going to be very hard to reach them with the message that you're trying to deliver. And you may not know that. And so it can, it could take a long time to get through that. And a lot of times then the conflicts comes in before you can even really understand one another's points of view. A simple one to one communication with an organization is A model. But if you think about then clustering people, ensure that you're recruiting people who are like belief aligned or belief diverse, depending on how you're trying to build up the organization, then you can optimize, like, how well you understand each other. And I think I call that belief coherence. So I think you can use this tool, belief coherence, to like, form sub modules within the organization or subgroups within the organization, or just like steward the culture of the organization itself. It gives organizational leadership a way to make sure that their message is understood and that everybody's working towards a common goal.

Cool. Now that's cool. Sorry, I gave you a super loaded question, but thanks for answering it concisely. So, yeah, for, for me, epistemic me, the name, what it means to me, and you know, like, why we're keeping it and stuff like that. I suppose what I'll say is why I'm so fascinated about the problem is all of all human problems, I think, come down to miscommunication, gaps in understanding, lack of shared understanding. And then what you said, when you have people working together, sometimes if there are different beliefs and you're trying to reconcile them, the communication could add friction. People can get emotionally flooded, there becomes tension, and the likelihood of getting shared understanding drastically drops to like almost zero percent or maybe even negative percent. Like, and you don't want to work with that person anymore. So when I think about that, I'm like, all of humanity really is kind of like working together. We just don't know it really. And there's different divides and, you know, continents or whatever. But it would be great if we could all reach shared understanding on lots of different topics. Like even, I mean, I'm just even thinking about, like, relationships are also a working unit where you're trying to get past, you know, different beliefs and meet in the middle. I certainly had my fair share of experiences there. And like, something like this, like, if you imagine the human version, it's like a therapist trying to ask you the next best question to understand why that gap in the beliefs exist. And then how could you bridge it? And so what we're doing with epistemic me is awesome because, well, why I think it's so cool is we are providing a way. And Jonathan, you said this quite beautifully the other day. I think I'm gonna steal the words to quantify subjectivity in beliefs to then bridge the gaps of beliefs between two people or parties or models of self. I think that's so cool because it sets up a scalable way to solve the problem of how do we get shared understanding across humanity. And my first exposure to philosophy was in college. And actually back then I was like, what is this shit? How's this going to make me money in this project with both of you is actually quite fun. And diving back into some of those concepts and looking at how philosophy has evolved over time is cool. And I think from a philosophical nerd perspective as well, I think the epistemic mean name is really cool. And then Dean, you've mentioned this like it could be in the future if this grows really well. You know, this is the biggest, coolest global open source project that ties together, you know, AI futurism, tech, how to think. Right. I just, I think that's like, it's so cool to think about in a future where epistemic. Me, epistemic. I don't know, sounds epistemic. X, Y, Z, anything like. I, I like. I just like that when you mentioned that. So cool. Yeah, I think so.

Yeah.

I think just to note on that information theory, information processing is really so expansive and all it can describe, right? Because there's this ability to view everything as information. So if you treat epistemology as the flow of information, there's really nothing it can't apply to. And it seems to be that idea that there's nothing it can't apply to is this kind of unifying idea that shows up in Eastern meditation. It shows up in predictive processing. It shows up in these different ideas. It's this larger theme, I think, that really fits in with the history of philosophy. Maybe that's the word that Chantan Mani is that we care about.

Jonathan, did you want to add to that? There are so many double clicks, even single clicks after information theory. I don't know if you want to add it where we dive in to one of those.

Yeah, I mean, so information theory conscient. And this is actually where I started by like introspection to subjectivity. I think this is 2006, 2007. You think about there's a sender and receiver on the channel. That's what information theory models. And what. How do you make sure that the recipient receives the message across the channel? And so it gives you the characteristics of information. And by the way, Claude Shannon both implemented information theory and he mapped Boolean logic onto the first circuits that we had. So this guy like took George Boole's philosophy of, you know, Boolean logic and both gave us communication theory and you know, the ability to like build circuit boards, which is Pretty incredible use of philosophy.

What a wire, what a gun.

Really. Yeah, impressive. But then. So if you look at this problem of subjectivity and how we understand each other, I think you have to go beyond the information processing of sender receiver. You have to think what's the semantic meaning onto the sender and receiver and how do we know that we're understanding each other semantically? And so then that's what you know, you need. I think of it as like a hierarchical representation of the internal belief system of each of the sender receiver. That's where you really can picture a mathematical way to extend information theory into belief systems and into semantic understanding. And it's a hard problem. It's like really something. You think you first look at it and you're like, why hasn't somebody solved this semantic communication problem? This seems like really needed and necessary, but it's just so many layers deep modeling themselves and decomposing it into useful predictive belief systems that give you a quantifiable measurement of ambiguity in a communication system.

Yeah, totally. Quantifying the complex is definitely a big part of this project. Maybe it'd be worthwhile, Dean or Jonathan, one of you to explain predictive processing and you know, extend off a conversation we had last night which was really rich around how can we guarantee that we will give you the next best question, you know, for your given use case, like the one we're grounding on health, like, how can we ask you the next best question to help you gain more information and evidence to form and evolve your belief systems? We talked about how predictive processing is one algorithm or one way to quantify subjectivity if you're going to allow it. Predict the next question. And I won't say I'm like a scientist and deeply know this. Jonathan, I think you, you know this probably more than both of us. Maybe you're equipped to like give some thoughts about predictive processing at a high level for the audience and then we can kind of tie that into epistemic needs and what we're doing.

Yeah, there's a lot to it and it's like a hard 101 topic. But I was introduced to it. I built this kind of theory of modeling subjectivity. And one of my advisors at T machine said, have you heard of Carl Friston? And it turned out what he was doing with some inference and the free energy principle was like very closely aligned to the work, but more rigorous in terms of the mathematics and the statistical mechanics that use to model. And what Carl Friston Noticed when he was looking at nerve cells through mri, he's a neuroscientist, was that the nerve cells don't wait for the environment to send a signal and then change the state of the nerve cell. Rather, the nerve cell's constantly predicting what its environment's going to be. And so it came. This was observed at the nerve cell level. And then he worked with a number of scientists. He got a very rich body of work extending this idea of the free energy principle. And so, like Freeway calls it, they're basically sentient beings or selves, what we call selves, anything that has, like, a boundary between itself and a set of internal states and the world. And in the, in the free energy principle, they use this concept called the Markov blanket, which is a statistical mechanics principle that separates and says, like, these are the internal states of the cell or the self, and these are the external states of the cell itself. And then there are, there are ways that sort of like, through the boundary of the Markov blanket, you can do perception and actuation. And so you get this loop. And if you go and sort of look on Wikipedia, you'll see this common. It looks a lot like an OODA loop, you know, in a sense, making loop. There's a lot of different models of this. But what free energy principle does is it takes internal external states and it creates a perception cycle such that the self is actively perceiving the world on an ongoing basis. And in doing this, like, the function of learning comes in because every prediction we make on the world then has a discrepancy which is compared with what sort of like, the actual observation was. And so we use that discrepancy as sort of like a learning tool and can characterize aspects of like, surprise. Is my prediction, like, very wrong? Am I just confused? Or was it just maybe a little bit surprising? How do I then update my belief system so that I can sort of make a more. A more accurate prediction on the world next? And I think one of the key things to understand about predictive processing, it came from a lineage of philosophical thought that's like, abandoned by analytic philosophy. It started with Kant's psychology, which is maybe considered a dismissible part of Kant's work by a lot of people in the analytic philosophy lineage. And there was. There was a philosopher called Himholtz in the late 1900s who like, built on the work and created like a version of cons velocity that had a generative aspect and then a discriminative aspect in. In a. In a system and in like our modern systems, we, we might have been all seen like these generative adversarial networks recently, where there's like, there's a probability system that's generative and the probability system that is discriminate. That's like one way to think about it. But what predictive processing does is it says that, that the self is actually using this generative set of beliefs to create its world. Like, so there's no like world model, there's no common world model in prototype processing. There's just the world onto the self that is being created and generated in real time. And extrapolate that to single cell behavior. You can extrapolate it to how systems in our body work, extrapolated to how the human works. And then you get to this point where you say all selves that just generate our realities around us based on an internal belief system. And that's the assumption that particular processing takes. And that's really to grapple with in your own lived experience. But the more you look at it this way, the more you see that this model is capable of explaining it has better explanatory power than I think, you know, considering that we all live in the same world in the same place.

Thank you so much, Jonathan. I think you said that in a way that I couldn't. As you were saying that around the, you know, predictive processing. One example more in like layman's terms that I was thinking about is I was just trying to. I was just thinking, like, where are there situations when your brain expects a certain thing, like sensory input, you expect a certain feeling or whatever. And then it just wasn't that. And so one example I was just thinking is, could be relatable when people move from like the city to the countryside, or countryside to the city, you have certain expectations. At least I have. Like when I've gone on a road trip to countryside, like, oh yeah, I should be able to go to the grocery store and grab my favorite peanut butter. No, Santa Cruz peanut butter is not sold here in. Or whatever. I was like, oh, my brain now needs to change my beliefs, right? So yeah, there's like, I think an intersection there. Not just on like patience, at least I have like, when I've gone on a road trip to countryside, it's like, oh yeah, I should be able to go to the grocery store and grab my favorite peanut butter. No, Santa Cruz peanut butter is not sold here in. Or whatever. I was like, oh, my brain now needs to change my beliefs, right? Um, so yeah, there's like, I Think an intersection there. Not just on like I suppose the algorithm, algorithmic side of predictive processing, but I tie it to my understanding of. And maybe this is my brain doing predictive processing right now. Tie it to like biases that we might have and biases can inform beliefs and that can inform what your brain thinks is going to happen next. I really like what you said about how we all don't have the same. There is no common world model or common actual schema of the world. I suppose like we're all interpreting it in real time ourselves. Dean, did you have anything to add to that? I feel like you, you might, especially when you were talking earlier about like how do we know we're not in a simulation.

Yeah, I think this is a very powerful idea that shows up across many different cultures. So I think maybe exiting the techn. The technical perspective for a second and just talking about this as like a, as an idea. I think, I think it's an idea that is so powerful because of the way in which it shows up across so many different societies and so many different philosophies, particularly in the Eastern tradition, in the Eastern meditative traditions there is a similar idea that if you reduce your belief system to zero, you realize that you're faced with just the world in front of you, just conscious reality and there is no reality that you can perceive outside of that. And as you add the belief systems in, you realize that your conscious experience is deeply tied to your belief systems. So it's this idea of single consciousness and this idea that thoughts, assumptions are influencing that conscious reality. So in it, it's a self contained system. So I think it's a really powerful idea because is very as, as Jonathan mentioned, it's highly explanatory for our own lived experience. It allows us to kind of merge these ideas of what is, what am I? And also it gives us a useful framework to start thinking about how to model other selves as well. So yeah, I think it's a, I think it's a very powerful idea and I think actually it's very interesting because I think you can map in, in computer science this is, there's this idea of reducibility, right? You'll have one algorithm and then you'll realize that you can reduce it to another algorithm. So for example, you might have these like more complex graph algorithms that are reducible to the language of breadth first search or depth first search. I can compose breadth first search or depth first search to produce these higher level algorithms. I think it's really interesting to. To. I think throughout our work in epistemic me, kind of like prying apart predictive processing and seeing what other ideas could predictive processing reduce to, or what ideas reduce to predictive processing? You know, and in reinforcement learning, there's a lot of really interesting ideas. We also have the idea of, like, the state action, like the state action space that an agent can perform, and then typically like a Markov decision process that enables the agent to determine how to act in its environment given the rewards that it's perceiving. Usually that reward is tuned by the programmer, but in many ways it's calculating a discrepancy between its expected state and the state which is sent by the system. So it'd be really interesting to see, I think if we treat these both as, like, epistemologies which are currently being used to create agents that reason in the world. Yeah. How can we, how can we maybe reconcile these, like, these different perspectives? How can we reconcile predictive processing with the bleeding edge? And how can we kind of like, tease out these subtle differences, like what makes what truly makes predictive processing somewhat different than reinforcement learning? And then how can we in epistemic me create the right benchmarks such that we can show that a system like predictive processing as actually delivering something new in the world of agents or the world of sense making? And that's where I kind of see epistemology, epistemic me as providing a lot of value. One, it, it gives us the interface to say, how does, how does an agent react in the world? You know, and I think it actually will probably end up looking a lot like this loop. Jonathan described this idea of processing information, asking, requesting information from your environment, and then making predictions like those are the core interfaces of an agent. And we want to get that interface really right. And then we want to say, okay, what are these different epistemologies being proposed by Carl Friston, being proposed by the reinforcement learning community, and can they all satisfy this interface? And once we start satisfying that interface, I think we're going to really start to tease out where the subtle nuances between these different ways of viewing agentic systems. And I think that's going to be really educational at the bare minimum. And I think at the, and, and the hopeful side of me is also that it will end up being. Producing like, very great algorithms for reasoning epistemology. She kind of sees the same thing and, and, and how to make ed agents that operate in the world and also how to model selves as a result, those all being ideas that deeply tie together Totally.

Yeah, I, I like what you said about how predictive processing seems to show up in different philosophies or applications or tools. It just shows up in different places in the world, which is an interesting pattern. And we had this really great conversation yesterday for the audience, like why are we going nerding out so deep on chain perception? And belief systems are like, we're building the SDK just as a reminder for anybody that's new, we're building the SDK called Epistemic Me link in the show notes on our website. It's going to be like the global open source. How can we create models of cells that programmatically represent unique cells, us like a billion humans on the planet and growing, and also every other organism also on the planet. That would be next next level. Um, and we believe that this is key to solving the AI alignment problem because if we don't fully understand ourselves and have common interfaces to handle knowledge, perception, beliefs, et cetera, then how can AI understand us? That's really our fundamental thesis. One thing I'll also add to that, just after explaining like why are we nerding out so deep on this? Why we nerd out so deep on this and we don't just do it on the podcast, we do it in our team meetings, is because fundamentally it is required for us as domain experts and builders building this thing, we need to understand these concepts very deeply. Not even just at a theoretical level, but enough to practically implement it into this utility epistemic knee, the open source SDK. So that's why actually we're not just nerding out on philosophies. During the podcast, a lot of our design sessions about technical direction of the architecture intersect very heavily with philosophy and epistemology and thinking through predictive processing and free energy principle and all these things. Cool. I was going to ask a follow up about predictive processing. So now we have the, I don't know, let's say Bachelor's level, PhD level kind of view on predictive processing from Jonathan. If we were to explain this to like a 10 year old, how might we explain it to them? Who wants to take first shot at that?

I think maybe explain yesterday about this concept of umvault and there's a great book called An Immense World to take the perspective of a fish that is needing to live its life and eat food. And the fish has sensory organs and it can sense sense chemicals in the water and it can detect, oh, there's maybe some food over in this direction. And then it's going to actuate its muscular structure in its body to move closer to that food. And I think when we think about this, we think about locomotion as like moving from point A to point B. But all that's necessary for the fish to fulfill this homeostatic need of feeding itself is that it, it is changing its environment. So from its point of view, it may not even know that it's moving. All it knows is that it can actuate itself and that food is then closer. And so single cells sort of operate in the same way where they have no concept of the world as a whole or like locomotion. And Carl Frischen sort of pointed this dynamic out like that something as foundational as moving from point A to point B is really a human construct. And. But if you come back to what's the sensory environment of this organism, it has its own sort of reality. And this. So the, the word. There is a word for the sensory environment or the German word called umwald. And there's a great book called An Immense World that kind of goes through the animal kingdom and it talks about like different animals, sensory worlds and just how different they are from the world that we live in. And when you start to think, you know, and you can demonstrate this with a dog who, as you take a dog on a walk, its entire world is, is like its social life is sniffing what's around it, right? Like it is seeing a completely different reality than we are. And I think you could talk to a 10 year old about like, why is a dog acting like, what is the world like, can we take the perspective of this dog and see like, you know, why does it act that way and realize that like, you know, the subjectivity difference between different animals is very stark. And then you know that there are also, even though we have similar sensory organs as all humans, there are a lot of like, you know, sort of umvault differences that are very different, you know, from the way that we grew up in our cultural context. And even the way that we perceive things as simple as color can be different between individuals.

I love that example of the fish and the dog, especially for explaining it to a child. To basically be like, hey, this dog has their own world, right? And it's not like your world that. I think that is a very simple example. It's, dude, do you want to eat dog treats, kid? No, but look at this dog. Would love to go ham on dog treats. Like, why is there that difference? Like, cool. I know, I love that. That actually simplifies the concept quite a bit. Thank you, Jonathan Dean, did you have anything to add to that particular thing? I do have a next thing that you might.

Yeah, I feel like. I feel like I might start even younger. You know, the. You. You'll. You'll often find with babies that they don't have this idea of object permanence when they're born. It's something that they're taught. So you'll often see, like, parents playing with their children where they'll, like, hold a ball up in front of them and they'll put it behind their head and then they'll like. Or they'll like, move it out of the baby's field of view. And the baby, like, they'll be so surprised when the ball comes back into their. To their perception because they don't have that idea that the object exists outside of their immediate field of view. This idea of permanence when they're not consciously perceiving something is a belief system that is under. That is inherently taught to them as a result of them interacting with the world. Because it's a very useful belief system, right. If you didn't think that things continue to exist when you're not perceiving them, you wouldn't actually fare very well in the world, right?

You.

You'd probably, like, fall or, you know, get eaten by a tiger if you're thinking about the, like, the, like the Harare view of the world, right? So. So this idea of object permanence is something that's taught to children. And I think, like, it'd be interesting to see if it's something that could be untaught to a child, right? This idea that it's your underlying beliefs that is actually causing this thing, for you to think that this thing exists outside of your unconscious experience. And I think actually Jonathan brought up the color question. That is really. I remember having that conversation when I was a kid, even though, like, on the playground, I was like, is your version of green the same as my version of green? And we all know that's actually not true, right? Some people are colorblind. But, like, is what. Is everyone around the world just seeing different, you know, versions of the color green? And just asking that simple question is so profound because it forces you to think about what is conscious experience, and is conscious experience generating to generate it differently for you? And that next question might lead to, like, am I the only conscious existence that is there? Right?

That's like, is everybody else an npc?

Yeah, yeah, yeah, That's. Yeah, that's one Logical conclusion for sure.

Yeah, yeah, yeah. There's a few ways we can go about this. Like we could double click more into like why we're placing a bet on predictive processing and tying that to epistemic me and some of plans to figure out a benchmark. Or we can go into. I was thinking about epistemic me with Jonathan a few months ago and I. I always go to. In my entrepreneurship career anyway and launching products and stuff. I always go to Peter Thiel's book zero to one. I think no matter what you think of the guy, like he, he has good thoughts and he has these seven questions and Jonathan and I went through them and I was like. Because the question with any venture is like why now? Like how are we saving the world now? And so there's a bunch of questions he has. I could like we could go through some of those because I think it does frame really well. Like why are we going after this and why is it the right time? And why should people give a shit about a philosophical toolkit for belief systems? Like what do you guys think? What direction do you want to go here?

That sounds fun.

Sounds. Sounds good.

Yeah, yeah, yeah, yeah.

I think I actually I had a thought. This is pretty concrete to wrap up the, you know, individualize the subjectivity world. I'll just carry. I think it's a quick thing.

Please.

When Dean was mentioning reinforcement learning and I think a lot of us were. Would play with the OpenAI AI gym which had this reinforcement learning frozen lake example. And you're in. The basic idea is it's a two by two board and you're navigating this actor through a frozen lake. And then they need to avoid where the holes are so that they don't fall in the lake and die. And in that example, you're taking the data which is this 2D map and you're running over and over again to train the system to learn how to navigate the frozen lake. What the system doesn't model is just like we talked about the fish and the dog or the child on the playground, what is their individual point of view? And like so it. Now if that actor on the frozen lake has their own sensory experience, which all living things do, that's information we can use to solve this problem. And so that. That rolls into sort of like the why now. But if you start modeling that as a. Is a sensory agent that has sensory information that has belief systems and you combine that with the world model that the agent also has about the frozen lake, then it gives you Complimentary information to feed into the learning algorithm. And that's how you would augment the, the, the reinforcement learning approach through you know, this belief system.

I think that's a great example. Like in summary the way I think about that particular problem to solve it it's like what is the data available? Right. And there's multiple ways to look at what data is available and that is a treasure trove of important data is that agent's sensory information and tying that to their belief systems to create a function to route them the right way. That makes a lot of sense to me. Like I imagine that function. The more data you have about the models involved in this interaction, the more complete your holistic view of it and efficacy of the solution I would imagine. Is that kind of what you were where you're going with it Jonathan?

Yeah, and I think it rolled into this kind of like the teals why now question. There's information about the world we're not harnessing and it is from the subjective point of view, you know.

Yeah, so. So yeah let's go through it. Very first question is can we create. It's the engineering question, can we create breakthrough technology instead of incremental improvements? And actually as I say this I should read the other ones. I don't know actually know if we'll have time to go through all of them. So I'll read them to refresh in our memories of this and then we could pick the most interesting ones. So can we create breakthrough technology instead of incremental improvements? The timing question is now the right time to start your particular business or project? The monopoly question, are you starting with a big share of a small market? The people question, do you have the right team? Fuck yeah we do. I don't think we gotta go too far there. No. But I have something to say about that. The distribution question, do you have a way to not just create but deliver your product? The durability question, will your market position be defensible in 10, 20 years into the future from now? The secret question, us being on this pod together and all of our talks. I think this is an obvious one. You identified a unique opportunity that others don't see. I actually see it as our mission to convince other people that hey, this is the opportunity everybody to align.

I'm curious how do you like maybe how those questions. I think those questions also apply to an open source projects but I. Yeah absolutely, yeah I think, I think they do. Yes, I think absolutely, yeah I think.

It'S just any venture, right? Could be a venture in the business of trying to make more capital or resources for any investors or stakeholders. It could be like a project where like Wikipedia, like they're not trying to churn out money. Right. And I, I see a parallel with Wikipedia and what we're trying to do with making this open source. But yeah, I think, Jonathan, maybe you could talk a little bit about the why now? Like rolling in from that previous example. Is now the right time to start your particular project? Why do you think now is the time?

Yeah, I mean a big part of it is just the revolution with AI and you know, transformers and LLMs and being able to process information. I think it unlocks some of the tools that we need for the complexity of subjectivity. And so while I think that Transformers was specifically sort of in like an analytic, a progression of analytic philosophy and realized in technology, it opens up new ways that we can look at each other and relate to and create a communication channel between individuals that wasn't there before. I think that's one aspect. The other is just like the richness of, you know, like I remember there, there's like a kind of a flood of theories of everything around 2008, which was shortly after Wikipedia was available to everybody. And you see these kind of. There's fluctuate increases in patents in metropolitan areas. Every time there's a new communication technology like email. We created a wave and I think we're just, we're living in wave after wave of like where we have access to so much information at the tip of our fingertips like that we never did before. So I think these are the things that say we can answer some of these questions that were previously unanswerable.

Got it. So that goes into the why now? Is the tooling has evolved for us to be able to take advantage of it to solve some of these bigger problems around quantifying complex sedratgivi. That makes a lot of sense. Dean, did you want to add to why now? Why epistemic? Me now?

Yeah, I think there's two big things. I think that as the, as the war between the big AI labs grows and everyone's trying to create these models of agents, there's the question of like, what is an agent? What is a belief system? What is the way in which they interact with the world? And eventually we're going to need a set of interfaces that enable us to understand how to interact with these things, how to model them, how to share them between systems. Like, and I think we're going to eventually need an open protocol that Multiple people align on as they're building these new systems to constrain the variance in which they propose these new ideas. Right? Because that's going to enable us to all work towards a common goal. In a way, it's a benchmark, right? It's to say hey. Saying hey. If I bound the way in which the system is going to operate, if I bound the sense of what is a belief system, and I tried to come up with the best definition I can, that's possible, that is still adaptable to these different ways of building agents. I think you have a really powerful way of enabling information flow across the economy, across ourselves. And I think that's really maybe like a larger theme in this podcast is enabling information flow, right? Good communication is information flow. If information is flowing in an organization, people are communicating well, you know, and oftentimes belief systems get in the way of good communication. So, so coming back to this idea of interfaces, I think getting the interface right is often, it's so simple, but it's so important in enabling everyone else to start to get information to flow effectively. And I think that if we're building agents, we want them to all operate in the same world. And I think that's why epistemic me can be very powerful in getting the right set of interfaces or a protocol together. And then the second question, the second reason is our ability to just process information has grown dramatically as a result of like Even what like ChatGPT is like shipping, like I was demoing it with my family over the holidays and like I, I, it just blows my mind. The kind of information processing that is available to me. Like I'm able to just like take a video of my surroundings and say turn this into JSON. And you have this system that's now like decoding external information into JSON. It's like a live transcriptor. It can decode all these different types of information it perceives into structured data. And I think going back to what Jonathan was talking about with like the, like the old versions of reinforcement learning that you saw with OpenAI, Jim, they were just so simplified because your ability to actually model the world's information was fundamentally limited.

Right.

And you had to kind of impose some kind of discrete structure to organize even simple systems like operating in a frozen lake or the balancing of the pole self organizing system that can take in this deluge of information, structure it effectively create that state system and then, and then go out and act in the world through a set defined interface. I think it's a really powerful Idea, you know, and I think, I personally think it all starts with the interface and getting that interface, really.

Right. As you say interface. Because we were talking about APIs yesterday. I was just kind of thinking there's going to be a Robert API, a Dean API, Jonathan API. And we're basically talking to each other through our model itself. Actually a funny thing for the POD audience last night we had a pretty. A deep debate about the whole where should we go with our technical architecture? And I had like, I thought I heavily disagreed with something about. I forget what it was even. And then Jonathan was like, well, what about this? And then I thought about it more, I said some things and I was like, oh, wait, we're basically saying the same things with different words. And It's. We spent 40 minutes. We spent 40 minutes on it. And then Jonathan had this question that like triggered my brain the right way and it was like, oh wait, no, we're aligned. Cool. I mean, but that's the story of everybody's life, right? There's a bridge, there's a gap in shared understanding. You build the bridge. And Dean, what I'm hearing from you is why now is there's so much information at our fingertips. There's a need for better information flow in the system of the world to optimize for. You didn't say this part, but I think just optimizing for. I don't know what Jonathan said, like information exchange, getting a shared understanding and hopefully humanity is united. And then the other part you said is an interface is going to be really needed for the type of data that we're talking about.

Right. And I think, and I think the most simple interface possible. Right, because you imagine you could build an agent interface that has a separate method for dealing with objects or dealing with people, dealing with environments. That's just so much information to try and get your. Get a hold of. But if we can come up with the most simple interface possible, something that's really beautiful, really elegant, that describes what it means to model a self and how it interacts in the environment, I think we're going to enable everyone else to start to build their, their versions of what itself is and help them sense of how that would work in the world.

Yeah, we'll be like the de facto open source personalization layer around the AI alignment problem. I think that's gonna be cool. It'll be cool to listen back to this podcast in 10 years or something and see us revisiting our beliefs around this and we'll probably see those evolve. Yeah, I'll answer the question myself. Why now? I think I so plus one to everything my co founders Dean and Jonathan have said around the tooling is there, the need is there. From information exchange point of view. One thing I'll add is just like the AI alignment problem, I can't emphasize enough like how I from the viewpoint of let's start with knowledge and let's start with perception and let's start with belief systems, let's start with epistemology, like the building blocks of how we perceive the world and stuff we're approaching. The AI alignment problem is just very different than what I've seen out there. And I would love just audience like if you're working on AI alignment, please come hit us up. We'd love to talk to you. We want to be very open about this and I think that's how humanity wins with AI in the future. I just think there's a pressing needle. There's going to be a lot of unknown unknowns as we broach into this next century, the century after that. And I think the time is right because there's a lot of existential risk, right? And I think it's waking up more people that oh wait, this thing's going to take my job like in 30 days. There's like this fear, right? And there's fears of oh, are robot's going to kill us? Are we going Terminator? Like there's a lot of uncertainty, a lot of unknowns, a lot of fear. And I think the only certainty is we all got to work together to make sure that we survive. Brian Johnson's don't die movement. We all don't want to die. That's the only common game that we all are generally playing together and have the same objective. And so I have now come to the conclusion like this is absolutely needed for us to survive in the future. Like that's where my head's at and that's why I have such a zeal for it. I'm like, oh, this is so purposeful. It really feels like we not only need this, but if we don't do it, then the risk is really great. So that's like the, you know, kind of corny humanity, existential threat sort of reason for why now? But that, that, that would be my addition to the conversation. So, next question, I suppose this one's an interesting one. There's two others that I'm thinking could be very interesting. We actually also answered the first question which is can we create breakthrough technology instead of incremental improvements yes. If we do this right, we're like the Nvidia for every application that needs any personalization in the future. Right. That's not, oh, we just made an Airbnb clone in a particular part of the world. That's not what we're doing. We're bringing a game changing perspective to this. So I think we've already answered that. Next one I'll ask is, will your market position be defensible 10 and 20 years in the future? I have a lot of thoughts about this. I'll just like go first. My thinking is the only way. So there is a lot of knowledge out there, domain expertise, I'll call it, around being human. Psychology, neuroscience, metaphysics. There's all these different domains where you literally have PhDs in sub domains with these domains, right? Like we just scratch the surface of predictive processing. But I'm sure There's like many PhDs around subdomains within predictive processing trying to move that field forward or the free energy principle or anything. And so what I've come to realize is that for this to be true, view the next best question that will help you evolve your belief system in the best way. That which is the conclusion we've come to. That's like the product constraint we want to get to. Optimizing for that requires deep collaboration between all these different disciplines and domain expertise that can go super deep and it's super broad. And then there's the, you know, other side, which is technical implementation. Like, well, how do you translate all those, you know, philosophy, epistemology, like all these different things, neuroscience, metaphysics, everything into okay, it's an SDK now. Cool. Like great. Oh, there was like, you know, graph theory, information theory, there's all these complicated topics and then translating that to a technical implementation. So our to answer the question, will our market position be defensible 10 to 20 years from now? If we succeed in building a global community around our open source project, then I believe we can be the Wikipedia of human belief systems and, or I guess even not even just the human belief systems, but certainly that, but beyond that, any model of self, any belief system. So the reason for that is because I just don't believe in a world where Facebook could buy all the PhDs in the world. Right? Nobody's competing with Wikipedia, nobody. Like that isn't. You can't compete, compete with that because it is a contribution across the globe kind of thing. And no central entity could provide enough capital to go make a Wikipedia competitor that wins. So that is our approach is you.

Could argue that maybe like Chat GBT just obliterated Wikipedia's value prop, but.

Actually we should talk about that a little bit more. Oh, I didn't even. Actually I didn't think about that. So that's, that's a great not question but something that is changing my belief systems in real time. Right. But I suppose either way. So Wikipedia aside, if we get global collaboration, because all of building an asset is really just getting people with the right skills with the right time to spend building the asset. And that's what any software is. So that's why big tech companies could pay more to keep building their assets. Um, but if this becomes like we get some of the smartest people in the world and that grows to a hundred percent of the smartest people in the world or a majority are contributing to this project. Like I think the intellectual property and amount of domain expertise that goes into it is. That's just, it's going to be hard to replicate. That's what I think. So that's how I think capabilities of epistemic knee like market position wise. That's why I think 10 to 20 years from now, the way that we're shaping this not as a private project, but as an open source project that will lay the foundation for scaled global collaboration which hopefully I think builds a moat against any, anybody else. Which with this kind of thing, I don't think you want fragmentation, right? What if we did have fragmentation? There's 30 different ways to do this in the world that just not be good. Wouldn't that just be replicating what we have today? Everybody's just off in their own world, think of their own things. We have divide across politics or even, or sexuality or whatever. Ideally we don't have 30, 50, hundreds, thousands of different models. We have like one common model for humanity. That's fun. Happy to hear other opinions though. I just talked a lot about us being Wikipedia Dean. Great challenge on like oh shit, did ChatGPT just obliterate Wikipedia? Maybe, I don't know. Gotta think about that more. But I do believe we'll reach a moat where other people can't build something better if we get global collaboration on this one thing together. That's my thinking. Does anybody wanna add to that question around, you know, how will our position be defensible in 10, 20 years?

I can break down how I think about it. It is and I think part of me says like that we don't. There's a certain aspect we don't want defensibility in is in that we want to just put up these ideas of subjectivity and create and just you know, like kind of tools that like, you know, are universally available. So I think ideally this comes across where like each individual owns their own belief system and because like why wouldn't you? And I think so that starts to sort of like turn the tables a little bit in terms of how we interact, interact with our tech oligarchies today. If you, if you can construct your own belief system, it's useful and it can help to ask the right next question in chat dialogues with LLMs and can predict what you're going to say so that it can represent your point of view on your behalf. That's interesting individual agent perspective. Now I think then you get into aspects of matchmaking and so how do we like use that belief system and model itself to people then I think there's a lot of diversity of ways that you could do that, you know, and, and this can come into place in say marketplaces. If I need to matchmaking supply and demand with one another, how am I modeling the belief systems of all the actors in the marketplace and then how can I make the market more efficient? And I think if you can show that you can create an emergent organization like this where you could solve a real problem like say increasing marketplace efficiency, then that creates a way that everybody looks at and says oh shit, that's the new bar to break. Like that's something we want to accomplish. And I think that extends to like next generation organizations too. If you can have an organization that can just like out compete other organizations because the belief systems allow for belief coherence and dynamics of execution that are just not achievable today through crime organizational structures. And that's like the proof that creates like a mode of sorts. But I think what you really get to is like just like ourselves have agency and they can, they have their own environment and their own predictive processing kind of reality. And we have, we are collective intelligence of our all of our embryonic process today. Like when, when you sort of matchmake people the right way and everybody is functioning in a way that creates sort of an emergent organism. So let's think of a culture now as being something that actually has its own sort of like sense making apparatus, its own way of modeling existential risks. And then I think you start to think about like how can representation system create nation states that are like more capable at navigating problems of existential risk and taking care of the people within that organ, within that, within the Nation better than our countries today do. And so when you get to this kind of emergent sort of the culture self, you're building a self out of the collective, right? And I think that's the thing where that's just completely different than anything that we imagined today. And that would be defensible in some sense, but hopefully replicatable in that those tools would be stolen and used and repurposed for greater good.

Yeah, yeah. One thing I heard from what you said is, and I think it underlies a belief that I think you have, which is if we do this right, then it should make people more money and just help every organization work together better increase efficiency. So there's that obvious values, economic value standpoint. Like I imagine if we can make our current markets just pick a big market, like financial markets, like if we can make that market like 1% more efficient, how many dips in the GDP like is like that? That'd be crazy. So yeah, I see what you're saying. Like if we are right in our hypothesis that this is the right way to solve this problem, quantifying subjectivity, doing better market making, matching supply demand more effectively, more efficient markets and then the other side, which is more effective organizations of people, then the value prop is really clear from those lenses. Dean, did you want to add to the moat question?

Yeah, I think plus one on the idea of getting ideas out there and having those ideas, if they're good, improve how belief systems agents models will self work as a whole. I think a great analogy here is Linux. Linux was also an open source project, but because of its underlying ideas which it promoted, right? Like keeping the core Unix philosophy alive, simple interfaces that ended up putting a lot of pressure over time on other competing operating systems to adopt similar ideas. You know, and today, you know, a lot of Windows, you know, takes or originally didn't, but today it actually embodies a lot of Unix philosophy because it had to compete with Linux on the enterprise scale. So in a way I think that there's an opportunity here to say how do we take the ideas that we maybe as the three of us or as the community grows, really care about. So like some ideas that I care about are how do we create models of self that are not abused and that ultimately for who the individual IT models can be used safely and in a well permissioned way. That's kind of an interface problem, right? And if you design the interface in a way such that you enable that auditability that does data controls, you provide this alternative in the marketplace that is going to put pressure on everyone else to adopt the same ideas in their interfaces as well. Some of these other ideas are really powerful as well. The idea that selves, the same data structure or interface used to represent you as a self can now model this collective self. And this idea of how do cells talk to each other? How do cells that are like organizational or government entities talk to their individual selves? What does that interface look like between the two cells? If you get that really right, you get that really simple by OpenAI or Google also start to adopt these interface and maybe they'll even choose to work with us on a shared protocol. And I think that's a real, that's a way to like evangelize your ideas by getting really good implementations out there that other people will want to copy. And I think that's the real value in open source projects.

Oftentimes that makes a lot of sense. As you say that it's. I can't help but think about our conversation about philosophy and the evolution of philosophy just being other thinkers building up somebody else's thoughts that they had centuries ago. Um, and I love the, I love the example with Linux. I think, I think all these things kind of run in parallel like our. The thinking is we ship something if it's valuable enough and people think it's value like the market. The world thinks it's valuable like Linux and eventually over time it'll be used more which is great measure of success and value delivered to humanity. And also like other players might adopt it or sponsor it or work with us as the value of a common data structure, simple interfaces, et cetera. Like these principles. I think yeah, that would transcend. Yeah, you know as I say that like it kind of goes back to our architectural principles I suppose like that we've been kind of forming in conversations. So I'll give a quick plug for the community that cares about contributing to this. Whether you want to build and like contribute code, get on calls with us to talk through some of these concepts and how we might integrate. Like minded that want to build this thing and I think um, I'd love to. I actually never deep dived into the history of Linux. Was there like an original OG community that like really drove the principles that are now still reflective today in the realized like value prop of Linux?

I think if you want to understand. Yeah, I mean there's how much of Linux when you talk about core principles is the Unix philosophy of Linux and to really understand Unix you have to go back to like Bell Labs and I think it's like Kenver and like all those guys who created the initial operating system and then. But Linux is definitely very opinionated. It's very opinionated in town in terms of kind of. Well, directly led by Linus Torvalds. And he, he provides a lot of really strong guidance on how to add things like adapters to the system. I think I'm using the wrong term here but like yeah, how you write an adapter for an operating system to work with, let's say a mouse, like a driver. Like, like there's, there's a lot of opinions over how you add drivers to the system, how you get the operating system to work with inputs and outputs. You know, in many ways that's kind of like a, like a self model or an agent that needs to do actions like how do you write these drivers or these, these plugins to get to work with the system, the world? And how do you get it to.

Right.

How do you come up with the right internal interfaces for beliefs? And, and how do you get it to plug into epistemologies the right way? In many ways it's, it's, it takes on that similar philosophy of like the right internal interfaces and the right internal composition. I'm not super well articulated on how Linux runs as an organization or what you would versus is, but it's, it's definitely there.

Yeah, yeah, that'll be an action I take is just like see how they do it because obviously they built something robust in terms of collaboration and a product over time. So. Cool. Cool. Okay, so question from the Zero to One book, Peter Thiel is have you identified a unique opportunity that others don't see? I feel like we have covered that a good amount. So I just want to check in. Are there stones left unturned that y' all want to bring up around that problem or that question? Unique opportunity that others don't see?

I think this idea of what the self is and how. And you know, I think Jonathan alludes to all these ideas that, you know, the self is this, it's not just you and me, it's this like hierarchical, like, like emergent phenomenon to describe these systems that process information in a certain way. So you could describe cells as cells. You could describe you yourself as a self. You can describe collective governments that are organizations of selves also as the self. That idea is actually really, really powerful. And if it can show up in a meaningful way in the interface, I think that's a unique perspective, especially if it's informed by philosophy.

Yeah, yeah, definitely a unique Perspective. Jonathan, do you want to add anything to that?

Yeah, I think we talked a bit about like, you know, how to, how can we bring more useful data into these AI systems? And I think like, you know the Schopenhauer quote, genius sees the target that nobody else can see. And I think when you start to decompose the self and model the self, you just start looking at the world in ways that create new data sets that you might not have conceptualized earlier. I think, you know, one way you can think about that is just like, how do we kind of break down the self in terms of all the hierarchies and say like one, one experiment that I, that, that I did was to say what are all the frequency domain responses that the nervous system takes to interact with the world? And there's a really interesting picture that paints and it's not something I had been introduced to like in psychology models that have collected data or are seen in papers, but you know, like this idea that it takes about 700 milliseconds for us to like cognitively think. In communication, our emotional responses can be at 300 to 500 milliseconds. Our reflexes in our nervous system can be as short as 10 milliseconds. And there's this really interesting thing that happens, your nervous system if you hear a loud gunshot next to you. If I'm in the middle of a conversation, it's going to shortcut this cognitive sense making loop that I have and it's going to go right to this fight or flight response which is going to happen at maybe 80 to 100 milliseconds. And it's invoke my spiral reflexes and my body's gonna turn towards the sound. And you know, there's data sets and all of that information. So if you break the human response down to these frequency domains, it's one just way of conceptualizing self and collecting data around all these interactions. And that's just one example. I think you can also say we know that all living animals have ancestors, right? What does it mean that we know who my parents are? What does it mean that I know who my ancestors are? What does it mean that I know that I evolved from this particular tree of life and that had to endure these certain environmental conditions and be and learn and pass this on through natural selection. That's all information that hardly any science uses right now. And it is fed into models of self and it's all out there. I have other examples, but I think there's just a couple that kind of Give you an idea of like, how we can like new data by just deconstructing the self in new ways.

Yeah, I wanted to complete the quote because it's such a great fucking quote. Talent hits a target no one else can hit. Genius hits a target no one else can see. I love that quote.

Schopenhauer.

Yeah, Schopenhauer, I think. So, leveling back up to the secret question, I think we definitely are looking at this with a very different perspective. And like I said, it takes a lot of understanding all these different multidisciplinary domains and fields of expertise at a good enough level to actually get to product design, technical implementation of the SDK and APIs. I definitely love that quote. That might embody our team, actually. We'll see. We're creating the future as our culture evolves. Dean, did you want to add to secret question?

Oh, what was the prompt? Secret question.

Oh yeah, it is the secret question. No, uh, have you identified a unique opportunity that others don't see?

Uh, for epistemic knee or for myself. Okay, yeah, I think, I think I answered that question, which is like, with the models itself and particularly like, like how we construct the interfaces. I think, I think the interface is the lens on the world. So I think it's, it's the novelty of the interface to me, which I think is. That is something that can really shine here.

Yeah, so, so just for the audience to track and make sure I'm tracking too. You're saying modeling. If I remember correctly from your previous response, it was like, if we can create a common data model for every agent, like cells to like humans to groups of humans, or nations of humans or the world of humans, then simple interfaces between those things to maximize information flow in a good way is going to be like the elegant, unique opportunity and perspective like that we have an opportunity to, to implement. Is that right?

Yeah, exactly. And I think, I think that idea of like the right data models that maximize information flow is a really cool idea because there's so many, like, there's like also like a lot of like low hanging fruit. Right? Like for example, how science is done today, you know, ever. Like, like there, there is a reproducibility crisis in science, you know, and part of that is because there's no like standard format for actually publishing how we do science. Right. Like people, like the primary way is to like put up like a PDF and archive, but that, that's not like a schema that I can actually interact with and plug into some other system and examine the data in a meaningful way and potentially hand off to another lab that's running somewhere that can actually reproduce that experiment. In order for us to get to that end world where every, every information that's being produced is reproducible and consumable. That requires having good data models of ontologies or good data models of what is science. If we consider science to be an epistemology, a good way of like publishing that as schemas that maximize information flow in the world. And I think it's getting. Yeah. If we get the. Yeah. The set of data models that can. I think it really all boils down to, in my opinion, epistemology and ontology. To get those two ideas, you model them correctly, I think we can. I think we can really maximize information flow in the world.

And because of our. If we get to that point where we are maximizing information flow throughout the world and I think tying that to what Jonathan said, there are so many more things that we can't even imagine can happen. Right. We build that as a platform. It helps us progress science more. And I think about things in terms of the macro and the micro. So at a macro level, I think this is, we save humanity, AI alignment. Cool. We solve that and maybe so many other things. Like maybe we could accelerate research and science in different domains like health or longevity. Or on the micro level, it's. I wonder if I could just like build a little therapist coach or whatever relationship coach on top of epistemic me that has had an efficacy where it could be like, here, let me talk to you and your partner and figure out the next best question to ask to bridge gaps in misunderstandings. Agro level, I think that type of thing would be cool. And then the thing that Jonathan is building is a prototype quick plug where Jonathan is building a reference app. Hopefully we'll have it available for demo. If you want to sign up for the beta, hit us up like, let us know. Definitely. Happy to get users on board. But a tool that helps you just like a chat interface like Claude or ChatGPT, to get a bunch of information from you to extract your belief systems and mirror your belief systems back to you so that you could start picking apart. Oh, okay, I'm thinking that way. And then we'll ask you like the next best question and evolve that belief system in front of your eyes. So like, we want to create a simple chat interface to kind of help you do that. And I, I kind of think of that as journaling. Almost like instead of me Googling journaling prompts, it's can this thing just predict what the next best thing is for me to think about so I could self actualize. Cool. So we're building that. If you're interested, let us know. So we Talked about the 0 to 1 like questions like why now? Why epistemic mean now? And hopefully people have a better understanding of the founders, our perspectives on why this is important now, not just to save the world, but do so many other things like enable so many other innovations can't even see today. And we think we have the right team, we think we've identified a good opportunity, we think that we can make breakthrough technological advancements and we think we can have this be defensible or it's so high value that leverage into the future. One thing I want to bring up to wrap the podcast is maybe a deep dive into thing we were talking about yesterday which ended up being like this should be the grounding product requirement. It is the idea that our benchmark should be what is the next absolute best question that you can get that we predict will help you evolve your model of self in the most optimal way. I don't know if I articulated that exactly right, but at a high level we want to be able to predict the next best question using our SDK for your model of self. And that could look like the relationship coach example. That could look like, I'm an AI health coach. I ask you questions about your sleep, diet, exercise. Maybe it comes down to you think that you only need three hours of sleep a night. The AI health coach, based on all the research it knows is like, maybe I should ask you about that. That. So that's what we mean by next best question to provide value to the user. So there's some things to consider about this, like what makes a question the best and then how do we ensure that those. That's where my mind is going. Do one of you want to add your thoughts on. On that? We were kind of talking about this last night, so we're cheating a little bit. But yeah, that.

I mean that is a very philosophical question. I think, you know, what is the question? What is the best question? I think there's certain ways in which we can ground the problem. I think both, both you, you are. I mentioned in the past this paper from Google that had done this experiment.

Of like how can it train a.

Model of self and then look to see does it actually answer the questions maybe in the way in which a user might answer a question that's predictive of creating a good model of self. Can you reverse Engineer something like that and create a benchmark for producing good questions perhaps. I think there's also a more philosophical view, one that kind of reinforcement learning has tried to answer in terms of like, how do you produce maximum long term rewards? Right. There's a lot of questions you can ask that can lead to like short term minimums or like local maximums or like I can make you feel good in the moment, right? Like for example, if I'm your therapist, I could juice you up and tell you that, you know, you're the greatest person in the world and you're going to conquer everything, but if I don't actually debug the belief system that is, you know, really getting in your way, it's not going to produce long term value for you, right? So the next best question, it may often intuitively be the question that gets rid of a limiting belief which requires understanding what your beliefs are in the first place, right? So it must first understand your beliefs and it must understand the algorithm towards improving your beliefs. So in that way I think you can view the next best question. There's a couple of components there. Long term rewards requiring an understanding of beliefs, then understanding what do we want to do with your belief system? And I think that that idea of what do we want to do with your belief system in a way is somewhat subjective. Which is why I think epistemic me might need to create room for having these pluggable subjectivities we had called them dialectical strategies where you could take a philosophy or a religion, encode that into a belief system and have that be the guiding, like a guide within the epistemology when operating in a belief system. So yeah, that sense of like what is best for a belief system I think ultimately will end up needing to be subjective in the short term. But as we operate at scale, we may start to create some of our own hypotheses over what is objective goodness when it comes to improving belief systems. And then that's where future epistemologies might evolve.

I love, I love the quote. I feel like this podcast even in like episode three is producing so many quotable things that I could put on a T shirt that would be dank. For example, we were sense debugging the belief system like holy shit, who is. How has it ever been said in the history of humanity we haven't programmatically created belief system so now we could say these things. Yeah, I love that I would put that on shirt that that'd be sick.

That's basically what therapists and yeah, And I think in May where that idea came from. I think it comes from the time spent in meditation and around meditation communities. You could argue that, yeah, there's certain meditation techniques that focus on link concentration or transcendental states. But many people would argue it's just debugging your own belief systems.

Many people that have done programming and attend meditation sessions. Yeah.

Which often overlap quite a bit.

I love it. And then so I got one quote from you and now I have the other one from Jonathan. I've been imagining this. Are you an npc? Sure. In my head. We're going to have to create these at some point. We're making memes already. This is great. Do you want to add to the whole concept of how do we determine best? Next question. Philosophically and then also technically some of your thoughts about that intersecting with epistemic knee?

Yeah. So I guess generally I think that the next best question is what's going to optimize learning in the system? What's going to make the belief system more representative of the user's actual beliefs and more predictive of the things that they want to do in the world? I think I've been thinking about this concretely. If we have a trained belief system that's a AI health coach and it, it, you wake up in the morning and it gives you a plan for the day that's consistent with your beliefs around diet, sleep and exercise. Now you can have really nuanced beliefs around diet, sleep and exercise. For instance, if, if, if you're extracting my belief system, you might get to the point where a great day for me is to optimize metabolic health and over a long term optimization. You know, I want like healthy mitochondria and I want great energy through the day and I have this kind of religion of my cells that I want my cells to be metabolically optimal for all the different types of energy sources they could process this kind of thing. Right. So that creates sort of an axiology about like what is a good outcome for a trained learning system. Now we back that up. And now, now like let's say the system is just getting to know me, like what questions should it ask? Right. And I think it's going to have, you know, some topics to explore like diet, sleep and exercise and just have some notion of like, have I completely gathered the, some ambiguous belief statements from this user that we can then start to evidence, Right. And then I think there's minimized returns at some point where you've asked enough questions, you've got enough like baseline sort of belief statements where we need to start evidencing these beliefs. And I'm going to start asking questions to evidence beliefs. And then I need to know which question to ask of, like, which belief to evidence is going to optimize learning next.

Right.

And so on until you get to a place where the belief system is predictive. And that's where like, you get to this point where we really want to be able to predict what's the answer that the user would give to the question and that then that defines like, what optimal learning is. So each of these stages, you have like this epistemology to navigate with its own sort of like manifold structure that creates, say, there's gaps in knowledge right here that we need to fill in. And one construct I was mentioning yesterday and I think is just useful is this idea of epistemic emotions. Epistemic emotions are things like surprise, curiosity, confusion. And so when the system asks a question, then guesses how the answer will be answered by the user is actually answered, then we have a discrepancy compare. That discrepancy can inform an epistemic emotion, which then can feed back into the aggregate sort of epistemological structure that then you can compute a function of what is the optimal next question or set of questions that we could ask right now. You know, in the sense as a developer SDK, like, we want to dimension these different sort of like dialectical strategies. And so depending on what the strategy is at a given point, we want to return like a potential array of next questions that the developer could like, decide, this is how I want to guide the user experience at this point and how I want to interact with the user and learn about them next, that strategy. So there's a lot of like, leverage to pull here, a lot of moving pieces. But I think that benchmark that Dean mentioned, I think it's actually a Stanford paper that they created, like they created agents of psychologists and sociologists to analyze people's chat logs and then predict the next. And I think that creates a benchmark for us to say, can we predict answers better? And then can we even shorten the amount of time it takes to collect that information to ask better questions and get better predictions? And I, I think that would be like a great concrete goal that we could benchmark and baseline and try to improve over time?

Yeah, absolutely. A lot to unpack there. I think it was actually Stanford and Google together. Okay, yeah, I'll link it in the show notes.

But it's called generative simulations of A thousand people. Yeah, a link for it.

So that gives us a target. So for everybody listening, if you're interested in that, we'll link that research in the show notes. We'll keep you abreast of our updates towards marketing, towards benchmarking against that with our own toolset. So there's another thing I wanted to mention. Oh, why does next best question matter to me? This, this, this story? Jonathan and Dan have already heard it, but I think for the audience it might be valuable. So four years ago or so, I was on a road trip to. I was in Vegas and I was planning to go all the way to Colorado. This was at the start of the, or halfway through the pandemic where my employer was telling me, hey, we're not going to require anybody come back to office for a year. And I was like, all right, hey, I'm just going to go rock climbing on the road for six months or whatever. That sounds great. Started renting out my condo in the bay and then just started like gone on the road with my dog Nibbler at the time. And then two weeks out of the six month road trip, I got into a mountain biking accident in the middle of the desert and we nearly died. I was bleeding out of my arm. I had an open compound fracture you could actually like see here. And I had a shattered kidney and nobody was around him other than my best friend who was like five minutes behind me. And there was no signal. That was the other bad part. We were freaking out. I had a lot of adrenaline, fast forward. Anyway, waved some riders across the ravine. They came over. They were locals. Luckily they pointed my friend to go hike up this hill. Found signal called the helicopter. I got shipped to the hospital. Forty hours later, I had open compound fractured elbow. Got surgery for that. Have some metal in there. And then I was wearing a pee bag for a while. Like they actually like fixed my kidney and then I was wearing a pee bag. So I. That was hard. I was very crippled and I couldn't lay down because I had a lot of internal bleeding. So it was very uncomfortable. I could only sleep like 12 hours a night. And so the reason why this is important is I was facing an uphill battle to get to my regular baseline health. And when I got the cast off, my orthopedic surgeon said, this is one of the worst breaks I've ever seen. You broke every bone on the elbow in many different ways. You're not going to have full mobility back in your arm ever. You're going to have tendonitis the rest of your life, you will not be able to climb for over a year. And I said, fuck that, hell no. And so I said, next time I fucking see you, I'm going to have full ass mobility in my arm and I'm going to be doing pull ups. And so little did he know, he didn't ask me a question, but he told me something I couldn't do. He told me that I had limits. And little did he know, he said the possibly the best thing for me as a person that's like, I see myself as an underdog in a lot of ways and I don't like being told by what I can or cannot do. So he told me actually specifically the right thing to get me fired up. They actually gave me a PT program. I looked at it and I was like, this is like 20 minutes a day, 30 minutes a day. Like, I set a goal to be doing one arm pull ups in 16 weeks from accident with the same arm that was broken. That was like my goal. I was like, this PT program is not going to do shit for me, dude. So what did I do? Based on what he said to me and my fire to prove him wrong, I like researched all the different PT protocols for your arm, whatever. And I built my own program. And it was eight to nine hours a day. Religiously, I was doing pt. And then four weeks I went back to his office, I was doing 20 pull ups even though I had a broken arm. I was starting to climb again. And also I decided to recover from the road. And I came back to him and he saw me and at this point I was like, still broken. My, my kidney's still broken, elbow's still broken. But I was more fit than the average human for sure because I was doing 20 pull ups. I did have a good level of baseline strength coming into it. But he saw me and was like, how did you do this? And he was almost like marveling his own work that this could happen, that somebody could heal in this way. He's dude, like, I fixed your arm. Like, and I'm like, yeah. And I showed him my PT program and stuff. And he was like, what? And so at that moment, I think he realized that I was not a normal person. I was willing to do eight hours PT. And so when he recommended the 30 minutes or whatever, when we go back to that health practitioner example, how can a health practitioner give a patient a plan that they adhere to maximize their health outcomes? How could. My doctor lucked into saying the right thing for me, but that probably Wouldn't be the right thing for other people. There's 8 billion other people and I'm sure many of them would have taken that the other way. Right. But for me it worked. And so I always wonder like, how can we optimize exactly what a doctor could say to their patient to maximize health outcomes or get a fire lit in them to just get after it with their health. So back to why is it important for this tool from a product perspective to produce the next best question for a given use case? I draw a lot of my experience of what has been said to me and like how it's affected me and how that's affected my actions from that experience. And so that's become like, yeah, this should be the product requirement. This should be like the absolute number one value prop thing that we can guarantee. We're giving you the next best question based on how we've designed the system. So anyway, that's a quick story on near death accident that led to somebody saying something to me that got me fired up. Probably wouldn't have gotten other people fired up. And then I think 12, 12 weeks later I hit my goal. I set a goal. I was like 16 weeks, four months and I'll be doing one arm pull ups on this arm. I actually hit it in 12 weeks. It was still broken too. It hurt, but I could do it.

I was just going to say, Robert, I don't think you can make this open source project work.

Well, fuck your shit. We'll see you in three weeks, bro. Oh man. Yeah, I think somebody did say that to me a few months ago or something. They were like, ah, this, this is gonna be hard. And I was like, all right, well we're gonna do it. So cool. Now you guys know how to motivate me. Just tell me I can't do it and I'll get after it. Dean, you were, you were going to say something. I think you both like, oh, yeah, I was just.

No, I was just commenting on your story. I think it's like, I think it's a great example. It's definitely a very like, it's an example that sticks in your head. But it basically illustrates the idea that not every patient is the same and depending on how the doctor interacts with you, it produces different outcomes. Like that occurs not necessarily at that level of. I would say yours is a fairly extreme case. Right. You know, and, and, and there may be a bunch of different factors. Maybe you heal better and maybe not everyone else heals better. But there's, there's There is no one size fits all prescription that I think oftentimes public health policy tries to take. Because, you know, at the end of the day, our medical system tries to minimize harm. You know, I, I don't know, I'm not trying to say it's is necessarily malignant inherently, but in trying to minimize harm, oftentimes it tries to create solutions that are best tailored for populations. But, and because it's so difficult to actually create the resources to do science or pretreat patients at the individual level. But I think we're, I think new technologies are emerging, perhaps like epistemic me that can actually model the user successfully and then take some kind of public health epistemology or some, some set of public health knowledge and we can model how those two things would interact so that we can produce the right health outcome for the user. And maybe we bound it so maybe we can all agree that insulting people might work for very few individuals.

Right?

I think that that, that makes perfect sense and it's something we can automate per se. It doesn't require PT having to be the kind of person to ask those kind of questions. We already know those and they know that information.

And, and, or we could give this tool to the PT or, and, or doctor to optimize their plan. Like I was just kind of imagining what would my user journey be in the future a hundred years from now if I went through the same thing and EM was successful or epistemic me was successful. It'd be like I'm helicoptered in. I already have this nifty model of myself on my phone fingerprint scan connect to the hospital interfaces and systems. They knew who I am now and whatever treatments they do, I could be knocked out cold. Cool. Here's my model of what I want for care and who I am. Maybe the doctor reviews that in like a nice report and then maybe that report says, you know, Robert likes being an underdog or he comes from this kind of background and he gets really fired up if you tell him he can't do something. So perhaps tell him he cannot recover to do his rock climbing trip and he will certainly find a way to do it. That would be so cool if we can get to that future. And that goes back to the whole like this is a protocol, right? The interfaces really matter. The way we design our data structure and model really matters. So all these things tie together even though they seem disparate, like philosophy, epistemology, AI technology, they all tie together really nicely in this project. So I love talking about this. I love working on it. I know we've been going nearly two hours now, so I will call it there. Hey there. Rerecording this bit because this part got cut out, but I wanted to thank you. Thank you so much for like listening or viewing our podcast on whatever channel you're viewing it on. Please check out our website, share it with other people. If you like it, like it, subscribe it, drop a comment, tell us what you like about it, what you don't like about it, any suggestions for future topics or content, things of interest. We'd love to hear from you. I read every comment and every email. Hope you have a great day. And thanks for watching or listening to the ABCs for building the Future podcast. Have a good one. Bye. Bye.