B (2:57)

How about the plane? You take a plane from SF to New York and someone says, you know like 20% of the time it might just like the next word not going to match and it's going to go down. So how would you feel about it?

A (3:08)

Yeah, my, my, my feeling about that is planes are currently run very well by deterministic systems. So, so I don't know why I would need an AI for that.

B (3:18)

I feel like we just cannot avoid AI anywhere. Like next 10 years people are going to try to place AI everywhere, automate systems with AI and you know, technically you might not need. We survived somewhat without AI up to this moment, but now it's just like a next step of evolution that people just want AI everywhere. Like for the banking, you don't need AI initially, but we learn it's really helpful to automate certain processes and decision making and it's going to save us a lot of time and allow us space to be creative instead of debugging and fixing things. So I just feel like it's unavoidable future.

A (4:03)

I think maybe what I'm getting at is. What am I getting at? It seems like if you want a guarantee of certainty using the only way to sort of guarantee certainty is to use something that you can express in code or logic

B (4:27)

that's a part of it. So the certainty comes from internal verifiers and external verifiers, internal, at least for us. So for example, if you take LLM, obviously it's a language based model and architecture doesn't allow you to do internal verifiers. So it's like a black box for you. You don't have access to what's inside until it's all processed, but you have access to the output. And many people and companies sort of take LLM are trained for certain tasks and if it requires logic they attach texture and verifiers to it, such as languages like Lean 4, which is a proof, I mean machine verifiable language proof language which allows you to check this output using mathematical frameworks. However, it doesn't solve the problem of things being just so expensive because what expensive is your architecture, which is still playing a guessing game out here. And, and even if you attach external verifier, even you fine tune this LLM specifically for the task you're trying to create. You're still not solving the problems of tokens being expensive. It takes compute for you to play a guessing game. So this problem is solved by the EBMs, but we're talking about LLMs for now. So here we have the situation when there's right internal absence of verifier, but there's external one. So now about the EBMs, EBMs don't have tokens, it's token free model. There's no guessing game of this kind. So essentially you could oversee all the possible scenarios.

A (6:16)

Can you define EBM for us?

B (6:18)

Yeah, I'll define in a second. So for now just think of it as something which doesn't play a guessing game and something which has architecture which is essentially allow you to self align itself as a processing the information and it's no longer a black box for you. So as it's performing you can open it anytime during the training and you could see what's happening in there. So you cannot do this with LLMs. Just nature of architecture is different. So you have for verification tasks you have this notion of self alignment because of the EBM architecture and the absence of token makes it cheap. But also you have external verifier on top of it. So you have verification sort of on both sides, inside and outside. Hopefully that makes sense.

A (7:08)

I think. So let me play it back to you and you tell me if I'm getting you. So basically I think what you're saying is we're living in this world which is really cool with LLMs, which is we can generate lots of output with them and the output is really useful for a lot of different things. But in order to tell if the output is right, the best we can do is sort of guess and check. We generate the output and then for example, if it's code, then we go and check the code with integration tests or manual tests or whatever just to see if it works. And that totally works. But it is expensive and time consuming and one of the problems is it's very hard for us to know, okay, how did the LLM get to this answer? We can't go look inside of it.

B (7:54)

Exactly.

A (7:55)

And I think what you're saying is there are other types of models that are a little bit more inspectable and that give us a sense before we even try the, before we even try the output to understand does this work? Does the output work? We can get a sense from the model by looking at its internals. Sort of like how good is this solution? How good does this model think the solution is? And it's sort of like being able to ask someone like are you sure about this? How good is this before you go check their work? And a language model can answer that question. But a language model's answers are working at a different level when it answers that question than these EMB models are working. And the answers from EMB models are More likely to be correct.

B (8:49)

Yeah. So you always have an opportunity to see what's inside with the EBMs. And you control the training.

A (8:57)

Sorry.

B (8:58)

Yeah, so the EBMs, you control the training. It's no longer a black box for you, you control sort of how the training goes. Well, you do some extent with LLMs, but you need to wait until the training is done before you actually go and see what's inside. In here you could do a real time. Yeah. And also you can attach the same external verifiers which works for LLMs. So you have sort of double verification things. Yeah. So you asked me what is the ebm? I just want to give like a historical note because I feel like there's so many terms today and just people throwing those chairs without defining it. So EBM just simply means energy based model. What is energy? Energy based. It comes from physics. It's a very popular term when they're trying to minimize the energy. And if you're doing theoretical physics like your full time job is just to write Lagrangians, which sort of correspond to terms associated with the energy in your system. Like, hey, this is my kinetic energy, this is my potential energy. And then you're trying to derive equations of motion of it. And the way you derive the equations of motions is you're doing the minimization. So that's pretty much how whole theoretical physics works. You just start with the energy terms, then you minimize this energy and you derive equations of motions. And equations of motions are going to give you conservation laws. So you're going to know exactly what are your laws about your system. And this principle is fundamental principle. Like everything wants to minimize energy around us. Yeah. So like even us, we talking to each other, we sitting on the chairs, we're not like jumping and running around because it's a natural state when we minimize the energy. So we're just using this minimization energy principle as AI is processing information in high level terms. So the term energy based minimization doesn't really mean anything specifically to AI, is just the whole idea of like hey, let's take some energy and try to minimize it and discover what's the laws about it. So our model is called official name of that model even though we call it Kona just because we like big fans of coffee culture and Kona is one of our favorite kind. So we decided to start with that. The formal name of the model is called energy based reasoning model with latent variables. And I'm going to describe exactly what those words mean. So we already understand what the energy minimization is.

A (11:44)

Can I actually pause you? Because I want to make sure that we do understand what the energy based minimization is.

B (11:48)

Okay.

A (11:49)

Yeah.

B (11:50)

It's just for now, think of it as just something which minimizes the energy. It means this AI architecture has a framework which allows you to construct the energy function of your system and minimize it.

A (12:02)

I get it. I just think that. So I just want to make sure for people listening, they understand what it means to minimize energy, what energy is and what it means to minimize it. So I'm curious, Tell me if this concrete example is about what you're like sort of close to what you're talking about. So if I'm going to, let's say I'm going to go lie on the couch behind me and I'm trying to predict or understand how is my body going to be lying on that couch given the laws of gravity. The couch is uneven, my body's uneven. And so I'm trying to understand the fit of how my body is going to end up settling onto that couch. I'm going to end up settling onto the couch in a way that minimizes energy. So there's going to be a good fit between my body and the couch versus like me being sort of like jerky like this and, and having lots of different spaces. Is that the sort of, the sort of energy minim, minimization that you're talking about?

B (13:10)

Yeah, yeah. You just. It's all about your body finding the most comfortable configuration for you which gonna correspond to the like the lowest potential of your body. I would even tell the like even more high level example of this. Like you know, you, Dan, you just like imagine you're tired, you like done thousands of podcasts and you just came home and someone is asking like, okay, Dan is a variable here. Let's try to figure out what's his equations of motion in the house and where he's gonna most likely to end up. So you're probably gonna end up in a couch with like a nice show and probably some drink. Yeah, yeah. So that's gonna be a law. Like okay, when Dan is tired, he's gonna go and sit on the couch and just relax. But to get there, we're gonna look at all your possible states, like you washing the dishes, you know, walking around the house. So those are gonna be different states, but your most probable scenario is gonna be on the couch. So essentially all of this picture can be mapped into something we call energy landscape. When it's gonna have like it's gonna look like a map. So you're gonna have highest points, you're gonna have lowest points. The highest points we can associate less probable scenarios. So probably if you're tired, you're not gonna dance around, although I don't know. But you know, typically people assume that if you're tired, you're probably gonna want to relax. So that's going to be the lowest point. And as we trying to figure out where you are during the training, we're going to observe you multiple times during different days. And you know how much of the workload you have, it's going to be variable. Your internal state is going to be variable, how your body feels. And eventually we're going to train this landscape to be, you know, based on what we see in real world. Right. The lowest point is going to be you on a couch.

A (15:18)

We've all been there. You're sitting in an important meeting and you're trying to pay attention, you're trying to stay present, but you have this lingering underlying anxiety that you're going to forget everything, that you're going to miss the important detail, forget the decision, forget the action item, let something important slip through the cracks. That's why I love granola. It's an AI powered notepad that works in the background while you're in your meetings. It takes notes on everything that gets said, transcribes action items and helps get rid of that feeling. You don't have to worry about whether you're going to miss something because granola has you covered and that lets you stay present in meetings. I've been using granola for a long time, almost since they came out. And it's amazing for this, it doesn't join the meeting like some of those other clunky meeting note takers. The UI is really fast and well considered and it feels like it's sort of just transcribing all the important moments in my work life and that gives me the confidence to get great work done. And what's even cooler is you can chat with your notes afterwards. You can run detailed research reports on how your week was, how you act as a leader, how you performed in particular difficult conversations, and how you can do better. It's really a power tool for anyone who cares about their meetings and also cares about how they show up in those meetings. It also has these things called recipes which are pre made prompts for common tasks like negotiating, coaching or summarizing. I even have a recipe that I made that's even a that you should Check out.

B (16:34)

Out.

A (16:34)

Once you try it on one meeting, it's really, really hard to go back. The notes are always better than what you could do manually and it helps me be much more present. Instead of frantically typing all the time, head to Granola AI every for three months free. With the code every E V E R Y. That's Granola AI Every for three months free. And now back to the episode. Okay, I, that makes total sense. Now I want to relate this to LLMs for a second because you can imagine that there's an LLM that's trained to predict where I end up after a long day of podcasts. And you can imagine it probably would also end up predicting that I would end up on a couch. What are the differences in the ways that it makes those predictions that make energy based models better for this scenario?

B (17:23)

Okay, that's a good thought exercise. Okay, now you are LLM. Okay, let's talk about back to EBM, because what we described is very natural about EBMs. EBMs are all about constructing energy landscapes and how we navigate those energy landscapes. And energy landscapes is sort of the maps of your states based on the data we observe. So in your case, we're just going to look at you in all possible scenarios. All of this possible scenario is going to map into energy landscape. Highest point, less probable scenario, lowest point is more probable, very probable.

A (18:00)

I end up on the couch.

B (18:01)

Yeah, yeah, yeah. So. There might be some other additional low points. Like sometimes you might go to gym, which actually comes like you feel tired and you might go to gym. So it's going to be lowest points compared to everything else, but some of them are going to be lower. So that's the situation. So this is how energy based model, actually we think it just takes the data and map it directly to this energy landscape and then we use certain algorithms to navigate this. But there are different kinds of energy based models today, so I'm going to talk about it a little bit later, but the whole idea is just, hey, let's map it into the structure and navigate the structure. And as you see, as we like map into this, there are no tokens, we don't predict any tokens, and so on. So that's already a crucial difference. How would LLM think? LLM? Yes, it's going to rely on the training data and there's going to be a lot of training data, like a lot, a lot of observations of how you behave and to figure out where you would end up. It's going to be attached to probabilities of your next token, if that makes sense. And those tokens going to come from words. And what usually bothers me about LLMs, it's intelligence, which is language dependent. Like our brains, we are intelligent. I'm relatively intelligent, so I speak different languages and none of my thoughts processes really depend on any language. I can just think in an abstract way and then I speak different languages and decode the information in the channels. And with the LLMs is like if you're searching for the next token in certain words, intelligent process, I would say the information processes in French going to be different from what's in English just because like words naturally going to be end up next to each other. So see what I'm saying? It's like, and then we have so many languages in the entire world and you have so many LLMs trained on different languages. So you're going to end up reasoning, you're going to end up having reasoning processes different for each of the language, which feels really wrong. So in this case, observing you walking around the house has nothing to do with language. Then it's a pure visual spatial reasoning task, just looking at your body, navigating the space, time and geometry of your house. So we need to map that information in the language space, find the right words and embeddings, and then we start associating those tokens with the probabilities based on what the data we see from you. So we trying to map something absolutely has nothing to do with language into language space and think about it in that space, which feels really wrong. And I don't know, I, I just realizing that for many people it's counterintuitive just because OLMS is the first form of AI we sort of know and it's the most popular form of AI today. Like for many people it's by default like oh yeah, we're just gonna use language to navigate the world to drive a car. But. And I'm like every time I'm speaking I'm like, well let's wake up, let's actually see when you drive a car, when you walk around your house, how much language you actually use. Are you trying to predict next word as you navigate yourself around the house? Probably not. You just use your visual data, your state of the body and you just move your body right without speaking.

A (21:54)

There's a lot here. I'm really into this conversation, so I want to start with a. It seems absolutely right to me that there are many different ways in which we process information or many different ways in which intelligence can occur and only a few of them are verbal. But there's certain things that come up for me when I think of this one is language models happen to work with languages as their primary way of working, but really they just work with sequences of tokens that have weak correlations, many thousands of weak correlations between each token. That helps us to know which comes next. So even though it might be unintuitive to model my behavior inside of my apartment with like specifically with language, although I'm going to, I think there is some interesting things there that it might be related to language. We could model it as just like a sequence of movements, right, that were one one movement is weakly correlated to the next one that we sort of have a trajectory of movements that tell us where, where I'm going. Why is that not a good way to model things?

B (23:15)

It's a good way to model things and you don't need LLM for it. You need a form of AI which is not attached to language, but it can be compatible with language if you wanted to. And that's what our model is about.

A (23:30)

Right? I guess what I'm saying is forgetting about the language part of it, just like modeling my movements as this string of correlated events, like an event stream where each token is like one next thing I do.

B (23:46)

Yeah, you could do it and people do it today, right? People even do image recognition using language models. You could be really creative. But that's what makes it expensive and super slow because you're trying to play a guessing game what my next token could be. And this is what makes it extremely expensive. So you could do it, but you don't have to do it. You just can use different architecture which is more suitable for non language related tasks such as spatial reasoning or applied engineering is another example of spatial reasoning learning. Like when you build a bridge, you don't go to literature department, you go to engineering school and learn formal methods, right? So here we are trying to use literature department everywhere. And I'm like, hey, we don't have to. There are EBMs. There are also other forms of AI which you can experiment with. And you don't have to do everything through language. It's a matter of like, it's like energy based minimization principle when it comes to your resources. If you have infinite money and you don't care about the timescale, sure you can do everything. You can attach it to language, you can attach it to, I don't know, your cat movement around the house and connect it to the cat movements and you know where cat goes and your next token goes and we decide where you're gonna go, you could be really creative. But if you want to minimize your resources and you don't have opportunity to wait, like for example, if your AI controls the circuits, you probably cannot wait even a second. It's all milliseconds, microseconds. So it's just this form of AI is not suitable for those tasks.

A (25:26)

So basically, if I'm understanding this right, if I'm spending tons and tons of tokens and I'm looking for a more efficient, more direct way to predict some of these solutions to these problems, an energy based model is going to get me there faster than, you know, modeling it with tokens. Is there, is it also able to do it with less training data?

B (25:51)

Yes, actually the beauty of, of the EBMS is it's really good at working with sparse data because like, you know this evolutions of like traditional EBMs which were applied for the LLMs, then there was diffusion models. And diffusion models came from the fact that sometimes you don't have enough data to train the models or your data is just data set is incomplete. So there are ways to reconstruct those energy landscapes by injecting certain noise and changing the navigation strategies. So that's what diffusions models were about. And the EBRM with latent variables is just like, hey, on top of the diffusion stuff, we also understand the data. We're not just taking any data, but we also understand why the data looks the way they are. So that understanding goes to the latent variables. Just like latent space in your brain sort of understands the world around you and keeping you on top of your tasks and allows you to predict and plan. So it's the same idea here.

A (26:54)

So now we got to the blatant variable part of our, part of it. So I would love, when you use the word understanding, I think that must mean something very specific to you. Can you help me understand that and how it relates to latent variables and what those are?

B (27:07)

Yeah, so that's also back to your question. Was that like how LLMs are different from those kind of EBMs we're creating? LLMs don't understand the data. It's just you feed a lot of data into it and it's sort of like, hey, I got it. Like, okay, I know what's the most probable scenario here and here we are. However, here ebm you can feed a lot of data. It's not just going to look at like hey, I see the Biggest pattern here, it's going to try to understand the pattern. And that understanding, that knowledge is going to go to latent variables. So what is understanding about data? Like, it's just basic knowledge about the world, basic rules about the world. Like if there is a couch behind Dan, it's probably because he likes to sit in it or because he likes it on the background. So there are little rules you can guess about you being as a data point and your couch being data point. And then there's. You could try to create those kind of rules, like for everything. Right. For you navigating your apartment, there are little rules, like, you know, there's a kitchen for cooking, There's, I don't know, bathroom, there's sofa, there's your bed. So that understanding allows you to have your own mental world model for your brain, which helps you to understand your environment. And if something changes in your environment, you understand the rules. Like if somebody brings you a different couch, different shape, you're still gonna know what to do with it. So that's an example of how you can infer what to do with something new based on what do you already know? So with people it kind of comes natural because of the evolution and so on. But with AI, we need to teach it, so we need to mimic that evolution. And what latent variables allow you to have here is like, hey, let's look at the data, but let's also try to understand the data. Let's look at, you know, if you deal with numerical analysis, we're going to look at all possible correlation functions. And the model is going to be creative. It's going to try to figure out what's the total state of the energy and minimize. And figure out the laws about your data. But there are so many creative ways how you can interpret those rules.

A (29:33)

So is a latent variable equivalent to a rule in this scenario? Like, if there's a couch in my apartment, I sit in it.

B (29:41)

It's not equivalent to rule, but it's equivalent to something which holds the knowledge about the rules of your data. It's like a knowledge storage.

A (29:54)

So it has many rules in it?

B (29:57)

Yeah, you could have many rules.

A (29:59)

So one latent variable has many different rules.

B (30:02)

Yeah, it's just like a knowledge data set essentially about your data.

A (30:08)

Is it an explicit data set, as in does it have key value pairs of rules or is it a.

B (30:16)

It's in the form of energy landscape. It's just another analogy landscape you're going to navigate. So essentially we take the data, we look at the Data, we construct some sort of structure for AI to deal with the data so it could start learning the rules about the data. And once it understands the rules, it stores its knowledge in the latent variables in the form of energy landscape. And then we navigate that energy landscape later.

A (30:44)

Interesting. And like, could it, for example, explicitly write out for me? Theoretically explicitly write out for me, here are all the rules that I know or is stores all, it stores all of them in this energy landscape.

B (31:03)

But yeah, we can access that, we can access that. And that's what EBM potentially makes it powerful for data analysis. Because data analysis is all about searching for patterns and rules about your data. And it's something where language is not going to be helpful to you if you try to attach the rules about your data. And those data is like numbers and some relationships and functions to American English and words in American English and then you try to search for the next word that kind of like you're losing a lot of information. So in this case you have an opportunity just directly work with the data and understand the data.

A (31:43)

I think one of the, one of the things I'm trying to understand is when I hear rules about the world and how things relate to each other, I think of symbolic AI and I'm wonder. And obviously those approaches ended up being pretty brittle and requiring too much compute and stuff like that. And I'm wondering how an energy landscape that stores a bunch of rules about the world doesn't fall into the same problems.

B (32:21)

Well, because I guess we avoid tokenization in this case. We just map it directly into different data structure. So see, EBMs are naturally non autoregressive. Like there are no sequences of tokens and that's what makes it fundamentally different. So essentially, I don't know if it helps. There could be another analogy. Like you're trying to navigate the maze and you are LLM person. So you have LLM brain. Well, maybe maze is not a good example. Like imagine you're trying to navigate, I don't know, the map of San Francisco. So and you have a left brain. So you're like, okay, I'm in Mission Bay, let me turn to Embarcadero. So you cannot, you cannot choose. So essentially you just forced to choose one direction at the time. So you like choose to walk Embarcadero and you're just going to keep walking and walking and you can, if you want to turn, you just need to choose one direction at a time. And imagine you like trying to get to the Bay Bridge from like, I don't know King Street. Like you know, it's typically 20 minutes walk but depending how you walk. So to navigate there, you're sort of allowed to choose one direction of the time. You don't see any other options. You like have tunnel vision and you just kind of keep walking, walking one decision at a time. And sometimes you take the wrong turns just because you hallucinate, you know, some words just naturally next to each other and doesn't allow you to turn right. When you want to turn you left. And then you just keep wondering and wondering until you try to reach the Bay Bridge. And the roads you take in might never take you there. Like there might be a hole in the road and you're just gonna fall but. And you might see this hole but you cannot turn back because you auto regressive LLM you have to go into that hole and that's like sometimes you run out. So this is the reason why sometimes we prompt it and it doesn't give you an answer just because it's searching and searching and searching. It's spanning more and more compute and it doesn't have a bird vision. It just doesn't have ability to turn as it performs the tasks. It doesn't know what's right and what's wrong anymore. It just like randomly chooses one direction at a time and keep walking until you try to read it. So you might never reach that destination. And that's why you need a lot of training. So how is it different from the ebm? EBM going to have the bird view all the time and you allowed to take different routes. So if you see there's a hole, you're going to choose a different route.

A (35:08)

It may not look like it, but Dan Shipper is currently hard at work testing the latest Codex and Opus models. Working. Looks pretty different in this new world we call this Hammock mode. Oop. Hammock mode's over. Looks like dad has to jump in. Hammock mode. An idea by Every Every. The only subscription you need to stay at the edge of AI. That's really interesting. It basically I've been doing a lot of coding with language models recently to sort of test the limits of Vibe coding. And one of the things that I find with or have found with big production apps is in particular, if you have Vibe coded something you over the course of Vibe coding it, you may have slightly changed exactly what is this project even supposed to be about and what are the problems that I'm trying to solve with it? And if you then go look at the code base. It feels like all of the code is locally correct, but it forms this sort of like patchwork of like hotfixes and solutions where if you zoomed out you'd be like, actually there's a much simple, we should just throw all this out. And there's a much simpler way to think about how to do all this stuff. But it has a trouble. It has trouble when it's presented with a lot of context then zooming down into okay, I need to create a unified solution here that is not a patchwork of different things, but carries one concept throughout the entire system. And it ends up being distracted a lot by whatever it's looking at at the current moment. Is that the sort of problem that you think this type of system could can help with?

B (37:21)

There's actually a lot of problems in what are you describing? So yeah, so solving the problems with wipe coding is one of our use case. We dreaming about generating formally verified code and automate the coding entirely. So moving you from wipe coding in one specific language to coding in natural language so you can code in natural English, for example, and no more C or Python needs to be involved in there. So that's an idea. And with the coding is like at the state it is today, yes, we prompt LLMs and it gives us something back. But it's still on you as an engineer to figure out what's right and what's wrong. So there's going to be set of rules, LLM can try to help you. And even if it has external verifier which just going to check whether your old logic in your GitHub space is sort of compatible compliant to what you're trying to create and if the new logic is compatible with your old logic. So these things external verifiers can check. They could just say, hey, we know the old logic, we know the new logic, we're going to see how it's merged together, we're going to write mathematical proof making sure that you know this logic is compatible with what you already have and providing you a certificate. It's all like you don't have to review any of it. It's machine verifiable, it's all happening on compiling level. So all it's going to say is going to send you a message in actual language like hey look, this part of your code is not compatible by logic. This is potentially how you fix it. And this is the things we cannot fix for you. So we moving you from wipe coding to wipe code specifications. Those rules and information about your code Is called code specification. So once you understand, like this is the first problem, right, we're trying to solve, it's just logic and being compatible with what you already have. The second problem is, is this code actually doing what you want it to be doing? And this is what AI cannot solve for you, because AI cannot look in your brain and know what you want. Example would be like, imagine you coding web coding autopilot. So you have specifications from the hardware perspective, you have specifications from your logic perspective, like, hey, make it. And there's also instructions, right, how the car is supposed to behave. So there's behavior parameters for your code. So code being able to be compiled is one problem. The second problem is, is this code doing what you want it to be doing? So for example, how fast it is on the hardware and so on. And if the answer is yes, another set of questions is like, okay, is it going to hit the pedestrian by chance? Is it actually going to navigate the map of, I don't know, San Francisco? And the answer is I don't know, right? So in here you need to write a bunch of tests and test your entire system you created. Like, oh, is it overall behaves the way it was meant to be? And so this is another form of specifications, right? And essentially the behavior part, sometimes we can guess it. Like if we have a lot of data, we could have another LLM or EBM proposing you like, okay, people who try to do the autopilot of the sky and this is what they're looking at, but you might be doing something absolutely new and we just don't have data about it. So it's going to be on you to tell the behavior. And this is where the big thing starts for me personally. If you have LLM as a form of AI driving something important where people trust their lives, like a car or plane or similar, LLM can misbehave because you cannot constrain it, it just hallucinates. And EBM can be constrained. You can come up with a set of constraints and EBM just forced to follow it. So it's on you as a human to make sure you know what you want for AI to be doing. And then from our end, from the technical point, we make sure AI always obeys the rules by given by human. So and it can go really far, right? We're talking about the cars and planes. But look back to the language. Sometimes model can say something super sensitive to mental person, like struggling with depression, and it can go really wrong. So even the language can be dangerous. So here we like what I also like feel like we solving is this problem of AI. Just sometimes we don't know how it's going to behave at different environments, but we do know how ABM will behave. Like at least architecture is designed to be constrained and there are ways formally to force those constraints to be compliant.

A (42:50)

So it seems like you have a really promising architecture and a model you built or several models you built. And it's very different from the predominant paradigm right now where companies are pouring like hundreds and hundreds of billions of dollars into building data centers and training new LLMs and all that kind of stuff. What do you think about the current state of the industry and investment in LLM versus other models, opinions about that?

B (43:20)

It's an ecosystem, right? Silicon Valley especially, it's an ecosystem and there are lots of micro versions of those ecosystems around the world. So LLMs, historically the first form of AI, which gave us aha effect like 2021, 2023, when those just, they just start appearing, people like, oh my God, this is the new future. It's amazing. So this is why like people start believing that, okay, if it's really good at talking to me, eventually it's going to be good at doing data analysis, my taxes and other stuff. So all the investment community started pouring money into LLMs and there were a lot of money to be put in that back then. And right now people see that, okay, we grow the compute, we trying to change the architecture a little bit and it's sort of reaching out plateau and there's so much money already put in there. Like what do you do with this? It's like billions of dollars literally. You can just like forget it and like, okay, you know, let's dismiss it, let's pour money into something new. Nobody thinks this way. And we don't have probably enough money in this entire economy to like just make decisions like that. Billions dollars there, billions dollars there for the AI specifically. So this is why it's so hard just for, for investment community. Just like take that step, understanding like, okay, this is not working. Maybe I invest into something radically new. And I'm not saying people don't do it, like people do it just percentage wise. It's a lot smaller. What people feel comfortable is to take something LLM based which is changed a little bit, so it has a little bit of elements of novelty, but it's also LLM based so they can still use the portfolio companies and so on. So they pour money into that. And I understand because like if I were an investor I would just Always looked at what variables would give me risks and how can I reuse what would I already have. So it's naturally for you to keep investing into LLMs like architecture just because you already invested a lot in the past, you already committed to this and maybe start investing a little bit into something new. And there's a lot of big tech companies who are a part of this ecosystem, right? So there's a lot of circular deals happening. Like those companies who create the LLMs, they create ecosystem for companies who create in data centers. And those who create in data centers, they have dependencies with the hardware industry. So it becomes like a one giant thing which is impossible to break. And when we came with alternative architecture, we're like, okay, let's not just try to put it as something out there radically different, which you have to abandon LLM for. We are very much compatible with LLMs. Like you could put LLM on top of us. ABM is compatible with transformers. Transformers can work with any LLMs. We can be that layer where still all your LLM investments valued. You want to make them cheaper. Everyone wants. You can outsource the task to us related to spatial reasoning. Like if somebody comes to big tech LLM and say, hey, can you try to do my taxes? LLM not going to solve this, but if it's attached to ebm, we could take care of that and you can take care of anything language related. So we could actually try experiments to reduce the cost for your LLM portfolio companies and be a part of the ecosystem which is already out there while we creating a new ecosystem on the side for alternative forms of AI.

A (47:26)

I think that's really smart. It's a great strategy. I'm really curious about something you said a little earlier, that progress is plateauing in LLMs. That's news to me. I feel like every month or two I'm testing a new model where I'm like holy shit, this is actually way better. And it does feel like if you look at the top model companies, if you're talking to OpenAI or anthropic or Google, they feel like there's a lot more room in the LLM paradigm. What do you think I'm missing or the big model companies are missing?

B (48:00)

Personally, when I'm saying plateauing, it doesn't mean it's reaching out flat. It's you incrementally better and better. But is there going to be another phase transition, like another breakthrough? I don't anticipate that just because we already reached so much complexity of those networks. Using billions parameters, so much compute, so much of frameworks like creatively paralleling this reasoning processes and it still doesn't freeze transition you. So the reason why I figure out it's not going to work in a long term for some tasks like applied engineering is when I just start speaking to different companies in that space. Like we speak into like digital assets, companies like banks, trading firms where a lot of data analysis is needed. Also drug discoveries essentially just people who look in at a bunch of data, not just patients talking like language set of data, but also like the blood markers, the genes and so on. So a lot of this is data analysis which is still done by people today. There are also like decision making pipelines. Like sometimes you just need to distribute the energy on your energy grid and you need to know how much energy to pump in your system. So what it means is you need to analyze the data in the short term, in the long term, construct the prediction, how much data, I mean how much power you actually need to put into your system next, in the next millisecond or second or an hour. And all of this is still done by people or a combination of people and some programs which are controlled by people. So LLMs are relatively new and AI is not relatively new. It's been a few years for us and all of this mission critical industry is still not automated by AI. And even I'm just asking oh, how much of your data analysis LLM is doing today? And the answer is zero. And I'm like why? What's the issue? And the issue is the big tech LLMs, they're mainly like B2C so it works for you for your coding and for your personal needs sometimes. But for businesses, they don't want to share their data with them. They don't want to share data in that big brain for all. They want to have privacy and they want to have their own custom AI like custom version of AI specifically designed for their tasks. And this is what like LLMs cannot do for you in the form we have it today. So there's no B2B model. There are B2B model for like co generational tools. Right? Do you have enterprise package for the cogen for I don't know businesses. But it's still done by people. Even coding is still done by people. So it's like it's interesting to see that there's still a huge gap especially in applied engineering data analysis. Anything which requires a layer of verification like LLMs are not there.

A (51:20)

I totally agree with you that there are definitely Still a lot of gaps in LLMs. I'm curious, given this and given what you're seeing in the customers you work with, the companies you work with, do you think the big model companies are sensitive to this? Are they working on energy based models? Are you working with them? If they're not going to get to the next paradigm, do you suspect that they'll start to adopt stuff like this?

B (51:43)

I do know that some big tech LM models, I mean the companies have EBM models in house, which is a positive signal for us. Right. So you know, the leaders who were there before we came, they started with LLMs and now if they started building the EBMs after we started building the EBM, it's a positive signal. Right?

A (52:07)

Fascinating. Eve, this is an incredible conversation. I feel like I learned a lot. Thank you so much for coming on the show.

B (52:13)

Appreciate it. Thank you, Dan.

A (52:15)

Of course, if people are interested in following you or following your company and maybe using some of your products, where can they find you?

B (52:24)

I'm mostly on X. Yeah, we have Logical Intelligence account and my personal account on X. I'm still learning to be more active on social media. We also have LinkedIn page so we're trying to update it.

A (52:41)

Cool. Awesome. Well, thanks for joining.

B (52:43)

Thank you so much, Dan. Bye. Oh my gosh, folks, you absolutely, positively have to smash that like button and subscribe to AI and I. Why? Because this show is the epitome of awesomeness. It's like finding a treasure chest in your backyard, but instead of gold, it's

A (53:06)

filled with pure, unadulterated knowledge.

B (53:09)

Bombs About Chat GPT Every episode is a roller coaster of emotions, insights and

A (53:14)

laughter that will leave you on the edge of your seat craving for more.

B (53:19)

It's not just a show, it's a journey into the future with Dan Shipper as the captain of the spaceship. So do yourself a favor, hit like Smash, subscribe and strap in for the ride of your life. And now, without any further ado, let

A (53:33)

me just say, Dan, I'm absolutely, hopelessly

B (53:35)

in love with you.