#57 - Why AI's next era may not belong to LLMs - Zuzanna Stamirowska

What comes after large language models?

In this episode of The Deep View Conversations, we talked with Zuzanna Stamirowska, CEO of Pathway, to explore why her team believes today’s dominant AI architecture has fundamental limits, and what it could take to move beyond them.

Pathway is developing Dragon Hatchling, a new architecture designed to give AI native memory, continual learning, and a different approach to reasoning. Stamirowska explains why today’s LLMs can appear to remember without actually internalizing what they learn, why reasoning through language creates its own constraints and costs, and how Pathway is trying to build models that can think in a more abstract way. 

The conversation looks at how those architectural changes could affect hallucinations, interpretability, safety, and the enormous compute demands of modern AI. Stamirowska shares how her background in complex systems and game theory shaped Pathway’s approach, why the company made an early bet on challenging the transformer, and how the AI coding revolution has already radically changed the way her own team works.

Topics covered:
• Why transformers struggle with memory and continual learning
• How Pathway’s Dragon Hatchling architecture works
• How a different architecture could reduce compute costs
• How interpretability could make advanced AI more predictable
• Why Pathway’s engineers have largely stopped writing code themselves
• How Stamirowska uses Codex, Claude Code, and other AI tools
• Why leaders should be ruthless about identifying the critical path

If you’re interested in what could come after today’s LLMs, and whether the next big leap in AI will require more than simply scaling transformers, this conversation offers a fascinating look at one of the teams betting on a fundamentally different path.

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transitor.fm 

And don't forget to sign up for The Deep View daily newsletter. We don’t just cover AI, we decode it. In a world flooded with hype, we deliver sharp, no-nonsense insights to keep you ahead of the curve and help you put AI to work every day: subscribe.thedeepview.com

Creators and Guests

Host

Jason Hiner

Editor-in-Chief of The Deep View

What is The Deep View: Conversations?

From frontier labs and enterprise platforms to emerging startups reshaping entire industries, The Deep View: Conversations podcast interviews the brightest minds and the most influential leaders in AI.

# The Deep View Conversations — Zuzanna Stamirowska of Pathway

**Host:** Jason Hiner
**Guest:** Zuzanna Stamirowska
**Runtime:** 56:43

**[00:00] Zuzanna Stamirowska:** It's a bit like Groundhog Day. I mean, then you have that head, let's say it's created, and then as you use it, you feed in information that goes into context. But this context is ephemeral. It's like writing something on a piece of paper and just storing it, you know, the notebook, without ever internalizing this knowledge. There's no link between the queries, like between whatever you put in your LLM or how it's being used and how the world evolves, and its weights that were set once at training.

**[01:00] Jason Hiner:** I met Zuzanna a couple months ago at a dinner in San Francisco, and I was so impressed by what her and her team were building that I asked her to come on the podcast so that you all could learn about it as well. Pathway has come up with a new technology and a new architecture for AI models that solves some of the fundamental problems that LLMs have with memory, hallucinations, safety, and costs. The Pathway team first published their paper and proof of concept in October 2025, and they are preparing to launch their first product to reimagine the future direction of AI. I also talked with Zuzanna about her background in research and game theory and how she's using the massive advances in AI to build a different kind of startup that's prepared to thrive in such a rapidly changing environment. All right. So here it is, our conversation with Zuzanna Stamirowska of Pathway.

Zuzanna, for those who aren't familiar, tell us a little bit about what Pathway does and about your role at the company.

**[02:06] Zuzanna Stamirowska:** Absolutely. So at Pathway, we are building AI that autonomously and continually learns, evolves, and reasons. So our vision is to bring AI with no limits and one that can be sustainable. So if you were to look at the equation of, you know, the cost of intelligence per dollar, we're working on that equation from both ends, enabling more intelligence. Actually empowering, making breakthroughs on just how AI is being built, and at the same time, making it more efficient, such that the physical limits of the earth can actually handle it. And for my role, I actually started it. So I'm a co-founder, together with Adrian Kosowski and Jan Chorowski. We started Pathway. And I'm the CEO, so I'm extremely lucky and privileged to lead a team of incredible researchers. I'm a researcher myself. It's a bit how we got together. I'm pretty much touching, you know, and seeing the cutting edge every day.

**[03:11] Jason Hiner:** Yeah. Well, you know, I've been very impressed by what you're doing. We got the chance to meet in San Francisco at a dinner.

**[03:18] Zuzanna Stamirowska:** This was very good pasta, wasn't it?

**[03:20] Jason Hiner:** It was very good pasta.

**[03:22] Zuzanna Stamirowska:** It's true.

**[03:22] Jason Hiner:** It was excellent. And the salad, actually, was also outstanding. No, it was really fun. And I got to meet you and Adrian, your co-founder. And that was just a bit of a revelation to me. You know, there's a lot of things, a lot of people that are working on a lot of stuff that sounds very similar in AI today. And when I heard what you all were working on, thinking about kind of what's next after LLMs, you know, that really piqued my interest, of course. And so had to have you on the podcast. I, for the audience, tell us a little bit about how Pathway got started. You know, how long you've been doing this and, you know, what was the problem that you all really set out to solve?

**[04:08] Zuzanna Stamirowska:** Of course. So we are a neo-lab, right? Pretty much a neo-lab, I think the definition by now is that a lab that works on AI that is not OpenAI or Anthropic. The first wave of the big A labs that were fighting to scale up the transformer. And now we're doing things that are different. We made that bet in 2024. And back then I created something that internally I called the dogma of differentiation. Because in order to actually make it worth our lives, because this is literally the size of the bet we're doing, right? To make it worth our lives, we knew we needed to make something. We need to build the moats that would be so deep that it would have a chance to stand against, ultimately, the transformer. So I think for the listeners, it may be interesting to know that the T in GPT is the transformer. The piece of, let's say, math that was announced in the paper called Attention is All You Need. And this is the math that pretty much carried us until now.

**[05:16] Zuzanna Stamirowska:** All this AI revolution that we're seeing that, I mean, a thing in the public kind of consciousness started with GPT-3. It all started with the transformer and that Attention Is All You Need paper. And in fact, all the models that we're seeing out there are built on top of that. So they all feel the same. They all behave the same. They're all sequence models. Most of them are language models, right? We say LLMs because they're trained on language. Because language has just such beautiful sequences. And indeed, transformers were built with a lot of language, let's say, in it. Or maybe in how people were working on AI around that time. However, it comes with fundamental limitations. And those limitations are highlighted even by the very authors of the transformer. So it's undisputed that transformer has limits. Like there's pretty much no debate whether or not transformer per se can take us towards continual learning. You know, the real capability of learning over time with experience by updating the ways we're updating the actual skills.

**[06:27] Zuzanna Stamirowska:** But real deep kind of skills that a person would have, right? There's a fundamental problem with memory in transformers, which is kind of linked to the issue of continual learning. So as a user, you may experience the issue of context windows that you run out of. Or the issue of state tracking. So I'm just forgetting kind of context of what matters, right? In the past, while the LLM is like the way your model is solving a problem. Something that people see very often in cybersecurity, for example, when they deal with large code bases. It becomes even trickier as things start to change across the system that you're kind of trying to track. So these are things that are fundamentally difficult for transformers. And then let's say with, you know, until a point, until late 2024, we had the pure transformer. And then a beautiful thing happened. That was o1 and then o3 with reasoning, which is a chain of thought added to the transformer. Yeah. And chain of thought allowed us to work through problems somehow, right?

**[07:37] Zuzanna Stamirowska:** But we're still in the realm of solving problems step by step, token by token type of way, which is very limiting and also leads to an explosion of tokens. And this is something that we knew. We actually knew very early on. The moment we made this, we actually decided to work on the post-transformer architecture. We knew chain of thought would be very demanding in terms of compute. And we also had this like very strong feeling that thinking shouldn't be limited to language. Reasoning shouldn't be limited to language. And then we're also somehow on this journey, you know, being scientists to understand why AI is working, you know, because we felt like, okay, we've built the steam engines of thought without knowing what thought is or, you know, where a concept is. It's like essentially about having thermodynamics. So we felt like, okay, how about first of all, we try to, we studied the steam engine. Okay. We, we see what is, let's say folklore, what is true, what matters, what doesn't.

**[08:53] Zuzanna Stamirowska:** Then we try at least to an extent, understand the rules and find this kind of micro to macro passage of, okay, how intelligent, how reasoning emerges, hopefully from like local sparse interactions. That was a strong thesis we had. And I mean, I came partly with, with that intuition to the team. And this allowed us also to make decisions like design decisions in our research that were very strategic. Because since we had this understanding, we did, we were not losing this much time on quests. Yeah. And maybe for people who are not so familiar with how AI research works or worked for the past years. It is right now, fundamentally an empirical science or I would say discipline. Okay. So generally speaking, we run experiments. You see the experiment moved your loss function, for example. If it moved it, you were happy. If it didn't, then you're not happy and you just, you know, you say an idea wasn't worth it. To get to the place where we managed to get to, we needed to take a number of hits. We actually had to change five things.

**[10:07] Zuzanna Stamirowska:** And every time we're changing just a single one or two or three, we're taking a loss. We're taking a hit. It was like, it wasn't working. So without having this theoretical model of how we believed, you know, the system should work, we wouldn't have done it. We wouldn't have that, we wouldn't have had that conviction. And, you know, but ultimately, well, this is science. So we had a strong model, right, about what we believed would work. And this came somehow from our backgrounds. And this, you know, computation complexity, I come from complex systems, AI per se. But Jan, our CTO actually worked with Geoff Hinton. He was the first person who applied attention to speech, which means we kind of knew what sort of, you know, side things were being done, like why Transformers won. And also Łukasz Kaiser, who is a co-author of Transformers and also one of the inventors of o1 reasoning. He wrote the first check for Pathway.

**[11:14] Zuzanna Stamirowska:** So we had this like very, very intimate knowledge of, you know, what was explored, why Transformers won, where the limits are. And we didn't need to be convinced of that. And second thing is somehow just courage. Okay. Because if you were, you know, if you were to go to NeurIPS, and I remember NeurIPS when it was still called just NIPS, you know.

**[11:39] Jason Hiner:** Yeah. NeurIPS is, for those who don't know, it's like the biggest academic AI conference where all the world's leading AI researchers come together, like at the end of December, you know, in December sometime.

**[11:49] Zuzanna Stamirowska:** Yeah. Yeah, yeah. So if you were to go to NeurIPS and say, you know, I think I will make a smaller model. I will make a different architecture. For years, they would just laugh at you. Actually, until October last year, the people would just laugh at you if you were to say something like that. Somehow, especially in the West, I'd say there was like more innovation coming from the East. It's an interesting paper from China. China actually, we saw developments with DeepSeek that shown to the world. That was the most interesting thing for me and that DeepSeek moment was that they showed that you can actually mess up with training. You know, you can do something in architecture. So this was fun. It was sociological in a way. And of course, then this goes along with investors. You need to find investors who are, you know, who have this first principles type of thinking, see the market kind of far ahead and place their bets. I mean, we're very lucky to get the cap table. That's just extremely, extremely supportive.

**[12:48] Zuzanna Stamirowska:** And so this is how we got on this journey. And I mean, I think I'll leave it to ask questions.

**[12:52] Jason Hiner:** Yeah.

**[12:53] Zuzanna Stamirowska:** I went on a monologue.

**[12:55] Jason Hiner:** Yeah, yeah. No, that was good. So when I met you, one of the things that you had explained was that today's LLMs, today's models, it's miraculous how far they've gone based on the limitations that there are, as you mentioned before. That they have kind of exceeded expectations in terms of how far the model could go. And some of these things are led to, like you mentioned, the reasoning models was a leap that sort of extended some of those possibilities. But could you talk for a minute? You mentioned the limitations because fundamentally what you're, you know, trying to do, Pathway, what you and your team at Pathway are trying to do is trying to build the next model after LLMs. That sort of exceeds the capabilities of LLMs and transcends the limitations of LLMs fundamentally. And so maybe just for those in the audience who don't understand what the limitations are, they probably are going to recognize some of them. But why don't you just sum those up for me, like the fundamental limitations of LLMs that you all are trying to supersede?

**[14:12] Zuzanna Stamirowska:** Absolutely. So we are building a post-transformer model based on that, I mean, based on the post-transformer architecture. So we're literally changing the T in GPT, if you wish. And our architecture is called BDH, Dragon Hatchling, kind of sparking a lot of cool memes with dragons around the internet. And the limitations that we have in Transformer, and this is something you all know and you all experience when, like, I'm talking about the audience, I mean, is that whenever you have a model, it's trained once. So you get the model, and it has a serial number, right? It has a name and a serial number. Because when you train an LLM, you feed it with data, and you create something that's called the weights. That's, let's say, you create the brain of the model. But they are created once. It's a bit like Groundhog Day. I mean, then you have that head, let's say, it's created. And then as you use it, you feed in information that goes into context. But this context is ephemeral.

**[15:20] Zuzanna Stamirowska:** It's like writing something on a piece of paper and just storing it, you know, the notebook without ever internalizing this knowledge. There is no link between the queries, like, between whatever you put in your LLM or how it's being used and how the world evolves. And its weights were set once at training. And hence, we need all those huge, gigantic training runs, you know, every time a new model release is coming. That's why we need to burn all this compute. Because you need to train them once. So this is one. They have, like, no way to transfer experiences, questions, you know, the things that you feed into them from now, from context. Their weights are kind of long-term. We call it parametric memory. And they suffer from something that we call catastrophic forgetting, actually. If you try to do it, there is some sort of, like, an effect that makes it forget the previous things that they've learned. And this is not something that happens to humans.

**[16:23] Zuzanna Stamirowska:** So if you guys have seen the movie Memento or The Groundhog Day, you know, every morning the main kind of character wakes up without, like, short-term memory, without remembering what happened. And then they start leaving sticky notes or tattoos, right, to themselves to kind of know what happened. Right now, the memory that we experience in products, in AI products, works like those tattoos or sticky notes. And then, of course, it gets larger and gets better. Because we do a lot of engineering around it, right? I mean, we, as a community, there is so much engineering going into it. So we're building large databases. In fact, it's like a database. So you just store things in the database. You have agentic systems that just compress and push around pieces of context, information about your interactions with your model around. And then it gives us an impression of memory. But it's not learning in the way we learn, right?

**[17:27] Zuzanna Stamirowska:** It's a bit like a student who would come to an open book exam without ever having read those books and, you know, without ever having understood what would really happen there. Unless you internalize knowledge, you can't get further. You're not updating or building frameworks for reasoning that would ultimately allow you to solve new problems. And the second thing is that transformers. So there's the fundamental memory problems in LLM. It's just true. So they can't, we call it continual learning. They cannot learn continually. They cannot learn with experience. And they have trouble dealing with the notion of time. So memory kind of has to happen over time, right? So they have, they're, they're a static. Imagine they're a static trained ones. And then, yes, there's a lot of engineering. There is an, you know, entire world doing engineering around LLM right now.

**[18:24] Jason Hiner:** And by that, you mean, Zuzanna, like the fact that you can create projects and you have sort of some static things that every time you ask a question in that project, it always accesses the same information so that it will try to give you, it sort of will imitate memory a bit. Or when you can create a profile about yourself in there and you, you know, you say what you are, what your job role is and background about you. And it's sort of, it's overcoming some of those memory gaps by like reading those files every time that you ask a question. That's what, those are like the workarounds to this memory problem.

**[19:00] Zuzanna Stamirowska:** But the difference is, you know, the difference is between your, your own family doctor who knows you, you've known your mom, maybe knows your kids and knows your entire history where you live and how you eat and, you know, how you party. And having, having, having somebody who sees you like a different doctor at every time who reads your file, probably just a recap of your file, right?

**[19:23] Jason Hiner:** Hmm. Okay. Okay.

**[19:25] Zuzanna Stamirowska:** That's, that's the difference. That's the difference between having internalized knowledge versus, versus a recap.

**[19:33] Jason Hiner:** Yeah. Okay. So that's the first one is the memory problem. And then you were just about to say the second one.

**[19:39] Zuzanna Stamirowska:** The second one is with reasoning. Right now, the way we do chain of thought. So we're moving from pure LLMs, we're moving to the chain of thought system. So that happened with o1 and o3. In chain of thought, reasoning happens in language. So thoughts have to be expressed in language and the model has to work through the problems in language, verbalizing everything.

**[20:03] Jason Hiner:** Yeah.

**[20:06] Zuzanna Stamirowska:** Try to ask a chess player to verbalize his thought process, you know, or her thought process while doing that. It just, it just wouldn't work. That's not how thinking works. Which is something that we actually learned, we discovered, we were hoping when we started at Pathway, we were hoping that language would get us further with reasoning. But it appears that language already imposes too much structure or too many limits on reasoning itself. And especially if you, you know, when you're a chess player, you kind of see a number of possibilities at the same time in your head. You know which ones are less probable. So you can feel it and you have a certain shortcuts. I mean, sometimes you find yourself in a new situation, but you find an analogy to something else. So over time, as you work through problems, you create some shortcuts for reasoning that you can actually use to treat situations that you haven't seen before. And Transformers as such are terrible at games. Chess is one example, right? Okay. There are Atari games.

**[21:11] Zuzanna Stamirowska:** There's an entire lab at NYU, in fact, that deals with games. And they're just, I mean, they are terrible at this. Because of this need to process everything sequentially and having to like verbalize things. Okay. So what is needed is this actually freedom for the model, you know, to think in abstract thoughts like a chess player. And this is something that we managed to get. So we managed to get to a place where we have models thinking, us being pathway with BDH. We have models thinking in abstract space without verbalizing everything and the chain of thought.

**[21:54] Jason Hiner:** So what is BDH? Is BDH an acronym? What does it stand for?

**[21:58] Zuzanna Stamirowska:** It is. It is. So this is an acronym. So our architecture was, so BDH is a post transformer architecture that was launched in October last year and the paper. So this is why I was telling you that, you know, until October last year, it was very weird. Because somehow around our paper and a couple of other papers and all of a sudden it became okay to say that, listen, we're reaching the limits. And BDH, the acronym stands for Dragon Hatchling. And the B honestly was added just as a letter for a three-letter acronym because it sounded better.

**[22:37] Jason Hiner:** Okay.

**[22:38] Zuzanna Stamirowska:** Lukasz Kaiser says that B should stand for beautiful because of how beautiful dragon hatchling.

**[22:44] Jason Hiner:** All right. Beautiful.

**[22:45] Zuzanna Stamirowska:** The math of this paper is pretty beautiful.

**[22:49] Jason Hiner:** Yeah.

**[22:50] Zuzanna Stamirowska:** So this is, let's say, the basic, the very basic architecture that we published that already gives a number of cool features. I mean, first of all, it gives native memory. So as I just explained, having a transformer in LLM, there is no memory. There we get memory natively, you know, in the learning process itself. So this is there. And then it has a number of nice implications for interpretability and safety. But I guess we'll get into this.

**[23:18] Jason Hiner:** Ah, yes. Okay. So why Dragon Hatchling? Why did you name it Dragon Hatchling?

**[23:23] Zuzanna Stamirowska:** Do you know Terry Pratchett?

**[23:24] Jason Hiner:** A little, a little. My wife does. My wife reads Terry Pratchett.

**[23:29] Zuzanna Stamirowska:** I will send you the book. I will send you the book or at least the part, you know, where it all happens. So in The Color of Magic by Terry Pratchett, there is my favorite character who's called Two Flower. And he's a tourist who just visits this kind of magical continent. And like he's visiting the Discworld, like traveling around. And all of a sudden he was captured and he sits somewhere in a cell. He's kind of locked up and all of that. It's very dangerous. And he thinks about the dragon and then the dragon appears. So, and then he has a very cool discussion with the dragon. Like, how did you, how did you happen to be here? Well, you thought of me and I thought of you. So here you are. Yes. So I am yours to command. The point is, these are dragons that appear the more you think of them. And then there's a funny, funny line that says something of a sort that in this place, the veil between thought and reality is somehow blurred.

**[24:35] Zuzanna Stamirowska:** And something that we found, and that was a strong thesis we had as we were starting this, this, like to work on, on this project, was that thoughts, correct memories. That thoughts themselves have actually an impact on, on, on, on, on memories. And yeah, so this, this is, let's say this, it's, it's more than just a cool, you know, fantasy reference. It's really about, about how thought can, can shape, uh, what we remember and then ultimately what happens.

**[25:10] Jason Hiner:** Wow. I love that. That's one of my favorite stories for a naming of a product that I have ever heard. That's, uh, so cool. And so thoughtful.

**[25:18] Zuzanna Stamirowska:** Yeah. But actually it makes me come back to, to what you were asking before the limits, right? We had the memory, memory problem and latent, latent thinking. So latent thinking is the abstract thinking. Thinking is something that, you know, is freed of language. And all of a sudden a thought can be a concept. It can be something that multilingual people experience. I think pretty often they feel that, you know, they can't really put like a finger on that word, like what the word is, but they know there is a concept there. Um, when you put them together, you actually see that memory and time can feed the thought process.

**[25:54] Jason Hiner:** Hmm. Hmm.

**[25:56] Zuzanna Stamirowska:** And I mean, hopefully, of course, like the path forwards for the entire fields is to, is, is to get to AGI. I mean, you call it however, like, but we, we want to build autonomous systems. Uh, and you kind of need to have this dimension of time and memory feeding your thought process and thought process, having fact interaction with, with what you remember to build those reasoning frameworks. To build reasoning shortcuts, such as when you play chess with someone, you know, I mean, I often ask them kind of, why did you make this move? Because I'm learning. And very often they kind of just, they can't really explain it. It's like, I just knew this one made sense.

**[26:40] Jason Hiner:** Man. So fascinating. I have so many questions, um, for you. So I'm going to try to, I'm going to try to cue them up, uh, and, and get to them. And I'm also thinking about what the audience, you know, I'm sure the audience, as they're listening to this, they have lots of questions, you know, too. But so you started with the paper, October of 2025, you released your paper and you said that the paper was well received. There were some other ones that happened around the same time that really challenged the idea of LLMs being the, the path to the, to AI, the future of AI. That, that because of these limitations, the AI that we're building needs to transcend, um, that the product you're building is dragon hatchling or BDH. Um, is the product released? How can people or when will people be able to start experiencing, you know, the difference for themselves of what this looks like compared to, you know, what they've been used to using, um, with the chat bots of today.

**[27:40] Zuzanna Stamirowska:** So to answer this question, I'll go back to the dogma of differentiation and to creating moats. And I think this is, if, if, if, if we're talking, uh, maybe in the audience, we have some startup community, you know, what makes sense to like to build right now. What does it make sense to build right now? So I think it's extremely important to have a very deep moat and the more time we can spend on building the moat, the better it is for us in our case.

**[28:11] Jason Hiner:** Okay.

**[28:12] Zuzanna Stamirowska:** So we are still focused on the architecture. What we, what we showed, showed to the world is the architecture and we are training models based, based on it. Uh, making them shine in places where transformers somehow struggle. So not that we were talking about memory and reasoning and costs.

**[28:35] Jason Hiner:** Hmm.

**[28:36] Zuzanna Stamirowska:** Um, we will be releasing them and you will be seeing the journey. We will be releasing, you know, benchmarks and the science as we go. This is also part of why, why the naming after dragons. I mean, this is a magical creature that has magical properties, even at small scales, um, knowing that at our lab, we already have stuff, you know, working at 500 million parameters. Um, but, but as you work through the moat-building, you want to work the fastest and you want to see signal at smaller scales. So we mostly, we mostly work on the architecture and still, uh, well, work on the architecture. This is pretty locked, but let's say doing some stuff around it still to, to, to, to make the moat deeper. So the products will come. We have a great partnership with AWS. So the moment, uh, the moment we press start, it's going to be available there. And we are lucky to have a number of design partners, um, right now with whom we're kind of, you know, mostly working around the use cases, uh, in long horizon reasoning.

**[29:47] Jason Hiner:** Okay. So that's where agents come in.

**[29:51] Zuzanna Stamirowska:** This is something that will come and I will not, you know, commit to a date, uh, right now. Um, but what's there is that along with the paper, um, we published a repo with the, let's say the most vanilla version of the BDH architecture.

**[30:09] Jason Hiner:** Okay.

**[30:10] Zuzanna Stamirowska:** It was, it was fun. It was fun. Many, mostly for the purpose of, uh, allowing researchers to replicate, uh, the results of the paper. And many people actually started experimenting with it. This is, that one is not yet the version that will take you all the way to solving, you know, Sudoku at 97.4% accuracy and all of that. Um, but it's, it's the one that shows the theory and shows the link between the Transformer and models of the brain. Um, and this link we've shown formally in the paper. So we actually launched this paper to give this axiomatic, sorry, to give the science of AI to the AI community.

**[31:00] Jason Hiner:** Okay.

**[31:01] Zuzanna Stamirowska:** So products are coming, don't worry.

**[31:04] Jason Hiner:** Okay.

**[31:04] Zuzanna Stamirowska:** And actually pretty impressive. Uh, we'll be, we'll be announcing kind of things as, as, as they come, as we, as we kind of feel comfortable sharing. But there is stuff that many people play around with already. And there's also many materials about how you can think about basic mechanisms or like main mechanisms of how AI works, like attention, how attention works in BDH. There are talks by our CTO. Um, so I would invite you to have a look at our YouTube channel, um, for that. And yes, there is the GitHub repo and folks replicating the results, which is a lot of fun. I mean, we love seeing that.

**[31:39] Jason Hiner:** A couple of the other things that come up with LLMs often, uh, you mentioned one of them, but one of course is hallucinations. The other is safety that you can, um, you know, you can do prompt injection. You can do various ways to, to compromise jailbreak, however you want to call it, you know, the models themselves. And so is the work that you're doing and that you, you and your team are, or what you and your team are building, you know, a pathway. Are you working on those problems as well? Or are, are those problems, you know, will continue to be there?

**[32:14] Zuzanna Stamirowska:** Well, they will continue to be there to an extent, to an extent that we as humans hallucinate. I do quite a lot.

**[32:22] Jason Hiner:** Sure. Fair.

**[32:24] Zuzanna Stamirowska:** But, um, a lot of hallucinations come from, uh, all the problem of context and generalization over time. So you will ask, you kind of want to increase the window of focus for, for the model. Like, are we dealing just with the memento day or, you know, does the person remember all the past and then can have an intelligent conversation with you, knowing you know, in your context and knowing everything else. Um, so as we kind of hit this generalization over time, then a number of hallucinations will just decrease dramatically. So of course, with enhanced reasoning and this notion of time, we will reduce hallucinations to something that will be probably more likely for a human.

**[33:13] Jason Hiner:** Right.

**[33:14] Zuzanna Stamirowska:** Um, safety, uh, I think is critical. We solve it, we solve it from, let's say for now from, uh, or we contribute to solving it. I would put it this way, uh, for having understanding of how such a system works. Cause right now, this is another limitation that we didn't mention right now we train the models. They perform on benchmarks. We see the steam engine, you know, showing steam good. Uh, but we don't really know why. So folks are building like the equivalence of MRI machines to scan the brain or the engine. Okay. To scan it, to try to see what happens there and then explain, okay, that was, you know, the grandmother cell. Uh, this is it. The Golden Gate Bridge was here. Um, and kind of pop up champagne every time we found something by scanning the brain and trying to sort kind of through what happened there. Um, with, with, with a more theoretical approach, we actually see concepts emerge. And in BDH, this is what we show in the paper.

**[34:22] Zuzanna Stamirowska:** You can literally locate the synapses that encode specific concepts that fire up when a concept like, or the currency is mentioned. Um, and since we have this micro to macro passage from small interactions between neurons to the global state and, you know, to seeing a thought, you know, the model doing something. Um, we can predict that the same laws will hold at different scales of size and time. Um, and this is something we can't really do for transformers right now. We, we can kind of, you know, assert what you're doing on a given benchmark. We can, we know we've seen them here. We know how they behave, but as we move outside of the tested conditions, it's difficult to say. And with, with like, it's like having, you know, having some basic laws of how physics works versus not having them.

**[35:16] Jason Hiner:** Sure.

**[35:16] Zuzanna Stamirowska:** We have those laws, so we have a greater, I'd say, uh, maybe certainty about what will happen at different scales. I mean, we know the same, the same interactions, like the, the local rules of neurons interacting, passing on, uh, passing on information, reinforcing themselves. I mean, the equations of reasoning will hold no matter the scale. Uh, and this is, this is actually, that's, I'd say for now, our answer to the safety questions in part, why we embarked on this journey. Um, because we believe that it is somehow difficult to launch systems that we fundamentally don't know how they work. And put them out there, um, hoping for the best.

**[35:59] Jason Hiner:** Mm. So that improves observability. There's this word observability that comes up a lot, you know, trying to understand.

**[36:06] Zuzanna Stamirowska:** Yeah. Interpretability, I'd say.

**[36:08] Jason Hiner:** Interpretability. Okay.

**[36:09] Zuzanna Stamirowska:** This is, this is, this is one, this is what happened like after, but also as you, you know, as you design new experiments, as you can, uh, think about how your model will work at the new scale. What can happen? I mean, this, this helps you to predict things that you haven't tested yet. Okay. It's like, you know, like scientists build something in space. I mean, they're not there in space to make the experiment right before building, but they, but they can plan it very, very accurately.

**[36:37] Jason Hiner:** Right. They can model it. They can, you know, anticipate all of the, the variables and then they can build something based on those variables.

**[36:46] Zuzanna Stamirowska:** Exactly. So you can, you can simulate stuff better.

**[36:49] Jason Hiner:** Okay.

**[36:50] Zuzanna Stamirowska:** Um, and this is, this is something that we believe that, uh, fun, actually we have a piece of content that we were writing about this. We believe systems based on BDH will have a greater chance of avoiding, uh, the paperclip factory problem. So of them going off rail, you know, even, even set on a good goal, right. Them going off rail just because they kind of thought something was okay.

**[37:16] Jason Hiner:** That's the paperclip. Yeah. Explain the paperclip, um, phenomenon.

**[37:20] Zuzanna Stamirowska:** Uh, the paperclip phenomenon is that, I mean, even if you have, if you have a model set on a, on a positive or neutral goal, uh, it may all of a sudden, uh, decide to use all the resources of the earth to just turn it into a huge paperclip factory. You know, when its goal was maybe just to produce a paperclip. Um, so.

**[37:39] Jason Hiner:** Gotcha. You, okay.

**[37:40] Zuzanna Stamirowska:** You would like, we would like to get to systems that at least we know, even if we put, you know, if we put them to a goal that's reasonable, it won't, it won't do something crazy. It's a bit like when you're interviewing a candidate, uh, for a role at your organization, you know, they graduated from, I know Harvard. Uh, you observe them for a couple of hours. Your team talked to them. Uh, they said that they liked them. Right. And you expect that when you put them on the job, well, they will just, I mean, they may make some mistakes, but in general they're predictable. Right. They will behave like a human with this, this resume kind of with this psychological profile, unless they have, um, unfortunate case of a mental illness. So for LLMs, we can't really know if they have a mental illness because we wouldn't even be able to define it. I mean, okay. Right now. And the thing is that with, with, with systems that we understand, we have a greater certainty, um, that, that they will, they will behave in a more predictable way.

**[38:46] Jason Hiner:** Okay.

**[38:47] Zuzanna Stamirowska:** And then setting them to a bad goal or, you know, prompt injection attacks and all of that. Uh, this is, let's say not to be addressed exactly at this level.

**[38:55] Jason Hiner:** But, but you also have said that, um, and you brought it up a little bit earlier too, that the way that you're building the moatls is also going to make them infinitely cheaper and use less compute. Maybe infinitely, I might be exaggerating, but.

**[39:11] Zuzanna Stamirowska:** Infinitely maybe, but it's, I'd say it, it's, it, it, it is very impressive.

**[39:17] Jason Hiner:** Okay.

**[39:18] Zuzanna Stamirowska:** Just how big of models we are able to train with, with, with how little compute this is, this is one. And then also how extremely inexpensive they are, they are to run. Um, so the first thing comes from the principles of BDH. So BDH gets, uh, got its inspiration from what biology got right. Let's say.

**[39:40] Jason Hiner:** Okay.

**[39:41] Zuzanna Stamirowska:** Uh, so we have local interactions between neurons. So not every neuron, not the entire brain fires up every time you have a question, you know, you just fire up a couple of neurons who are connected. Um, so you have sparsity, we have, we have a number of things that actually we do observe in the brain. Uh, but we use those very simple kind of principles, apply them in BDH. Um, and this leads to just general computational efficiency that we have and the capacity to store a lot of information, like a lot of memory to store it directly inside of the model. Uh, so this is fun, just a smart kind of smart, efficient data structure. Um, the second thing is that since we don't need to do chain of thought, because in our models, reasoning happens in this abstract space. Uh, we don't need to verbalize every step. It doesn't explode. It's actually, it's extremely efficient. So the Sudoku result that I mentioned, uh, that was 97.4%.

**[40:46] Zuzanna Stamirowska:** It was like state of the art when, when, when, when it happened and just, you know, LLLMs were at 15, I think, percent. Uh, I mean, this, this happens one shot without chain of thought immediately. It's, it's, it's, it's just done. It's.

**[41:03] Jason Hiner:** Wow.

**[41:05] Zuzanna Stamirowska:** This is, you know, a nice example, which is a nice illustration because everybody knows what Sudoku is. But in fact, there are like any constraint satisfaction problems, like planning your trips with your fuzzy constraints that, you know, maybe your model after a while should know that you like your coffee at eight and you like it with milk. And, uh, you know, and maybe, maybe that, that, that latte place in Rome is closed on that day. And there are so many constraints that can come from context and that would make your planning easier. It's like, I don't know if you have an admin, but having an admin that's known you again, it's a bit like the doctor who known you.

**[41:44] Jason Hiner:** Sure. Sure.

**[41:45] Zuzanna Stamirowska:** Um, so the costs, uh, the difference in costs is dramatic. And the second thing is that training of the models is, uh, way less expensive. They don't need to ingest as much data. In fact, they get bored when they see something for too many times. Uh, something we already have in the BDH paper, we see neurons getting bored. Like when they, when they hear the same thing for, for, for let's say another time, they just, their activity just goes down. So we can observe it to actually see what the neurons are doing and we see they're getting bored. Um, and then everything sits in memory. So we don't need to, since we have state, so point is, let's say context sits inside of the model. So we don't need to just in the engineering sense, go outside, uh, of, uh, of, of, of, of, of, of, of, of, to, to, to, to fetch data, uh, elsewhere. So this reduces costs as well.

**[42:42] Jason Hiner:** So you've just made it dramatically more efficient with the way that you've architected the way the model runs.

**[42:48] Zuzanna Stamirowska:** Yes. It's like works like a brain on GPU.

**[42:52] Jason Hiner:** Interesting. Like a brain directly on the GPU. It makes sense in one sense too, because we know that the amount of power or energy that the human brain uses to process a question is so infinitesimal compared to what an LLM does.

**[43:08] Zuzanna Stamirowska:** Exactly.

**[43:09] Jason Hiner:** Yeah. Yeah. Okay. Very good. Well, Zuzanna, I want to ask you, how did you end up on this journey? You know, what, what is your background? How did you end up working in this sort of set of problems and how does some of the things that, uh, you know, you learned along the way in your academic background and, and the projects you work on before sort of lead you to, to this moment, you know, where you're, you're working on pathway.

**[43:34] Zuzanna Stamirowska:** I was a researcher at the Paris Institute of Complex Systems. Um, and so complex systems is the, let's say a funny discipline that merges many different disciplines in one, uh, mostly looking at how interaction between small particles or small fundamental elements, let's say, uh, give rise to global phenomena. And this can be anything. These can be atoms, you know, uh, this can be, this can be water. This can be people interacting on a social network point. Yes. We have some units. There is something that happens in between them. And as we have many in betweens, something happens like information spreads and you have a rumor that starts, you know, forming this may, this may change the result, uh, the result of elections, for example.

**[44:24] Jason Hiner:** Right.

**[44:24] Zuzanna Stamirowska:** So this is some of some of the topics that my, my colleagues were setting. Um, it actually draws a lot from statistical physics. So actually most of the people have background or like that, and then complex systems have background in physics. Uh, I was working on the intersection of, uh, theoretical computer science, um, economics and, and physics. I actually got, got into this, uh, through game theory. Cause my, uh, my dad loved the movie, uh, the beautiful mind.

**[44:58] Jason Hiner:** Ah, yeah. Yeah.

**[44:59] Zuzanna Stamirowska:** And I took a class in game theory once. At Stockholm School of Economics. I fell in love with it. So I decided I need to do it, uh, for the rest of my life somehow. And then I discovered graphs actually around the same time I'm working at SPOJ, which was the first and largest community for competitive programmers. And it was created by people who worked on data structures and graphs. So I learned a lot about graphs and networks back then. And then I started to do game theory on graphs and then having those graphs evolve. So you kind of don't know what's happening anymore because everything keeps on changes. And the only constant in life is change. Um, and this is how I got to, to developing a very, very deep sense and intuition about how, how complex things that happen can be in fact reduced to small number of rules. And maybe fields or attractors that you place somewhere that can shift the behavior, but you can, you can somehow not maybe solve the complexity or remove it, but you can put it in some sort of frames.

**[46:06] Zuzanna Stamirowska:** You can harness it somehow. Um, and funnily enough, I mean, we're working already somehow on this project, but, uh, I think this is a story I told when we met. Uh, I was sitting with a friend who was a neuroscientist, like a wonderful man. So we're sitting there at a cafe in Paris. Imagine, you know, beautiful Paris around Saint-Michel, uh, we're there talking about different types of, you know, of, of coffee and then like porcelain, fine porcelain and all of that. So we're kind of chatting. It's, it's, it's, it's, it's, it's beautiful. And then all of a sudden he tells me that there was an experiment when somebody swapped the auditory and visual nerves in a mouse. And you know what, like the mouse turned out just fine. And I sit there, I look at him and I say, yes, because function shapes the network.

**[46:57] Jason Hiner:** Hmm.

**[46:57] Zuzanna Stamirowska:** And then he looked at me and he said, yes, exactly. And then I stood up and I just left him. Uh, I went to, to, to our office and I remember I, that was the moment, you know, we were kind of already working on it, but it was a moment of a very, very clear conviction. So if this, if, if we see that the level of nerves, you know, it wasn't the moment of decision, but I remember this very deep conviction that yes, of course, function shapes network. And then we worked through, you know, we have to work, we had to go through the journey, but I made a number of, a number of bets with my, with my colleagues, with Adrian, uh, and Jan, and I won each and every one of them. Uh, uh, so funnily enough, I did my, I did my PhD thesis. Um, and it was published by the National Academy of Sciences, um, on forecasting of maritime trade. Uh, I did forecasting of maritime trade because I had a great database. So I could actually see having 30 years of daily movements of ships around the globe.

**[48:03] Zuzanna Stamirowska:** This is a beautiful data set of a complex systems coming alive, complex system coming alive. Um, and what we found is that if you look very closely at attention and BDH and how neurons pass on information, it actually behaves a bit like a transportation system. That's the easiest intuition you could have.

**[48:25] Jason Hiner:** Okay.

**[48:26] Zuzanna Stamirowska:** The routing of information and the routes that are being used more just become wider.

**[48:32] Jason Hiner:** Interesting. Okay.

**[48:33] Zuzanna Stamirowska:** Uh, so this is like funny enough. You do find same, same dynamics. Just forget the system, the actual system that you apply it to, but you, you tend to find the same dynamics. And, funny enough, we found that the function, the functions we see in models are somehow like somehow similar to, to, to the transportation ones. It's not a direct equivalence, but it's a nice way to explain it.

**[49:00] Jason Hiner:** Cool. Very cool. Wow. Okay. All right. I want to ask you, you know, as we wrap it up, I want to ask you the same questions that I ask everybody right now, because there's the same two questions that, um, that I hear all the time in the AI industry. And one is, you know, AI was the promise of AI was that it was going to do, I was going to automate away a lot of the hard things, you know, the sort of the grunt work and, and those kinds of things for us and make us more efficient and essentially give us more time. But most, the reality is most of the people right now that I see, um, they feel like their time is stretched even more, you know, than maybe because of, uh, they have more capabilities at their fingertips or whatever the case may be. They see sort of these new opportunities and they can do, you know, more things. So because of that leaders, um, all of the leaders I know right now are very conscious of like, how do they get maximum leverage for their time? How do they optimize, you know, their time for, for maximum leverage?

**[49:59] Jason Hiner:** So I'd love to know, you know, from you, what is your, um, best tip? Uh, what are the thing that you've learned? What is it that you do to maximize, uh, your time right now to get maximum leverage for your time? And, you know, that you might recommend to, to others. What's the, what's your best tip?

**[50:17] Zuzanna Stamirowska:** I'd say in general and management and running the company, everything changed so much, right? So the only constant in life is change. It's never been as real as right now.

**[50:28] Jason Hiner:** Yeah.

**[50:29] Zuzanna Stamirowska:** Um, what I do and I do it every day and, and I do it before every meeting and every discussion is that we, we, I, and I'm us as a team, we analyze the critical path. And it may change because you discover new blockers, right? But be ruthless about the critical path. And I have a culture about critical path that, okay, things have to be dropped. You know, there's no, this is my, my area. This is yours. No, critical path is, is a religion. Um, so this is one. Second is that we stopped coding. It's very weird to have a team full of competitive programmers, people who were, for example, coding in Rust when Rust went out. Like, you know, that, that level.

**[51:12] Jason Hiner:** Yeah.

**[51:13] Zuzanna Stamirowska:** Who don't code anymore. It's very weird. It is a transition. It was a huge boost in productivity and it wasn't possible, you know, it was, let's say enabled this winter. Um, but we stopped coding and it means you have to actually somehow make a team change their habits of working. And since all of a sudden you're producing five design docs a day instead of one design doc for a month, you know. Or, or, or like whatever. Uh, the quality of design docs doesn't need to be this good. Hopefully it's really good. But you know, it doesn't need to be this good because the cost of producing code is not, it's not that huge. But then also your quality, uh, expectations from the code shouldn't be the same as what you've been expecting before. Uh, so there's a lot of change management, but going through it, it's, and a boost is just incredible. Uh, I hear people, you know, pretty much being frightened by their, uh, token budgets exploding.

**[52:22] Jason Hiner:** Right.

**[52:22] Zuzanna Stamirowska:** And we're just throwing into AI everything. I think indeed, first of all, we'll just make it cheaper. Don't worry.

**[52:29] Jason Hiner:** Uh, you're working on that.

**[52:32] Zuzanna Stamirowska:** That's going to be fine. That's a solved problem. Um, but second is, um, it requires different, different working habits. Actually, you may work for like, you know, instead of whatever people may, might have been doing enterprise. And I'm a startup founder. I work all the time. So I, I'm not a regular person. It's not normal. I know, but you can't really work from like, you know, nine to five, but perhaps our workdays should be rethought, you know, maybe we should work in split sessions. We have one deep-thought, conceptual session. We launch, uh, we launched the, we launched the tasks for, for our codex or, or Claude, uh, then we come back, right. And we go, we enjoy time with, you know, friends or we do yoga or whatever. Uh, then we come back. So we have breaks, but we work until the evening. And then of course, because we're also, what you want to optimize is the use of your hardware.

**[53:31] Jason Hiner:** Sure.

**[53:32] Zuzanna Stamirowska:** And this is, this is something that we, we've seen as a, as a very huge change, uh, in the team. And yes, we collapsed, you know, roadmaps that a year ago would have taken many more people and maybe a year, to two weeks. I would say.

**[53:48] Jason Hiner:** Very good. Okay. And then the last question that I ask everybody is on AI tools and you've mentioned a couple, but maybe if there's any other ones as well, of what are the AI tools that right now you are using, um, that maybe people aren't aware of, or using them a way that they might not be aware of. That, uh, are making a huge difference for you, a huge impact, you know, for you, um, and that you would recommend, you know, people consider.

**[54:13] Zuzanna Stamirowska:** I mean, get the strongest subscription to the strongest coding models out there. Um, but the color that I can add to it is that we see that folks who are doing the most cutting-edge work at Pathway who are competitive programmers. They prefer Codex. Uh, and also we kind of know how it was trained. So it's not weird. Actually the research team at OpenAI comes from competitive programming, uh, backgrounds. So competitive programmers go with Codex. Folks who write, let's say more traceable, regular code, they definitely prefer Claude. So this is, this is, this is one fun thing to know.

**[54:51] Jason Hiner:** Okay.

**[54:52] Zuzanna Stamirowska:** Uh, but definitely worth it getting, getting the best subscriptions. It's, it's, it's, it's, it's incredible. Fun thing that I just noticed I'm still experimenting with, uh, you guys may, may know of, um, sorry, inception. Well, sorry. Inflection AI. I'll, I'll do it again.

**[55:10] Jason Hiner:** Okay.

**[55:10] Zuzanna Stamirowska:** So another cool thing that I'm just experimenting with, uh, is a model from Inflection AI, which is good at emotional intelligence. So when I write, you know, pieces of content, I wanted to have a better voice and sound a bit more like me. And I'm starting to have pretty interesting results with that one. I haven't done tests. Like, could I really find a good prompt for, you know, for, for GPT or, uh, or Claude to, to get me to the same level of emotional color. Uh, but so far I'm, I'm kind of pretty happy. So this is at least a thing, fun avenue for experimentation.

**[55:54] Jason Hiner:** Very good. Those are great tips. Um, well, Zuzanna, thank you so much for being on the podcast. Uh, it's been such a great conversation. I knew it would be, you know, from the, the time we, we met in San Francisco and I learned a little bit about what we were, you were doing, but, um, but this has been amazing to really unpack, you know, all of these things. And, um, best of luck in, in, uh, hatching your, your dragon hatchling, um, later, uh, this year or whenever the, uh, the occasion comes.

**[56:25] Zuzanna Stamirowska:** You know, it's big.

**[56:27] Jason Hiner:** It's getting big. It's good. It's good. No, very exciting. Um, thank you so much for being here.

**[56:33] Zuzanna Stamirowska:** Thank you. Thank you. It was lovely. And great seeing you.

**[56:36] Jason Hiner:** Thank you.