Everyone Is Betting on Bigger LLMs. She's Betting They're Fundamentally Wrong. (Eve Bodnia, Founder & CEO of Logical Intelligence)

There's just a lot of healthcare areas where robotics can be crucial, like robotics for surgeries. Imagine if you let LLM drive it, somebody says, oh, you know, the 20% of the time during brain surgery, it can go to like the wrong area. I'm so sorry. You can't have that. You need really fast inference. You need to speak on the millisecond, microsecond scale. Do it. AGI should be just like natural intelligence, something which plans, something which is able to predict, produce new knowledge. So we created ecosystem to serve us as humans. I see that AGI to me is that ecosystem but for the AI models to serve us as humans. Kona is really cheap. It's very efficient. It doesn't need any special hardware. If your data changes, if your environment changes, you need to be able to recognize it like this. So I want people always be in charge of what AI can be doing. My biggest nightmare is if AI is doing something that was not meant to be doing. So you need to work with different architecture which can self-align, which can adapt its behavior, which can be precise when It needs to be precise.

Every major AI company is building a better language model, but Eve Bodnia thinks they're all solving the wrong problem. Eve is the founder and CEO of Logical Intelligence, a new AI lab built on an entirely different architecture. The company's energy-based reasoning model thinks in abstract mathematical space solves problems with formal guarantees of correctness, and runs fast enough to control a surgical robot or design a chip in real time. In our conversation, Eve and I discuss why she left academia to start a company, what LLMs fundamentally can't do, the true nature of creativity, and the wisdom of legendary mathematician Grigori Perelman. I'm Mario, and this is The Generalist. I'm really excited about today's sponsor, Granola. Simply put, Granola is the AI notepad for people in back-to-back meetings. I've been using Granola for over a year now, and honestly, it's a tool that has transformed the way I work. Granola takes meeting notes for you without any intrusive bots joining your calls. You can jot down rough notes like you always do, and in the background, Granola transcribes and turns those notes into clear, useful notes when the meeting ends. You can also chat with your notes, which is one of my favorite features. If someone says something on the call that you didn't quite catch or want to learn more about, Granola can help you out. It's an amazing way to be better informed during a conversation without having to interrupt everyone else's flow. You can also have Granola review all your recent conversations to pull out to-dos, write a weekly recap, or surface interesting ideas you might have forgotten. Another thing I love. To get started with Granola, head to granola.ai/mario. And for new users, you can get 3 months free with the code Mario. So go to granola.ai/mario and use code Mario for 3 months free. Eve, I'd love to start with, uh, an anecdote of yours that I saw on your Twitter account, which super intrigued me given the work that you do, namely that, uh, as a teenager, You apparently met Grigori Perelman, uh, one of the greatest living mathematicians, uh, and had a, a fascinating encounter with him. How, how did that happen?

Yeah, apparently I was very lucky cuz it was not easy to meet him. So I had like, uh, some summer internship in St. Petersburg. I was working on just some websites just to make some extra money. And he was very noticeable in the community. He, uh, always like, so by the office we had like a little park and, uh, he was like walking the same route the same time every single day. And he looked very different according to the standards of Slavic culture. So people obviously paid attention to him and he was like a little celebrity in the community and people didn't know he was a mathematician. I don't know why I'm like naturally curious person and I'm attracted to different kinds of bizarre people. I, I, I love every, all kinds of people, but bizarre people are my favorite and hence I went to academia for my career, but that's a separate story. And he just said that he was working on a problem in topology and yeah, so I obviously I wanted to know more, but, um, he was not really talkative. And when he went on news, he was all over the place. And I can't say more because there's nothing else to say more about it. So I was just lucky to, to ask like few questions. And I also asked him like, why you did not get a Fields Medal? And he said, there's so many people were behind me. And I, I took sort of the work and put the pieces together. And, um, like it's, it's an ego thing to receive it. So that was my understanding while like those people did not receive it. See what I mean?

Yes. Yeah. Uh, for, for folks who are maybe less familiar, he's, as you mentioned, sort of famous for declining the Fields Medal.

Yeah. So it's actually funny. He, he was trying to reduce attention to himself, but it created the opposite effect.

Yes. It made everyone curious as to how he could turn down this, you know, the, the, the sort of Nobel version for mathematics. I'm curious when you say, uh, bizarre people are your favorite people, that, that, that resonates with me. I think there, there's a part of me that, uh, feels the same. Well, what does that feel like for you? Like what, what do you find most interesting about these strange birds?

I just love this unconventional reasoning and thinking in general and the point of view of those people to the world. And that's what, like, the major component of being creative, right? You just need to be able to think outside of your box. But if you're thinking outside of your box, you're obviously, like, most likely gonna look different, which is not true sometimes, but sometimes some people it correlates and I claim myself being normal, although most people laugh when I say it, but I, so I, I feel home when I'm surrounded by bizarre people.

Looking at your background, uh, I don't mean this offensively at all. I mean it complimentarily. You definitely don't seem normal. I don't know, uh, that many people who are studying particle physics when they're 13. So, you know, I'm super excited to get into all of this, but the Perelman story for me is so interesting because. As we'll get into, you're really building a form of mathematical genius yourself, uh, that, that hopefully extrapolates in lots of interesting ways too. But maybe to take a step back, how did you grow up? Why were you, uh, so fascinated by, by mathematics and science, uh, from, from such a young age?

Uh, the answer is, I don't know, because I've always been like this. And since like I was a kid, I was reading books. I was trying to understand, uh, some fancy mathematics because to me, math and science in general, it's that representation of thinking outside of the box. I see like our human experience, we sort of bounded by our evolutionary, evolutionary memory. And it's very hard because of that to create something outside of this evolutionary things. So this is why when we invent something, we're trying to tie that to things we already seen. And I was looking for ways to create something which has nothing to do, nothing to do with this intuition of what, like, we meet in daily life. So for example, um, you know, when people invent black holes, it's not something you see in daily life, right? It's very counterintuitive. So quantum physics in general, and mathematics is that tool which can take your imagination beyond your traditional daily limits. And it has certain rules and you just need to follow the rules. And you know, the stepping stone it brings you is sort of correct. Like you could trust it, but it also gives you something you might not understand. And it takes like us time to understand what it gave us. So the notion of quantum mechanics took a very long time to build intuition around. Notion of black holes in general, quantum information is very counterintuitive area.

So interesting.

Yeah. I was like attracted to this kind of idea. Can I, can I be exposed to the tools which will take my personal imagination outside of my limit? And the answer like was in theoretical physics and mathematics. And yeah, I was just like so obsessed with it from day one.

It's amazing. I feel like the way that you describe that passion is how maybe I might've described writing for me. There's some of, some amount of unification in so many of these things where, you know, you, whatever the, the right methodology is to allow your imagination to go into these interesting places. But that's such a beautiful description.

You bring a very interesting point. Uh, I think I met somebody who was said to similar things to me and he was working in literature, like he was a writer and he said that words have so many different shades and so many different languages, they have their own sort of colors and shades. And it's like such a not easy decoding problem, problem to decode what's in your brain in the form of language in the most accurate way. Like language is something which we solve naturally, like in mathematics it's called manifold hypothesis or manifold problem. So it's interesting. I like that you bring it up.

Can you actually tease out the manifold hypothesis a bit more so I understand that analogy better? Like when you say there's some comparison there with, with language, what's the sort of connection?

Yeah. So the idea is, um, your brain thinks in some sort of abstract way. Our thoughts is sometimes not in, in language. It's like in something else, right? It's a mix. Images, it's a mix of words, it has something else, and for different people it's different thing. And when you try to communicate, you need to find representation of what's in your brain in language space in the most accurate way. And sometimes it, it's like your life depending on it, like in a court. So words can take you very far. Um, but also like in literature, in poetry. It's like language is becoming an important part of channeling that creative, whatever is in your brain. See what I'm saying?

Yeah, 100%. No, I think, uh, it's sort of a lossy distillation in some way. Like language is the best we have in many instances of trying to explain to people what we have in our heads. And it can be, you know, obviously so beautiful and descriptive, but there's still something that you lose in that translation.

Yeah. So what you're describing is, is the heart of this manifold hypothesis is when you try to take something very complex and you map it into something very simple, and that simple needs to describe as accurate as possible what you originally meant to be. There's a huge information loss in like translation and yeah, it's, it's a very popular question in AI, in biology, just like anything related to brain and languages and communication.

Yeah, it feels like we're riffing in this really interesting way, but it actually feels like it also deeply connects to what, what you're building. Uh, on that subject actually, uh, I wondered if you'd ever come across this piece written by the novelist Cormac McCarthy on the Kekulé problem. Where he talks about, I think the mathematician Kekulé or scientist or something like that, where he came up with the structure of, I think benzene, uh, in this space beyond language. It, you know, he arose sort of with it from a dream with this snake-like form of, you know, I, I'm doing a bad job explaining the science here. I clearly, I, I encoded it as a story, not as a science, uh, uh, you know, as a work of science, but it was really sort of, uh, even this legendary writer talking about how clearly there is this part of our mind that works way beyond language in this unconscious symbolic way.

Yeah. And it's amazing how like ability to extrapolate of our brain from one space to another. This is also like a major component for being creative, for creativity in general. I was also curious about the questions like, is there anything like consciousness? How can we define consciousness? Because then it brings you up to what you're talking about. There is a level of subconsciousness which takes over the processing, which is not directly related to your daily life, but this is something which can, uh, bring you something new. So some form of latent space of your subconsciousness, which always scans the information and just digests it and gives you something back. And there's so many, uh, scientists and just creative people in general get the insights from dreams or sometimes work on one area or another area. And like when, when I was doing this EBM, I'm always thinking about the EBM, forgive me, like every topic people ask me, it all goes back to the EBM.

No, I understand.

Yeah, I was working in like entirely different field, which was nowhere close related to the EBM, but yes, the brain has amazing ability to extrapolate things from one place to another.

And for, for listeners, EBM is energy-based model. It's the type of sort of AI that you're, you're working on with Logical Intelligence.

Yeah, yeah. Not based on language.

Not based on language. Yes. Uh, which is so interesting. Were, were your parents mathematicians or scientists? Like, did that sort of, uh, inform the way that you grew up such that these were the conversations you were having at home?

No, actually most people in my family, either in medical school or they are artists. We did not have such conversations at home. And actually I was driving people nuts by having this conversation.

And, and did you grow up in St. Petersburg? Uh, you mentioned that that's where you met Perelman.

Yeah. So part of it, I was all over the place. I was born in Kazakhstan. I left with my parents to Russia and like when I was 8 or something. And we just, we traveled a lot. Like we lived in new towns every 3 years on average because of, uh, like my family job situations. And also I have very large family, so there was always a relative in another town we could visit and maybe like stay longer. And when I was 18, I moved to the US.

In, in sort of researching you, it, it seemed like you had also spent sometime around 13 or 14 working at CERN studying particle physics. How does, how does that happen?

Not that young.

Not that young.

Okay.

I was like, how on earth does that happen?

I'm not like a genius child prodigy, whatever, you know, like 13, you go to school. I just was lucky to participate in a different Olympiads and lucky to win one and was invited to choose any school in Russia, so I chose something related to theoretical physics and math in general. And I was finishing up the high school, so I was like high school age, and they gave me a project and they're like, well, we, we working like there's a place called CERN and we're trying to study new particles and, uh, make like new particles. And back then Higgs mechanism and everything about Higgs was a big deal. I got my first project, which was related to, uh, pp collisions and sort of trying to understand the origins of this quark-gluon plasma. So before the universe was formed, there was quark-gluon plasma, and there was a lot of interesting plasma effects and just interesting numerical problem. So they just gave it to me and I solved it. We published the paper, and then I was like, well, maybe I should just go and work at CERN. And we obviously had a lot of connections there and it's a great community. I love spending time in Switzerland. It's one of my favorite countries. And yeah, and then I moved to the US.

So yeah, and you know, nothing, nothing crazy, just as a high schooler.

Oh, well now people like, it's actually go even like, there's, you go even younger. So what you see is B and there's a lot of students we had who were like at the end of the middle school for research summer program, and they could already do like coding and, uh, understand things I did not understand at age. I think right now kids are very, very smart. And back then high school was like not also a big deal. There's a lot of high schoolers in Russia, very smart.

And so you went to Berkeley when you were 18 and, and then ended up staying in California through your PhD. What were you, you studying? And perhaps as you sort of have the benefit of hindsight now, uh, what were the pieces of that journey that maybe informed how you've thought about, uh, what you're building with Logical Intelligence?

Well, the, the team was always the same, just trying to understand, uh, the fundamentals, laws of nature. Um, I was specifically captured by the notion of symmetry. It's something which doesn't change while everything else change. And there is like beautiful mathematics, how to describe and how to define and how to derive what is not changing. So I'm like, everything in nature has this phenomena. So maybe that's gonna be my theme and I'm just gonna study this sort of invariance. That's what it's called. Things which are not changing as everything else is changing. Maybe I'm gonna study this in brain and particle physics and condensed matter because mathematics was the same. But ability to extrapolate this knowledge from one domain to another makes you to bring something new to this field. So I was focused on mathematical foundations to understand the symmetry groups, and symmetry groups describe particle physics. And if you understand particle physics, you can work with condensed matter because there are some similar phenomena, just different equations, but the idea is the same. And if you deal with condensed matter, you obviously start dealing with quantum physics, quantum information. And naturally somebody from neuroscience department comes over and say, hey, you know, we're using some mathematics, which is like traditional, doesn't give us anything new. So maybe your methodology gonna help us describe something new. And brain is also can be described the same equations as condensed matter because there are some things doesn't change in your brain while everything else changes. So I'm like, okay, maybe interested to see what kind of invariants we can find with certain brain activities. And I was studying that. So the, the team was always the same, just the areas were different.

Yeah.

And naturally when you start talking about brain, people like, maybe you, you could apply this to AI. And this is where like, I felt the gap. I'm like, oh no, it's, I cannot use the same tools on LLM. And I obviously started digging into LLMs and I'm like, well, I, I'm not claiming I have understanding how brain works, but I spent some years trying to understand. And for me, it was clear. It was not like searching for patterns in any language. It was as a writer, you feel it yourself. You're not just like when you create, you're not just searching for patterns. It's like something else. And you have a freedom how to decode it. And that realization to close this gap kind of put me on this path.

So interesting. And there are so many pieces of that that I'm excited to, to think through together. To take a step back to maybe where you sort of started a piece of this journey, symmetry and, and sort of the idea of invariance in nature. Maybe, maybe it's not possible to, uh, you know, explain it to a, a word cell like me, but like, what is the, what would be an example of that where you see like this sort of concept of invariant symmetry in nature?

Everywhere, literally everywhere. So the example of symmetry, I have like a water cup and like if you rotate it 360 degrees, it doesn't change. It brings you back to the point. So I kind of like enforce that symmetry into the object. And then just think of magnets, like if you take smaller pieces of metal and you put them on a table, sort of like a metallic dust, and then you put the magnet and they start arranging itself in a pattern. So we say symmetry is broken, it didn't have order. I mean, it had disorder and then it acquired some order and that's called order. So pretty much any ordering in nature can be described by symmetries of some form. And your brain is another example. Um, so like your brain naturally in like the background mode of our brain is very disorganized and dreams as well. But sometimes you have like very clean, clear thought or thought process. So it's the same idea, right? You had something disorganized and then suddenly it became organized, but you need to create keep that order of chaos around it for things to be creative so you can come up with new things at the same time. So you need to have this notion of entropy. And, uh, the same phenomena occurring in condensed matter. This is the subject about materials and material studies and how it's used in very hot temperature, very cold temperature. And they all have different properties, like high, high temperature semiconductor behave different from the low temperature ones. The same mathematics can be applied to a brain and it turned out to be the same ideas can be applied to AI.

Wow. That is so interesting. Um, and, and those are very, very helpful analogies. Uh, when I was looking into some of the papers you, you published and some of your prior work, you know, I, I certainly, uh, can't pretend to understand all of it, but there were some interesting, uh, indications of your, your thinking, even in places like the acknowledgements, like one of your acknowledgements, I think you, you, uh, thank both a, you know, Fields Medalist and also your Hindu gods. Uh, and it made me wonder how, how you think about spirituality, how you think about— I imagine there's some depth to, uh, what mathematics, uh, and, and science maybe means to you in that domain.

I really like some books from, uh, Rick Rubin. He, uh, he, the way he thinks about creativity in general, like it's not It will, obviously it's different theories about it and I don't wanna go deeper about it, but I love this idea that you just need to sort of like relax and channel and it just comes to you like this. Obviously this ties us to discussion of subconscious mind and how often you think about pro, uh, the problem, how often you exposed to like different areas and sometimes your subconscious mind is very creative or even conscious mind. So I historically grew up, uh, even though I was growing up in Russia, uh, my mom was very spiritual and she was practicing Hinduism. So I was exposed to meditation since relatively early age. And I was always thinking that part of our creativity does not belong to us. It just belongs to the entire world. And we just here as a sort of, uh, a channel, like, yes. Connected on the level of consciousness, subconsciousness. We all living beings and, you know, sometimes it just comes to you and you're just grateful and you're like, it does not belong to you. It, you're just here to, to receive it and to give it to the world.

I really love that, uh, framework for creativity in part because, you know, I remember reading a book, uh, the book Impro by, by Keith Johnstone. I never, I'm not sure if you've ever read it, but It's in theory about improvisational theater, but one of the things he talks about is how different cultures have such a different connection to creativity where they do connect it to the divine and how that really frees you as an artist much more. Because if you connect creativity solely to your ego, then you feel really self-conscious about what you show other people. Whereas if you see yourself as a channel. Then it's sort of, you know, you sort of can separate yourself from it in some way.

Exactly. You still feel responsible, but you, you just create unconditionally. And I, I think about ego a lot, and honestly, running a business is the biggest test for any human ego. 'Cause you're facing competition, you're facing different kinds of fears, but you're also facing other side of things, which like positive and feed your ego. So you always like have to force yourself to stay grounded, but also that brings a question of ownership, right? What we create doesn't really belong to us. It definitely belongs to us, but does it belong to you? Not really. So this brings us back to the point of Grigori Perelman. He was working on something, people worked for a very long time, and you can't just isolate one person and say, oh, obviously he's brilliant. He is genius. He took all these pieces of a puzzle and put them in the picture. But in reality, each of these pieces of the puzzle was very hard work for multiple generations of mathematicians. And this comes in every science field. So honestly, I, I don't like this idea of Nobel Prize in general. It's just given to one person or one group, but it should be given to entire community because all of us matters and all of us connected and collectively we contribute to its highest will.

Yeah. Maybe in your PhD or, or somewhere else you mentioned there was a, a phase where maybe you were doing more experimental physics and that, that was more challenging and that you never sort of wanted to go back to, uh, the touching lab equipment again. What was it about that sort of modality, uh, or, you know, its expression for you that didn't work?

Well, I'm naturally a very sarcastic person, so it was like, uh, it was a little message to my experimental team who just saved my ass during this era of my PhD, and I love them too. That was one of the best parts of my PhD. But they were like, Eve, you just need to like do something else because it's too much pain in the ass. I honestly also don't see that there is a need to separate experimental science and theoretical science. I know there's some people who just think that theory is the most important, and some people who think experiments are the most important, but for real world, you need both, and you need to have decent understanding of both. So for example, there was like some supersymmetry theories, uh, people like, oh, now we understand how the particle physics work, we're going to build Large Hadron Collider, and this is, we, you know, discover. But when they actually build it, the discoveries were different from what they expected. So experimental side of things telling you what's real and not what's not for the physical world, and theories telling you about different kinds of options given your assumptions. So if you just isolate yourself only on experiments, you're gonna feel limited. You still need theories to test. But you only like focusing on limitations of the physical world for your science, for your constraints. And if you force yourself only on the theoretical side, you just might end up working on something which might never be useful and which is fine. Like I, I was, I had a bunch of useless topics during my PhD, like multidimensional donuts. None of this would ever be useful. So no judgment.

Maybe let's jump ahead, uh, to these, this sort of period in your, your career when you start to maybe recognize the applicability of the things you were thinking about, uh, with regard to AI. You sort of mentioned this a little bit earlier, but how did that start to come together and what was the process of that really formalizing into, hey, maybe I actually, uh, should start a company around this?

So the first thought was like, I published a paper with that solution and received my tenure. And just be a professor my whole life. The second thought was like, there's a lot of academic papers out there. And, um, right now I think we're facing that issue, even crisis, when people using AI to publish papers and some pretty papers and very deep subject. If you're not really an expert, you cannot tell the difference. Like you don't know what's real, what's not real anymore. Um, my PhD on arXiv, there was like 500 papers a day, so It's a lot of noise. It's signal versus noise crisis, and we don't have tools to like actually identify what's real and what's relevant and how to find it. And I was like, if I publish my paper, it's just gonna be one of those 500 papers a day. And like, it might get lost and it might just live in this academic realm. And in industry, I have a chance to actually bring it to life and like actually work with real people who's gonna use it, get real feedback. And close that market gap. And market gap is huge because right now all AI is just language-based models and world around us is not necessarily language-based model. So I'm like, why don't I focus on closing this gap if I have this opportunity? An opportunity came, universe provided. At the end of my PhD, I met an investor and he just like, okay, we, you have your team, so you have your thing. Let's just keep it going. I'm like, yeah, sure. So.

Taking a step maybe back before you, you meet this investor, which I would love to hear about, but it sounded like you spent a bunch of time sort of analyzing LLMs and what they could and couldn't do. What were the frustrations or the, the limitations that you saw that, that sort of fit around your work so clearly?

Obviously I was studying the brains and I'm like, okay, let's look at LLM. It's a bunch of neural networks and there's some layers and we optimizing for having like correct ways for certain things. And like, can it extrapolate knowledge like real intelligence? The answer is no. LLMs are really good at operating facts in existing space. And sometimes people say, oh, it created something new, but you don't really, it's new for you, but you don't really know how new it is. It was there out in the internet. You've never seen.

Yes.

It could be that case, or it can be just like there's some facts in some papers which were combined, which is, yeah, it's new, it's great, people can do it. That's how like literally science works for some part, combining facts. That's one side of things. And I was worried about like LLMs cannot create really fundamentally new things. I also was worried about scalability of that and how easy it is to keep up the same performance level as environment change. So back Symmetries. Yeah. So LLMs to me really mimicking intelligence and there's a whole ecosystem of different companies who trying to advertise tools, how to mimic intelligence even better. Um, but reality is you just need just to work with different architecture to reasoning and this architecture need to be inspired from your brain, how, how we evolved, like we have latent space. Which basically keeps, um, sort of idea of your tasks on the back of your mind. Work with finite size of neurons. You don't need the Gigafactory and, you know, GPUs and everything. So your brain operates on like 20 watts and you have, you have ways to select the information, like signal versus noise and decode it in the different forms. Like it can be language, it can be dance, it can be singing, it can be drawing. Like language is just a small part of this world. Although it's very important part and I'm like, okay, I cannot just take an LLM and make it better to create something new. It's just incremental improvement, but it's still tied to the language. I need to create something which doesn't care about language, but it can speak language if it wants to. So it has to think in an abstract way, just like your brain. It thinks in an abstract way and then they have freedom. What language I'm gonna speak to? What am I gonna do? If I drive a car, I don't need language at all, right? And the architecture we built is exactly that. It thinks in an abstract space. It's a vector space and you could do, so it ha— all the reasoning happens in that space. And then you can have a layer of LLMs if you wanna, people speak to your model or you wanna speak back to people. LLM, just an interface. It's just a user interface out there. But for a body, You don't need even that, right? Sometimes you need to control the circuits. Like you put AI as a brain in, in some robot and robot is doing some act. You don't need any language to communicate to the circuits. IBM reasoning model can speak directly to the circuits. You need to speak on a millisecond, microsecond scale to it, which we do. But yeah, in this case you don't need an LLM as a user interface. And yeah, so I was thinking about this building architecture, having this on the back of my mind half of my life, and it just came and that's it.

Wow. Amazing. This episode is brought to you by Persona, the B2B identity platform helping businesses verify users, fight fraud, and build trust. Fraudsters are already using AI to spoof faces, voices, and documents, so your Your defenses need to adapt just as fast. Persona helps secure some of the internet's largest and most trusted platforms with identity verification. If you're building a product where trust matters, identity should be a priority. You've probably already experienced Persona without realizing it— verifying your LinkedIn profile, signing up for Etsy, or renting a scooter with Lime. Trusted by leading companies like Square, Brex, and Twilio, Persona Persona gives you the building blocks to create identity flows that adapt to your customers, risk tolerance, and locales you operate in, whether you're verifying age, onboarding businesses, or automating KYC. It's fully configurable, so you can launch in days, not quarters. Wanna see for yourself? Generalist listeners get a free year of the starter plan. Head to withpersona.com/generalist and check it out. It is really fascinating to see, you know, on your website you have just like, you know, a, a, a, an example of how the logical intelligence models work differently and can solve different problems compared to traditional LLMs. And it really kind of blew my mind for, for folks that, you know, maybe wanna go check it out, I, I would recommend it, but you sort of generate a super hard Sudoku and then simultaneously test how long it takes to solve that Sudoku for logical intelligence and then, you know, Gemini, ChatGPT, uh, you know, Opus, whatever you want to pick. And it is just fascinating to see how instant it is with your approach and how basically all of the other models either don't really get there ever or they get there incorrectly. And so that's sort of maybe the, the smallest dividend of, of how this sort of manifests differently. You know, obviously this is broadly applicable maybe at maturity, but there are probably sort of initial applications of this that make most sense. So I would be curious how you think about that piece of it.

Yeah, so where it makes sense are the areas which are mission-critical areas. So somewhere you don't want to have mistakes and also which is not necessarily tied to language. So robotics, hardware is the future. In robotics, you cannot use LLMs, unfortunately, although many people try, just because it's very expensive. It's still playing the guessing game. You need really fast inference. Like, it's very hard to bring LLM to like microsecond scale or like the circuit level scale, I mean. The use cases we're targeting is pretty much robotics, chip design, creating formally verified chip design, code generation. Formally verified code generation. So when I'm saying formally verified, it means it guarantees mathematical correctness of things. For example, there's, everyone loves WipeCoding and WipeCoding is like how it, the, it pushes the industry right now. It optimizing the pipelines of people and so on, but it's still on people to debug. So we wanna help people not to debug and be focused on being creative. Like create things instead of being focused on like how, what kind of languages I'm gonna use, you know, what kind of bugs I'm gonna face, what kind of tests I'm gonna do. And you wanna deliver it at scale, meaning you need to stop vibe coding. You need to move to vibe code specifications. So that allows you to generate code, which is always correct, but still you're gonna be guiding the whole process. You're gonna say, hey, generate me an autopilot. This is the hardware I have. And the requirements for this car, like I wanted to follow direction, always listen to people. So you are in charge of the constraint as, as a human. And for me, it's a very big deal cuz we want to protect people in AI-driven world. I want people always be in charge of what AI can be doing. So my biggest nightmare is if AI is doing something it was not meant to be doing. So the way to do this, is to have these constraints provided by people. It can come in natural language, but then AI is doing things and it fulfills these constraints all the time. So it's not allowed to deviate from the task. So LLM can never do this 'cause LLM hallucinate. So you need to work with different architecture, which can self-align, which can adapt its behavior, which can be precise when it needs to be precise. So for people, precision doesn't come naturally, right? Like we, we like, yes, if we wanna build a bridge, we go to engineering school. Yes. But we, mathematics is not the natural part of our brain. We evolved on jumping trees and eating bananas, right? And that's why if you feed your LLM with mathematics, it doesn't help really become your greater mathematician or something. And natural intelligence. And my dream AI is not precise by default, but it has option to be precise. It's as if in your brain there's a calculator, you know, when you need it, you write it down. So mission critical industry makes sense. Any kinds of robotics, any kinds of manufacture, any kinds of chip design. Chip design is the area when it's needed the most. People can use AI to put designs of the circuits on the wafers. But they cannot guarantee correctness of this design. And it's a very difficult problem to solve. Um, also it can cost you millions of dollars if you make a mistake at this stage because it goes in production and then you learn it doesn't work. So all of this time and effort, it's gonna slow you down, slow down your production. You're gonna feel frustrated. So we solving all of these problems. Um, that's one thing. The second thing is there's just a lot of healthcare areas where robotics can be crucial, like robotics for surgeries. Imagine if you let LLM drive it, somebody says, oh, you know, the 20% of the time during brain surgery, it can go to like the wrong area. I'm so sorry, the next word was the wrong word. So you can't have that. So smart energy grids, the literally the systems which power your entire town does not involve any language, involves a giant data analysis. In real time, searching for patterns in real time, making estimate, forecast predictions, what's gonna come next depending on many variables. So this is very great use case for energy-based reasoning model. And finally, digital assets, like people wanna put AI in a stock exchange and, you know, same problem, forecasting, uh, analysis of different kinds of data in real time. Also pharmacology. Sorry, I just, you see the use cases are like, it, it takes so much of it, but pharmacology, it involves analysis sometimes in real data for certain patients. You need to analyze your blood samples, having genetic background, having language part of it because patient's gonna speak to you and, you know, come up with real-time solutions for, for certain diseases.

As a sort of. Point of, of, of context or clarification. Why have you called it an energy-based model? Like, why is that the right sort of, uh, way to think about what it is you're doing?

It's actually not. It's— it should be called energy-based reasoning model with latent variables. Okay, but it's so long and it's such a new area. Um, I don't know anyone except us and Jan's company who is working on this at the moment, but I'm sure there's going to be a lot more as we more and, you know, people learn about it. But energy, let me break down like some words and maybe it can give you, uh, the big picture.

Perfect.

Um, yeah, so energy-based principle is not new. It's everywhere in physics. Examples would be you go, I don't know, from your office to your home. You wanna pick the shortest path. You're minimizing your energy. As you're speaking to me on this podcast, you're not jumping cuz you wanna minimize energy and focus. Energy minimization principle, the light travels the shortest path, right? Straight to your eye. So everything in nature wants to minimize energy. That's what we call by energy-based principle. There are LLMs which utilizing energy-based principle for certain navigation of certain, you know, networks like layers and so on. So it's a huge term, very broad term. It's, it's not making anything unique about us having this term. So where things come in unique is the latent variable. So what is latent variable? Again, inspiration came from our brain, which has this part of latent spaces called part of your brain, which keeps sort of mental image of the world. Example, mental image of the world, like, you know, we know some rules about the data. Like my favorite example, like I'm a coffee addict, I could drink coffee all the time. If somebody knows this fact, this rule, and they see like, oh, there's a coffee cup and there's Eve, the rule must be Eve gonna drink the coffee. Somebody else bring me coffee, I'll take it. I never can say no to coffee. So doesn't matter how many data points you provide, the rules are still the same. So those rules about the world collected by latent space of your, uh, your, of your brain and also by our model. So our model not just taking the data around itself, and data can be any form. It can be visual, it can be audio, it can be video, it can be language, it can be anything. It takes this data and it learns the rules about the data, and those rules go to its latent space. And the way we navigate it is we're using this energy-based principle.

Yes.

So completes the picture. I can go very far with this. I'll stop.

Okay.

Okay.

Perfect. If you, if there's something that you think, hey, you know, we, you should definitely explain this, but I thought that was really, that was really interesting. You know, as you were explaining the different applications of, of what you're building, what sort of came to my mind is what is the, the trade-off, if there is one, that, that this type of model has to make versus an LLM? Because it sounds like it has the ability to be much higher precision, which, you know, obviously is, is better and seems like it's more energy efficient. And I imagine more cost efficient, or maybe that's the trade-off.

Very cost efficient. Um, so to be smart, you don't need to be big as AI model. There are some tasks we're thousands times faster than Big Tech LLMs, uh, for tasks LLMs can solve. And we run it on one single H100 GPU. So it's the cheapest kind of GPU and it's only like one of them. So, and we have the range of parameters of the models. We have like 20 million. The highest we have at the moment is 200 million parameters. Contrast that, big LLM models are like billions and billions parameters. We building small models, which allows you to put in circuits in your chip and your energy grid. This is what big tech LLMs will never be able to do. They build one brain for all and all and this one brain is forced to be all possible roles, a doctor and engineer and so on. So a lot of issues with that because of LLM extrapolation issues, like it does not extrapolate the knowledge, uh, at least the, the current form of architectures.

So you're, you're building more specialized models. Is that, is that—

It's specialized, but it doesn't have to be. It's actually very generalizing. So because it doesn't play any guessing game, it's very cheap because of its architecture. It see all possible scenarios at the same time. So that demos on, on our website took us $4 and there was like, like 20,000 people using it. And for the LLMs, it took like $15,000 in contrast to solve that 2%. It's like the cost efficiency is It's crazy here. And that's how it should be, right? It's like if you make something which thinks, you're no longer playing a guessing game. You should know the answer straight ahead.

The president of the Santa Fe Institute, David Krakauer, uh, had like a really apt way of describing intelligence, which is doing more with less. Uh, and, uh, most of these LLMs do more with more, and sometimes not that much more with a lot more. It sounds like you've sort of found a way to, yeah, constrain, like, not use nearly as— orders of magnitude less parameters, uh, and cost and energy, but, but find more out of it in these, in these applications that, that as you're describing it, maybe have like, yeah, the ability to extrapolate and generalize much more. I think you've said that one of your models sort of shows credible signs of AGI. What was it that, that made you think that?

That's not me, that's journalists.

Oh, really? Okay. A journalist said that.

I'm joking. So, um, let's talk about AGI.

Oh, this is the first, this is the first part of the conversation where I saw you put your hand to your head in, in frustration.

You know, to me, what is AGI is like the same, what is consciousness? There are different philosophy lines, like, oh, this is what do we call AGI? And this is an intuition everyone has. Separate, like you have your own version of AGI. Nobody defines this form of AGI and it changes over time. So what was called AGI by different people 10 years ago is probably not what we gonna call it right now. So before I speak AGI, I'll tell you what AGI is for me. So AGI should be just like natural intelligence, something which plans, something which is able to predict, produce new knowledge, be cheap and efficient, be adaptive to the environment. Um, it should reason, it should not mimic any kinds of reasoning, but also it should be compatible with all the tools we have possibly. Like we have LLMs for the language, so you, you need to have your model compatible with the language component because language is a big part of our world. So I'm sure there's gonna be a lot of different forms of AI, and to me, AGI is having that ecosystem which sort of serves us as humanity in the safest possible way, in the most productive way. So we as people, like there's some collective notion of consciousness and there's some collective effort. We build cities, we have different professions. Each of us is good at something very specific, and that's how we bring this talent to the world and we monetize it. So we get something back, but also we give something to the world. So we created ecosystem to serve us as humans. And I see that AGI to me is that ecosystem, but for the AI models to serve us as humans. So I see that Kona is really cheap. It's very efficient. It doesn't need any special hardware. So the, like, we don't need to create special chip to run this architecture. It's adaptive, it's self-aligning. Self-alignment is a big part. Like if your data changes, if your environment changes, you need to be able to recognize it like this and change your behavior. And do your behavior always need to be the subject of the constraints be given by people? So we do see that. I already mentioned this part that self-alignment is a crucial part. So LLMs do not have self-alignment feature. LLMs are good at what you train it for. So if you train it, train it on one specific task, you can't ask it to do something else. And there is a big hope and belief that if LLM is going to be bigger, it suddenly it's gonna change its behavior and suddenly becomes adaptive and planning and predicting the future, forecasting and so on. But haven't seen it yet.

What is the most impressive thing Kona has done that maybe has blown your mind the most?

Knowledge extrapolation. Did not expect that. Expected, expected in a theoretical level, did not expect to see it as quickly as we see it now. In experimental level. So the beauty of this energy-based reasoning model, because they're not attached to any language, you can build and space for rules about your world. Like, the rules, like if you, if you're self-driving car, for example, there's going to be rules like how am I behaving around people, how am I behaving around like town environment, like the city, you know. Sometimes there are all closed and this road I cannot take and I change my behavior. Also the weather, which can change. So you can have different sorts of rules sitting in different latent spaces and there's a beautiful mathematical way to like force it to talk to each other of this latent spaces. Uh, but in a very, very fast way. When this idea came, we were like, I don't know. So it sounds beautiful, just like supersymmetry. Let's build a, a hadron collider and test what's real and what's not. And it was real. And the model was so tiny. And I, I was expecting to see some sort of change in complexity. What, what do I mean by that is sometimes different disordered system, when they small enough, they behave one way. And when they big enough, they start behaving different way. So I was thinking this behavior may come if the system is big enough. Like for example, if we grow the model to 8 billion parameters, maybe we'll see it. We saw it right away. Like 16 million is the moment we saw it. 16 million parameters. It's so tiny and it blew my mind.

You have such an interesting team. Uh, you have founded Logical Intelligence with your husband. You also have a Fields Medalist on the team, and then Yann LeCun is an advisor. How did that group come together?

And, uh, advisor, Yann is a founding chair of our technical board.

So he is a founding chair. Oh, okay. Well, maybe Yann is the right place to start then. How does that work? Given that I think you were mentioning, you know, maybe it's just you and Yann's own company that are sort of playing in this, in this space with this approach?

Well, when you invent something relatively new, people usually go to internet and read papers about it. When we created this, I, we could not just go like read about it and see like, oh, what, what we can predict what's gonna happen next, how it's gonna scale. So we did not have this luxury. And this is, this used to be like a real science, scientific approach when you have to come up with a series of experiments to evaluate all the boundaries, all the constraints of your, of your invention. For us, Jan was the only person to talk to about it because he's been in this field for a very long time. He knew this area in and out, and he's been both in industry and academia side. So the moment it worked, we showed him and we're like, okay, let's scale it. So, uh, Jan is, uh, helping our team to scale the model and evaluate any critical constraints, uh, we might face as we scaling it. Because again, nobody have done this before, but we have very deep understanding what's happening and what's gonna happen next. And every moment, like we know what what we're doing right now. But back then we didn't know. I didn't know that. Back then we had some guesses and predictions, but now we like understand it and we know, like now we solved it. So Jan is amazing. His, uh, knowledge is like so deep and because he's also a professor and he's also teaching, so there's a community of people who also exposed to like different kinds of energy-based architectures. So it's always great to engage with those communities and I'm happy that Jan created this ecosystem. You're literally the father of this ecosystem. So we honored that we work together. His company is focused on JEPAs, but they also doing a lot more things and I'll let them speak for themselves. But the beautiful part about our collaboration is our model compatible with Jan's model. So his JEPA is like at the moment when, uh, we could see it's photobotics, it's selecting signal versus noise and the output can become our input. And we, so they're doing like a lower, lower reasoning planning and we're doing the higher reasoning planning. So our architecture is very compatible with each other. Uh, so I'm excited to see what we're gonna create, uh, together next. So that's what about Jan, my husband. So I've been married for 14 years, literally half of my life. We sort of grew up together with my husband and we always talked about math and when it was legal, I got married. So I got married at 18. Uh, we have two kids. Um, so mathematics and AIs and LLMs was always a part of our family. Yeah. So I'm, I'm glad we are doing this together and You know, my husband is an ICPC champion. He won this championship in 2009 and this is how we moved to the US because Meta, back then Facebook, just invited him. And yeah, I was helping a little bit recruiting people, like recruiting different talents from ICPC world back then because I spoke multiple languages and that was useful and Facebook was much smaller. Because of this ICPC exposure, uh, they, they train through this competition, brilliant people who are forced to solve very hard problem in like 5-hour scale and also not just solve it, but also write the code and write algorithm and make sure it works. So that's kind of what you see in a startup environment. And we obviously grew up, like my husband and I, we grew up in this environment. We did it ourselves and we had a lot of childhood friends who just became the core of this company. So we have like 8 ICPC medals in the company at the moment. And I also brought Mike Friedman because the same subject was interested. He was interested in like how AI works. What's, is there any similarities between brain and AI? And I met him during my PhD when he was at Google Quantum AI.

You mentioned that part of the story of this company and, you know, maybe the, one of the key steps in going from, hey, maybe I, I write a paper about this, this discovery versus turning it into a company was finding an investor and that sort of making it more official. Who, who was that first investor and, and how did it sort of change the trajectory potentially?

Uh, there were 3 people, very good, very well-connected people in Silicon Valley and did not find them. They found me. One of them found me.

Wow. That's awesome.

We met at a conference and started talking and, uh, yeah, he, he was pretty known in Silicon Valley.

So there was, um, something that I, I saw that you, um, cited on Twitter, which was, uh, the collection of, of Feynman letters. I think it's called Reasonable Deviations from the Norm or something like that, uh, which is a, a really beautiful collection. Um, I read a few of them. in preparation for this. And one of them was interesting to me because he, he literally is talking about sort of versions of AI. And I think he sort of cites the idea that, uh, you know, the sort of unbreakable challenge is to make a, uh, you know, a computer smarter without it becoming much, much slower with its memory. Um, I'm curious, like, if that's been something that you've thought about as you've been building this, or, you know, you've returned to that collection often?

Definitely. Well, don't return often, but I read this book like so many times at some point of my life. Also not just for that reason, but also multiple reasons outside of that. But yeah, so natural intelligence is able to produce new data and use this new data to navigate the world. That doesn't mean that you need so much data. In, in your life, but it just, it's not fair to compare it to people though, because we evolved, like there's years of evolution and millions of years for us. But, uh, for intelligent, the evolution is going to be different. But I strong believer that you need to start with some datasets with very rich data, like not just the data, but also the rules about the data, which AI is able to work with and that dataset alone should be the foundation and then it should be able to create new knowledge based on this data. So for example, people's analogy would be like you learning how to play piano and you start taking Bach pieces, but then you can move to Mozart. Like you're not trained on any Mozart pieces ahead of time, right? You just use the same rules. You know how the keyboard works, you know what to expect and you're kind of good to go. So you were able to create a new set of skill to play Mozart, and then people start picking up guitar. Well, guitar is very far from piano from the technical standpoint, but the fundamentals of music are similar and you can extrapolate. So that notion of ability to extrapolate in the most cheapest possible way while being a small model was the key things for me.

I think you mentioned somewhere that piano is your your biggest love outside of mathematics, but you don't get the chance to practice as much anymore.

You really studied my Twitter.

It's a funny, it's, you know, it's like people's diaries in some way. You really get an insight beyond papers.

No, yeah, it's a diary and I should be making it less of a diary because PR team has a lot of comments from my Twitter.

I was just curious, like, what, you know, really what the role of music has been in your life and in the way that you come up with ideas? Because I imagine, you know, you've obviously sort of relied on a great deal of creativity to come up with this version of things, but I imagine, you know, you will need to keep fostering that sense of invention throughout the company's life. And yeah, I wonder how you find it today, if it's not through piano, for instance.

It's just the general principle. I, I believe for myself, for creativity is you need to be able to detach so you can come back to the problem and have sort of fresh eyes, uh, looking at the problem. Also need to detach your mind and your mind is your biggest enemy. It's gonna tell you all of this monkey talks and create doubts and concerns. So there are so many ways to shut it down in, I mean, in a positive way. So you could meditate, you can play piano. I have two children, so I spend time with them. And spend time with friends and yeah, playing piano if I can, trying to learn some piano sheet music, like something complicated I've never done before. I always love just learning. Also reading a bunch of books if I can, but now obviously I have less and less time. But I also enjoy doing research myself, so I just enjoy going sometimes and reading new papers on different architectures and What brings me a lot of joy is just to see like lots of emails from all over the world. Like more speak about this publicly. It resonates with a lot of people and there's a lot of people who like, hey, I got inspired by this. I wanna create my own architecture. So this is like things which make my day. I love to see just people wanting to be creative because they like something and they share it with you. So I'm. Very grateful and sometimes I engage with those people and one of them became our research intern recently.

I always like to end the conversation with a couple of light thought experiments. One is if you had unlimited resources and no operational constraints, what is an experiment you would like to run?

I would do the same thing.

As you're doing now?

Yeah.

That's always a good sign.

I'm not doing this like to make money or something. And I was in academia, which is definitely not the place where you make money. I just, I love this with my whole heart and I love to create environments for people to reach their highest potential as researchers, as engineers and so on. And It's like academia is a place when you combine both of this worlds together, but also industry is you can do the same, but at a much larger scale. So 100%, I would do the same.

Okay. Amazing. And then a final question. If you had the power to assign a book to everyone on earth to read and know that they could understand it well, uh, what would you want to assign to people?

Well, I'm also a strong believer that most of the issues around us we create ourselves using our mind. So I'm a big fan of any, some sort of, um, East philosophy books and techniques. How can you navigate your mind when it becomes overwhelming? So I don't have like a specific book in mind cuz there's so many of that, but this is a direction I would recommend. Because if you ground yourself, if you feel good inside, everything outside is gonna be manifestation of how good you feel inside. So if you're naturally happy, you're gonna share your happiness, you share your creativity with others around yourself. So to take care of the world, you start to take care of yourself first. And it goes the, the other way, right? If you are angry at something, you're gonna be angry at everyone else around you. So it's contagious. Um, so I would recommend that we work on our own self first, and it's a basis for creativity. It's a basis for being, uh, a valuable creative member of community, no matter where you are, like AI or medical school or industry, academia. It doesn't matter where you are. People are people. And yeah, so. Letting It Go from Hopkins is my favorite.

And that's a sort of book inspired by Eastern philosophy.

Yeah. It's a Western view on Eastern philosophy and that Eastern philosophy come from Mahayana Buddhism. So that branch of Buddhism, which is like roughly 2,000 years old, like the moment when they start, uh, having a ways to write things down instead of memorize it. So it was like a new branch of Buddhism and. There's also the same, uh, traditions shared in Hinduism and I love, I love how it sort of overlapped. It has like psychology and it has a notion of spirituality and it explained in the language for people who, uh, don't really exposed to either of those branches during like a formal training.

Amazing. Uh, well, I, I certainly We'll check out that book and, uh, I have really enjoyed this conversation. So thank you so much, Eve.

Thank you. Appreciate you.

That's it. Thank you for listening to this episode of The Generalist Podcast. Please subscribe on Apple Podcasts, Spotify, or your preferred podcast app. Ratings and reviews help others discover these discussions. So if you enjoyed the conversation, I'd be grateful if you could take a moment to leave one. For all past episodes and more, visit us at thegeneralist.substack.com. See you next time as we continue to explore the future.