Building a More Efficient AI Company with Pathway

[00:00:00] Once a year, I open up a handful of places for people who are serious about launching their own podcast, but have never actually gotten around to it. And over the last 11 years, I've helped entrepreneurs, business leaders and Fortune 500 companies all launch their own shows. And here is how it works. You record your conversation on Zoom, Teams or whatever works best for you.

[00:00:23] You send me the audio, an episode title and short description, and I will handle everything else. From setting up the podcast host and distribution to getting your show on Apple Podcasts, Spotify and the many, many other podcast platforms out there. I will commit to working alongside you to help you launch your first season of 12 episodes in the bag. So you can start 2027 by ticking a season of podcasts off your bucket list.

[00:00:52] Think of me as your not just personal podcast producer, but also your accountability partner to make sure that you actually deliver those 12 episodes into podcast land. Now, I only offer this once a year with limited places available. So if you want 2027 to be the year you finally launch that podcast, head over to techtalksnetwork.com. Send me a message, send me a DM or even an audio message. I'd love to help you make it happen.

[00:01:18] How can a startup compete in AI when the largest companies can spend more on computing power? Well, one answer is to question the very architecture itself. And today here on Startup Builders and Backers, I welcome back the co-founder and CEO of a company called Pathway. And today we're going to have a conversation about building a research company around that very bet.

[00:01:47] She would describe Pathway's work on AI beyond transformers and its ambition to make reasoning and learning much more efficient. And yet we will discuss the commercial decisions around that research, the investors who understand its timescales and the pressures that founders carry when the result is still being built.

[00:02:09] And we'll also discuss how a startup or any business should judge the value it receives from AI rather than counting the tokens that it consumes. And yet there is a big scientific ambition here alongside practical advice for founders who need to ship a product right now. But what do you build now and which assumptions about tomorrow deserve a closer look first?

[00:02:36] And on that note, let me introduce you to my guests. So a massive warm welcome back to the show. We last spoke, what, three years ago. But can you remind everyone listening with a little about who you are and what you do? So I'm Zana Stamirovska. I'm a co-founder and CEO of Pathway. And we are building the post-transformer AI.

[00:02:58] So if you think that, you know, the transformers like the math under the transformers in TNGPT enabled this revolution that we're seeing right now. Like this is literally the breakthrough that happened in people's heads, actually. You know, a couple of folks at Google Brain and Time invented the transformer. It opened up the market, showed everything that's possible. But it's a bit like the V1 of the steam engine. The one where we keep on burning coal by the ton. And there is, like, we definitely can do better.

[00:03:26] And what James Watt did for the industrial revolution is that he came up with an engine that was dramatically more efficient. The V2. And that was actually the one that got, you know, massively adopted and enabled the full industrial revolution. This is what we need to do. This is what needs to happen in AI. Even we do it or somebody else. Point is, it will happen. It will have to happen. And we're building that V2. That's it. This is a post-transformer AI.

[00:03:50] And this will be systems that are capable of actually reasoning, evolving, and adapting with time. There will be actually learning with experience, just a bit more like humans in a way more efficient way than we're already seeing right now. And it is an incredibly exciting time for you right now because you're doing all this at a time when most of the industry is investing enormous sums in scaling transformers.

[00:04:16] So I've got to ask, what was it that convinced you that improving the existing architecture alone just wasn't enough and that maybe a different approach was needed here? Everybody. Everybody. I mean, everybody. So it's very funny because the machine learning community is mostly a community that's driven by experimentation. So you do a small change. You see, you know, it makes it better on a certain benchmark. You're happy. You pop up your champagne because you saw. You saw your game.

[00:04:44] Of course, if you change an architecture, if you change something very fundamentally, you won't see it so immediately on the benchmarks. You actually, we actually took, you know, a theoretical approach from theoretical computer science to optimize that. So, I mean, but we were in 2024. The world was very different to what it is right now. We saw all the, you know, firsthand all the problems big enterprises had with managing context or like having the first, you know, AI applications. How would that work?

[00:05:10] Like people were barely putting their heads out of, you know, out of the chatbot. Right. It's like, would AI be more than chatbot? I mean, they were not even necessarily asking themselves this question. I mean, Transformer won because it kind of won that hardware lottery. It scaled very, very well on GPU. Yeah. Grammatically well for language. We could say even that it's like optimal for language. Like it's hardware optimal for language and GPU.

[00:05:38] But in 2024, O1 came out. O1, it was like the first reasoning model. And, you know, there were the papers about the fact that having a transformer with chain of thoughts is too incomplete. And there are proofs of that. So that was a huge moment. But if you look at reasoning, it's dramatically inefficient. It just consumes so much compute by definition.

[00:06:02] And in reasoning, as it is being done with like on top of the transformers, every word, like every path of your reasoning has to be verbalized as a word. Like it's as if you were playing chess. And instead of playing and looking at the board, you would be literally writing down every sentence. Like the rook is here. And I like that would be impossible. It's like this sort of thought as you're solving constraints, it doesn't work for language. It's very cumbersome.

[00:06:31] So this process, this computation is actually not very optimal. And very early on, we had strong conviction that it's reasoning that is the key to intelligence and not knowledge and language. Or like which one doesn't mean that knowledge and language are not important. But it's just that they are not the key for intelligence. And this is because if you want to have autonomy, it's reasoning that gives you autonomy.

[00:06:59] You should have some sort of patterns of thinking and you can adapt easily to new situations to learn new things on the fly. And this will go through reasoning more than knowledge base. And then, of course, transformers went very, very far with scale. And this is something that it was almost a religion for. I think it stopped being a religion this much a bit like a year ago. When actually finally never on the time when we published our first paper.

[00:07:25] It is an absolute religion that you would be mocked if you were doing something outside of the transformer paradigm. Because scale, scale, scale, scale. And then, you know, as we reach scale, it's going to be just awesome. And then, I mean, if you need smaller, then we'll distill. That was that with the main. The main even like sociological, you know, argument. You know, folks popped up. I think I was very happy when DeepSeek, first DeepSeek went out. Yeah.

[00:07:52] First DeepSeek went out and all of a sudden, oh my gosh, they didn't need to spend this much. Oh my gosh, they actually innovated around the kind of, you know, how you train. Oh, you can do that. Yes, you can. You should do that. It's smart. And, you know, it is difficult to change even the equity story. If pre-regnant story is just we need so much cash to kind of, you know, load compute and kind of. It's really good to have a lot of compute. I'm not saying no. No matter how efficient your architecture is really, right?

[00:08:20] Because if it's more efficient, you'll just get way more from that compute. So that's awesome. But it was literally amongst the researchers. It was difficult to, you know, at the party to say that, hey, how about we don't do scale thing? The people, the first people who moved beyond that were actually, who were very open to it were the Transformer co-authors. That was funny. They're like, so Heon Jones from Sakana, right? He created Sakana saying that Transformers, like, he's annoyed with the Transformers

[00:08:49] because they just like that all the room from me, all the air from the room, right? So we actually organized some cool events with those folks who are friends. And actually, but the post-Transformers, we have like Transformers versus post-Transformers boxing fights on an actual ring, you know, between literally like the biggest researchers, you know, who created the field. It's very funny. Yeah. So what convinced us, I mean, we kind of knew that there were many things that we knew.

[00:09:16] And we kind of had this strong conviction about kind of where AI was heading, that people are still wrapping their head around the chatbot. Whereas we were really so reasoning, we, you know, we saw enterprise kind of how enterprise deployments could look like. That I had like long, long, long time ago in 2023, I had a strong thesis that every data pipeline would be an LLM pipeline.

[00:09:43] Like, I think that the kids today, you know, we don't know what I'm saying. And I'm just talking about 2023, right? But it's true. Like right now you see it, it actually happened with agentic applications, right? It's effectively the realization of that thesis. Yeah. I don't know. Very long answer. What could be, you know, there are stuff that convinced us on a scientific level.

[00:10:10] So our, our Adrian, our chief scientific officer, like he, he actually went to, in 2019, he went to a summer school. It was like, I think 20 people in Corsica, but those people included Jan LeCun, Giorgio Parisi, and like this level of folks. And he became obsessed with the brain. So from, we were obsessed with the brain from like 2019 computationally, and then from kind of, you know, machinery perspective.

[00:10:38] And the brain is such a beautiful thing. Like, you know, around 20 watts, it consumes, has memory, learns over time, doesn't recall everything precisely, but just learns to adapt, right? Learns to be, to be thrown at a new, at a new situation, figure it out. And it's a physical network system that, you know, we, we literally can see, we can scan the brain and see how, how the things fire up. And we see that they are local. You have a signal that's being propagated literally through a graph, right? This is very nice.

[00:11:07] And at first it works, actually. It works. It just, I mean, you don't need to recreate the brain precisely with all the problems, like all the chemicals and whatnot. If you look just at the simple principles of neurons passing on messages, a bit like buses, you know, in between them, then you actually go very far. Yeah. Yeah. And before you join me on the podcast today, I was doing a little research on you and I was reading how one of your criticisms of Transformers was that you described them as disconnected thinking. I love that line.

[00:11:36] But what does that mean in practical terms and how does it contribute to greater compute requirements, energy consumption, and potentially disappointing AI or ROI as well, which is another hot topic right now? This comes at, we can say that thinking, reasoning was somehow bolted on top of the transformer. Right now it's being externalized. I'm not sure if this is what you're asking about. But so as I said before, you have to express every thought in language.

[00:12:06] Yeah. This is inefficient. Of course, like for visual thinking, this is very weird that we have to describe a board. Like some examples, we have Sudoku is a great example. And Sudoku was very, very difficult for LLMs until actually very recently. The simple Sudoku that your grandma can solve over a coffee was almost impossible for the things that are about to change humanity forever and cure cancer. Interesting. But this is literally kind of a structural thing for them.

[00:12:33] This was kind of solved in Astra, mostly because of some elements or partly so. But so this is like reasoning. Because of the constraints of language, reasoning is inefficient. That would be my answer. And this means that it's actually difficult to have a bridge between the thinking that happens in language that is informed by reading a lot of recipes.

[00:12:58] Because ultimately, as you read all the recipes for everything in the world, you're going to know a lot. You're going to know a lot of recipes. And the chance of you actually drawing the right recipe for something is kind of pretty high. So and potentially may by chance recombine two recipes and it's going to be awesome. So there, I'm not saying there is no value. There is value in scale as I started, you know, at the very beginning, even if you have a more efficient way of learning, the scale is still cool. But just how much do you need it?

[00:13:28] How much of it do you need? So right now, I'd say there are like the criticism for, you know, for like just how much compute we need. Yeah. I have two companies and there is not enough energy right now. Right now, if you, if I'm illicit by GPU, right. I was, I was just kind of had an offer to buy GPU chips. And it was difficult to find a place to put them. So the problem is not, not even the chip. The problem is a data center space.

[00:13:58] And it's not even data center space. It's data center with a covered energy supply. And folks are fighting over energy contracts, like literally having electricity. Right now we are constrained, like Anthropic is constrained by contracts on electricity, like ultimately. And this is, it tells us something, right? Because we're probably still very early in terms of how far AI could integrate it into our lives.

[00:14:28] Consumers are not paying for it much. Like the consumer market is totally not penetrated. It appears. Like if you look just like how many households actually do pay for AI, there are many free services that are being used, but to actually pay, you know, just the business models are such that, I mean, it's, it's difficult to have a margin on, on this in consumer. And it's like, you really, you really need efficiency.

[00:14:53] So funnily enough, I think that the, you know, the way more efficient AI will enable actually further consumer adoption. Even though it will start with enterprise. And then also in the, like, especially, especially like in the U S and in Europe, people started to be very aware about token consumption or like just how much they're spending for the ROI. Because sure, it's cool. Oh, we're getting it automatically. But why did finding this one little bag that maybe I didn't care about cost me 20K? Right.

[00:15:23] And that was a realization that funnily enough came pretty late because for some time folks were kind of token maxing. And I find it weird, but you know, whatever. Right. Maybe you have big organizations to manage them somehow, but, but also folks were doing that to encourage people to use AI, right. For like having change management, but that honestly, what's the cost? Maybe having a human do that would have been cheaper. Right.

[00:15:47] I'm being asked a lot by, by people about how do I measure if this makes sense? Right. Am I making the best choices, the best model for me, for the specific use case and actually measuring the value you're getting out of AI is non-trivial. And you actually should, like everybody should at least spend, you know, I don't know, an hour thinking about how to measure and probably constructing by coding their own success metric

[00:16:12] on like the ROI and calculator for a use case that they would, and that they would be able to convert different models. Do you need AI agents that you can trust? Well, with an AI data layer providing real time connection within your data platforms, you can trust your agents to provide accurate solutions. So scale your business by trusting your agentic AI accurately getting the work done for you. Trust its capabilities with Denodo.

[00:16:41] And you can do that by simply visiting denodo.com to learn more. I was going to say for startup founders that are listening here, I would imagine that creates almost an interesting strategic question when enormous incumbents control so much of the compute, the capital and foundation and the model market that we're talking about here. So for those people listening, where do you see opportunities for startups to compete through

[00:17:06] architecture efficiency or specialized research rather than attempting to outspend them, which is a battle that they'll never win? And this is it. So for us, it was a strong thesis question, right? So, I mean, I guess your question is about our startup, not necessarily the other startup that would be using us. Naturally, it wouldn't make sense. I don't think we could win, I guess. And it wouldn't make sense to start the fight against, you know, open AI and entropics on their own terms, which would be scale.

[00:17:34] So it had to be different and had to be more efficient to bring sufficiently strong difference to the market for this to be interesting, right? Because otherwise, which is yet another model. And I'd say it's in the market that's kind of already somewhat shaped up, right? It's difficult to see a potential building, building a huge company. So yes, this comes directly by building the V2 of this team engine. That's the opportunity for us.

[00:18:03] And this goes for supplying models that will be finally personalizable, that will have great data sample efficiency. That means they learn from way fewer data. That means that, you know, an organization can have their own model. I mean, in the Silicon Valley, we call them sovereign models. But by this, we only mean that this is my own, right? Not kind of outsourcing my intelligence, but I just have my own model. I guess elsewhere in the world, when people hear sovereign, they think national, right? That's not the case.

[00:18:32] I mean, sovereign just means for an organization, I have my own model and I'm not giving my intelligence to anyone. I'm not sharing my problems, right? This is kind of my sovereign model. So this AI sovereignty is important. Personalization is super important, especially as enterprises have their... It also holds for people, actually. But first of all, for enterprises, they have their own data that's highly valuable. Their own specific use cases. And their intelligence, first of all, should stay with them.

[00:19:01] They would like to draw intelligence out of it. And they need to create their mode. Because the situation is that if we have two players on the market who supply all of the intelligence, you submit all the prompts to them, right? Effectively, where are the modes of the companies who are using them? Everybody is a rapper. At the end of the day, it becomes so dependent on Entropic or OpenA, I don't know, those companies. And then if the use case becomes interesting enough, their lunch is being eaten.

[00:19:31] So I was chatting with one of the biggest automation companies. And they were like, it's like, I'm paying the executor and digging my own grave, right? Because I pay so much for the tokens right now. It's actually my business is fully dependent on that. Whereas actually, I'm exactly the market thereafter. So the question is, okay, can we find a way? Can we solve it technologically that people will be able to have their own differentiated

[00:19:59] models in which they actually create their own modes? And then they trade, right? They actually exchange these kind of unique capabilities between them. And not just by having a weaker model, but actually being able to have the smartest ones. Can this be at the frontier instead of being a destination and always, you know, always kind of losing against the big ones. So it is a vision of like somewhat more distributed intelligence, which I think could also just,

[00:20:28] you know, lead to a healthier college society by abusing the term, right? Because you have more diversity that has to collaborate somehow. And like the power is not just so concentrated into organizations. And looking at Pathway is described as an AI lab pioneering post-transformer architecture. And I would imagine that building a company around research that challenges the industry's dominant technical approach must create a very different kind of startup pressure.

[00:20:58] So again, for founders listening, looking to be inspired from you, how do you balance scientific ambition while also proving commercial value, attracting investment and deciding when research is ready to become a product? I appreciate this question is almost an episode on its own, but how do you get that balance right? It takes very specific kinds of investors. Yeah.

[00:21:21] And it takes very, I mean, assuming this is for the founders, this work is so much in our heads. Yeah. So anecdote, we called our models dragon hatchling. The architecture is called BDH. The paper is dragon hatchling, the missing link between transformer and the brain. And this comes from Terry Pratchett. And that was because in Terry Pratchett, The Color of Magic, Two Flower thinks about a dragon like he's being imprisoned. He thinks about the dragon and hence the dragon appears.

[00:21:51] And he's like, I thought about it and you just appeared. Yes, my Lord, or something like this. I am. Hence, I am yours to command. So actually the thought produced a dragon. Right. So it's kind of a meta thing that by just thinking about something, we may shape reality. So yeah, like, and we literally had to think about the dragons that we're creating. Right. And our dragons are actually reasoning machines. So it's very meta at the numerous levels. So we needed to have very strong conviction.

[00:22:18] And this is one of the reasons why, you know, coming back to your first question, like, why did we do that? I had a very strong scientific conviction about the directions we're going, about the direction that we took. And that took really, really a lot of courage and kind of commitment just at the human level, knowing that, you know, all the social pressures around us, everything would be, would be literally against. But I mean, is there a point you trust your science and you trust your gut?

[00:22:45] I mean, if you, if you see, you know, sufficiently, if we know about complex systems, like interacting stuff, how, how to model state, et cetera. And honestly, it was mostly our, our, like Adrian, our CSO, who was for a year, like, like a robot, like, and with unshaken will, I'd say. And so a lot of this is psychological. A lot of this is really psychological. Then I must say, it's not for everybody.

[00:23:11] If you're a founder and you're building a business, don't do what we do, what we do. We would have other ways of making money that would be way faster and bigger. Like, just look at our CVs, right? But what's very clear is the opportunity and the reasoning process behind, okay, where does the wind of history is blowing? We're like, where does it blow right now? Right. And in which direction? And of course the breakthrough will need to happen. Like the breakthroughs just happen.

[00:23:41] Thinking that they want is in fact silly. So this is obvious. The fact that people are just, you know, unused to change is kind of funny. So the breakthrough needs to happen. Scientifically, everybody in the transformer was inefficient. Everybody. It's like, it's not questionable. You know, we had that boxing fight. We had the code, like, Łukasz Kajsar, the coder of the transformer. He was actually fighting for the transformer. And of course he wouldn't be defending the transformer in the sense that, you know, this is like the best.

[00:24:08] And he's like, he's more like, I'm actually waiting for the post-transformers to kick in. But so far, but we'll need to just show such a huge gain for the entire industry to shift. Right. Because if a gain is too small, then the switching cost is too big. Of course. But it will happen. So, I mean, if you know how you just see the future and it's kind of easy to see per se, it will have to happen. Yeah. So that's it. That's our conviction. Right. And then it doesn't make sense to even questions because if we questioned and let's say do something

[00:24:37] halfway or kind of just a little step, like, you know, just do some transformers just slightly differently. Then, of course, we won't be able to compete with opening in tropics. It doesn't make sense. So maybe something else should be done in life. You could chill or I don't know, find a different profession. So that was this conviction element was dramatically important. And then the second thing was to have investors who actually understand that. Right. Like, this is the bet we want. But you are not to go after customers now because, of course, you could get, you know,

[00:25:07] it would be so easy to get pressures from your investors who would say, yeah, yeah, yeah, go around customers. Like, go sell. Like, what do I sell to them? A shape of a matrix, you know, at a certain point. And actually, the very good investors, even like some that are not on our cap table, but they chatted with. And they were like, it's actually harmful if anybody pushes you to talk to customers because this is a waste of time. You will sooner go and be able to have kind of massive impact of the market.

[00:25:36] And the sooner, the sooner you have, you have finished product. So you need to kind of run, kind of run lean fast and a lightweight as for, for the R&D complete this, train the model and launch it. And honestly, this is, this is what we're doing. It's kind of Navy SEAL type of approach. We're not dramatically religious about the stuff that we do. We can, we can break things, et cetera. Knowing that there are some kind of first principles, the first principles, I guess, is the dogma differentiation?

[00:26:05] Because otherwise we like, we don't want to build another open aeronotropic. I mean, the thesis is what it is, right? At least until we, I don't know, the forces of universe decide that we need to reevaluate. But I mean, that's, that's the thesis. That's the bet here. And, and move, right? And then just preserve the critical path. And I mean, many, many rules of just building a business. I mean, we kind of had to bend them somehow because they just given the pace of the field, right? That it's like, it didn't make sense to, to, to maintain it also as kind of automated

[00:26:35] coding kicked in, right? Like at the beginning of this year, it is completely changed our work. So this is, this is like us very, very difficult psychological work on the team. I realized that our, like the output of our team is somehow very closely correlated with my own kind of mental or like psychological state. Journey speaking helps if I'm just kind of cheerful and upbeat.

[00:27:02] And then, then all of this, but, but really, because given the weight, given how huge this, all of this is, any doubt can slow you down. It's easier to fail quickly and recover and, you know, and move than, than kind of, than be doubtful and scared. And we have no time to be scared. You kind of have to act as, as, as this sort of a situation. So to founders, yeah, I mean, be a warrior kind of, and just move on.

[00:27:30] And then the, the, then the investors, the investors that we have right now are blessed with the fact that investors that we have on the cap table are the biggest fans and the believers. And like, this is the bet they're like, they're proud to have made that bet. I have one guy on our cap table made his branding after our architecture. He inspired his own branding, like, because, you know, he's like, he's a smaller fund and yeah, he's like, no, I love BDH so much. He implemented it even, you know, right.

[00:27:57] It's, it's almost a cap table that are your fans. And again, they actually fuel us also. Well, of course they help so much in different, like breaking capital. It's a big help, but, but also on a human level, I think the two things that, that investors can bring to, they won't solve our math, right? They won't make the model work faster. They can help with intros, but this is, and they do, of course, whenever we need, but it's us who's responsible for the job.

[00:28:25] So, but the two biggest differences that they can make and they, they're actually uniquely positioned to deliver is capital. Actually, they know other capital guys, they know how to deal with them. You know, they know how to do it and mental support, honestly being cheerful. This will help for like for founders. And we're, we're lucky to have, to have those very supportive, supportive investors who understand what sort of venture this is.

[00:28:52] And we have talked today about the speed of technological change over the last 12 months. And for people thinking it is difficult to keep up with, this is probably the slowest it's going to go. It's going to be even quicker next year. So for, for people listening, we'll record this, what, at the end of the year, next year, what can we expect from Pathway? What is going to be keeping you busy and where can people listening carry on the conversation and find out more information about what you're doing? Yeah. So I definitely follow us on X and LinkedIn.

[00:29:23] We have a lot of things happening, you know, Dragon's doing a lot of fun stuff, but we're not, not sure yet. Like when, when, when this is going to be, to be announced and how I can give you the general direction of the field of kind of how I see the future will look like. Yeah. I'll definitely move towards AI that's dramatically more data efficient. So this personalization bit of AI models that it, it will be out there and sooner, sooner

[00:29:51] than you would expect kind of by just looking at the market right now. And this has important implications for the economics in which AI is built. Right. And then naturally serving will go down. Serving are already like inference, inference costs are going now, but, but the, this will like, this will be a step change. So I'd say this is probably all good news to the, to the founders, but people ask me, okay, how should I build right now? Right.

[00:30:21] Given, given the new architectures, the new breakthroughs that are around the corner. And my answer is just build with what is here because you also don't have time. There is a question of capturing the market, getting users, getting things right, learning. So it's our job to make our models fit, kind of be an easy swap for you. And once, once we launched, then I mean you plug in, but right now built around, around what's, what's best, what's available and, and kind of get, get the market.

[00:30:50] This is, that would be my message. And as for kind of speed of change, I mean, you know, we kind of need to understand the exponents, right? So in AI, we're all an exponent and it's like, these are literally, they're accumulating kind of compounding growth rates. Right. So we, and people don't see is that they say, oh, AI is moving so fast. It's like the fact that AI is moving so fast, this is like, okay. The issue is, and the issue, the thing is that AI is by the tool.

[00:31:19] It's not possible to enable other technologies. So the change will happen, not in AI per se, but in everything that it's going to disrupt. So this speed, this compounding growth rates, they're going to just go all around other disciplines. Imagine things within, you know, happening within years or months that, that we wouldn't, that we wouldn't dream of. Wow. I think that is a powerful moment to end on so much food for thought. And I would encourage anyone listening to check you guys.

[00:31:49] I'll include links to your LinkedIn and the social channels, and of course, the website. And you did leave us with a couple of teasers as well. So I think we need to get you back on in the new year. Certainly not leave it three years till we get to speak again. But thank you so much as always for sitting down with me. Really appreciate your time. Thank you, Neil. I love my guest's argument today for a different AI architecture and how it's also a business

[00:32:13] argument about what a smaller company can offer that incumbents cannot easily match. And getting there requires research funding and investors who understand the work that they have chosen to support. And I also loved her advice to founders waiting for the next breakthrough. Build with the tools available, learn from users and keep moving. A research team's roadmap should not become the reason to postpone learning about your market.

[00:32:44] And for anyone buying AI, I think the question remains whether that result will justify its cost in a particular task. So big thank you for joining me once again on the podcast. And please share your thoughts with me, techtalksnetwork.com. Where could a different technical approach create a business advantage for you? And what evidence would persuade you to give it a try? Let me know. I'll be back again real soon with another guest. But thanks for listening as always.

[00:33:14] Bye for now.