10 Top AI Founders building the Autonomous Economy share their secrets

Billions of agents, billions of robots doing real work and transacting. That's the autonomous economy. And if you haven't heard our last episode yet, go back and give it a listen. We unpack the full thesis there. But today we're going to go one layer deeper. Welcome back to the GenTech Podcast. I'm your host, George Bandarian, general partner at Untapped Ventures. And last month at our 2026 Annual General Meeting in LA, right after laying out our autonomous economy thesis, We put our portfolio founders on stage, the ones actually building the infrastructure, models, robots, agents, and even mind uploading that makes the autonomous economy real. And a bonus session on longevity, exploring how humans can live better and extend our lifespan in a world where robots and agents are taking on more work. So today what we have is 4 panels, 10 companies, 5 layers of the stack, and one question I asked them running through all of this, which is how is your company pioneering the autonomous economy, and you're about to find out. So let's go. I want to invite up our first portfolio founder panel. So we're going to do the compute panel. So I'd love to invite up Gil from Extropic and Haik from BlueQubit. Let's give them a round of applause. So, yeah, I mean, let's start with the basics. Why don't we take a minute? If you guys can just introduce yourselves and tell us a little bit about yourself and your company. So we get to know you. Haik, you want to go first? Yeah, sure.

Hey everyone, nice to meet you all. I'm Haik. I'm a co-founder and CTO at a quantum computing company called BlueQubit, and we do software and algorithms for quantum computers. Before BlueQubit, I was at Google, and I left Google 3 years ago to start this company with my co-founder. Who's coming from a quantum physics background. I myself come from software engineering and computer science background, and we started BlueQubit to basically understand what can be done with quantum computers that exist today. And as they are getting more and more production ready and more stable, what can be done, what value we can get from these computers. So happy to talk about that today and share this with you guys.

Great. Thank you. Welcome, Heike.

Yeah, so Guillaume Verdon, the founder CEO of Extropic, originally as well from a quantum computing background as well from Google Quantum. That's where I started my career. And then I ended up working for Sergey Brin, the founder of Google, for a couple of years on special projects in quantum tech, physics, and AI more broadly, kind of looking at a broader landscape. And about nearly 4 years ago, I left Google X to found Extropic and pioneer thermodynamic computing, which is not quantum, it's not deterministic computing, it's a not-so-secret-now third thing. And yeah, so thermodynamic computing came from essentially me realizing that most people coming to quantum computing wanted to run a probabilistic workload, whether that's optimization, simulations of finance, biology, chemistry, and of course artificial intelligence. And so essentially there's an opportunity to build a probabilistic computer for the probabilistic era of software. And 4 years ago, you know, that was very non-obvious that most software would be probabilistic. But now we're living in that era where most compute and most wattage for compute is being put towards LLM inference and test-time compute, which is diffusion models or reasoning.

And when you talk about probabilistic compute, so what is that versus the others, just so everybody knows?

Yeah, well, so probabilistic is like a broad class. We do thermodynamic compute, which you could view as a probabilistic computer that is programmable through equilibration. The way to, the way to view how this computer works is Traditionally, you know, usually I have these boxes with, you know, there's a partition. My ball is in the left partition or the right partition. Definitely, it's definitely 0 or definitely 1. In our case, it's an analog stochastic computer. So the way to view the computer is I have this landscape of hills and valleys, and I have many, you know, balls in this landscape, and then I have— I can shape this landscape, and I have some wind that has many gusts and has some randomness, which is, you know, in our case the balls are electrons and the gusts are just from the randomness of the hardware, of thermal fluctuations, and essentially shaping the valleys, we do that through voltages. And through this you could do many algorithms, a broad class of algorithms known as Monte Carlo algorithms, or or probabilistic inference. So that's very relevant to anyone in finance, anyone in defense, robotics, edge computing, and of course GenAI. So it's a, it's a very broadly applicable paradigm. And essentially by going probabilistic, you save a ton of energy. Because it turns out that maintaining knowledge, maintaining determinism, fighting sort of entropy, right? Entropy is lack of knowledge. It costs you energy. And traditional computers, you pay energy to maintain your computer in a deterministic state. So if you learn to let go slightly, not fully, but slightly, you could save a ton of power. And so by the projections we released in October publicly, we get about a 10,000x energy efficiency gain for diffusion-like models. Running on our hardware.

And explain why that's such a big deal.

10,000x— using 10,000x— not 10x, 10,000 times less energy, not percent, than a GPU. That's a big deal because you get far more intelligence per watt. You can densify intelligence massively, so you can have much more compact systems that are much smarter. So in order to have AI chips in everyday wearables and our robots, having robot brains in order to have intelligence at the edge. And of course, even our data centers in order to serve the insatiable demand for intelligence. We're running out of power right now. It's very difficult to scale up the power supply to keep up with the AI demand. It's putting a lot of stress on our grid. By some projections, we definitely need to double our grid, if not 10x it. If we want to serve the whole population, never mind trying to achieve having, you know, a billion manufacturing robots, humanoid robots. We can't, we don't have the power grid for that right now. And so instead of just scaling up the Korean paradigm, which is what everyone's doing, which is a short-term strategy, we're taking the longer-term strategy, which is to densify compute. We have to reinvent the whole stack from scratch. That's a huge challenge, but we've made a ton of progress. And now that, you know, we can use AI to even accelerate our own work, that progress is accelerating. So it's kind of, yeah, it's kind of the final form of AI assembling itself for the future through us in a way.

Yeah. Great overview. And I'm going to come back to Haik in a second. The reason I brought that up is because that's what you were just alluding to, is we are entering this world and the current compute, everybody buying NVIDIA's chips, is not going to work well going into that, and that's what you're solving for, right?

Well, it's like, you know, 200 kilowatts, 200,000 watts per model that, you know, serious large-scale model that you have to train, or more, right? It's almost a megawatt. You can't have a billion robots with this paradigm. You just can't. Like, not only are you going to run out of power very quickly, but eventually you're going to run out of like the heat exhaust capability of the planet. And so our work is needed basically yesterday. So we operate with sort of wartime urgency because we're, you know, we talk to various hyperscalers and they're like, hey, yeah, we're spending a quarter trillion doing these buildouts. If you can start making a dent into NVIDIA's market share and we could run more of our models on your hardware within 2, 3 years, we're going to save hundreds of billions of dollars. That's a big deal.

Yeah.

Right. And so they want to support us.

Yeah, that makes sense. So coming back over to Haik, so Both of you came from the quantum world. You're in it. So give us an update. What's the state of quantum? What's the current and future? Where are we heading from your perspective?

Right. Yeah. Thanks for the question. Yeah, I totally agree with Gil about the power consumption of classical computers. So in our quantum world, we make this distinguishment between regular computers that we have now, which are CPUs and GPUs. We call them classical compute, and then we have the quantum compute. I wonder how Gil thinks if their machine is classical or quantum.

Thermo, third thing, right?

I don't know much about thermo, but, um, I can talk about the classical versus quantum. So where quantum computer computing is, uh, today, it's basically— there is— we have a bunch of companies like Google and IBM spending billions of dollars to build scalable quantum computers. And last year there was this important milestone where we noticed and we observed how multiple quantum computing companies, uh, namely Google, IBM, Quantinuum, and few others reached to a point where they have quantum computers today that can run things that no amount of GPUs or CPUs can replicate such results. We call that scientific quantum advantage. So these are specific problems that these computers can solve in hours or seconds. But to replicate that same computation classically using like Elon Musk's Colossus or the largest data centers in the world using like million GPUs hooked to a nuclear power plant, still they would need like thousands of years to replicate that computation. Now you might ask, if we have these powerful computers, why are we not seeing them everywhere? Why are they not revolutionizing everything? And the answer is there's a little catch. The, um, there is still a lot of noise in these computers and you can only run very limited applications on them. And that's why we talk about these very specific problems that, problems that we have found so far. We call them scientific quantum advantage. There is 5 such examples, a few of them from Google, a few of them from Quantinuum. One of such claims is from our company, BlueQubit. We published a paper on it. It's called Heuristic Quantum Advantage. Uh, one of the things that our company does is expanding and exploring more such algorithms that can be run on a quantum computer. So as computers are getting bigger and their capabilities are getting bigger, we can run more and more algorithms on them. And the goal is to go from scientific quantum advantage to commercial quantum advantage. Which will require much more gates and much more qubits. Yeah, last year was a very important milestone. And looking at the quantum computing companies' roadmaps, we expect that in the next 2 or 3 years we'll start seeing the first large commercial applications of quantum computers.

Yeah. And so just to make sure that we capture a key part of what you mentioned. So, so let me just take this in. There is 5 companies in the world that have achieved scientific quantum advantage. One of them is Google, one of them is IBM, and BlueQubit is one of the 5. So tell us about that. How did you guys do that?

Yeah, this is my favorite subject. I love talking about this for hours, but I'll try to limit myself.

It's a big deal. You should talk about it.

Yeah, I'll talk about it in a— I'll tell you guys a little story. So there's this conference called Q2B that happens in Silicon Valley every year. It's the largest quantum conference. So when we had this result, like a month before the conference, we were thinking about giving a keynote speech so that we can announce to the world that, hey, we found this amazing result. We can run this algorithm on a quantum computer and it would take like 10 years on the largest supercomputer to replicate this result. But the cost for this keynote speech was $40,000 and we're like, okay, this is an important result, but We are a startup. We're not gonna pay $40,000 for the 30-minute slot. But instead, there are two scientists, very well-regarded scientists in our world, Scott Aaronson and John Preskill, who had the first, the keynotes of the first day and the keynote talk of the second day. And both of them mentioned our result alongside Google and Quantinuum, and we were shocked and impressed. That was very great. We didn't even know that they were gonna mention it, so it was like a completely a third-party objective verification of what we've seen ourselves. And yeah, it's, it's just an amazing result. And we are working on similar results internally now. In the next few months, we'll be announcing more such scientific quantum advantage results, specific problems that can be run on a quantum computer today that will take months or years on the largest supercomputer to replicate. And our goal is to go from 1 to 5 to 100 such results And eventually some of this will start spilling into the commercial realm. We're actually already starting to see how some of these problems have relevance to material science and how larger versions of these problems can actually lead to discovery of new materials and to actually commercially viable products.

That's exciting. Excited to be on the journey with you on that. Gil, so you've presented with our investor community a number of times. I think the thing is Everyone's very excited and it seems like there's this path to abundance and amazing results. And you create— you created that execution plan which you should— we shared and everybody kind of dug into that. At this point, what can go wrong?

I think it's like traditional execution risk at this point. I have pretty darn strong conviction now that thermodynamic computing is happening no matter what. And I want us to to be the ones to keep being in the lead and pioneering this and scaling this as aggressively as possible. And so, you know, right now we've tested our components that we can manufacture them. There's a—

Tell us about that.

Yeah. So the thing that's really nice and, you know, the reason I left quantum computing to go to probabilistic and thermodynamic computing is because essentially we are computing with noise, you know, whereas in quantum computing you assume the system has zero noise. So you have to you spend most of your computation filtering noise, which is quantum error correction. In our case, we can leverage the noise. So we've characterized the noise very well. We can manufacture them at scale with TSMC and GlobalFoundries. We can operate them at room temperature. We can control our components such as to do the algorithms we want to do accurately enough, because again, probabilistic algorithms, they have randomness to them. And so you don't need as much precision as a traditional computer or a quantum computer. And so, you know, just like our brain is a weird wet chemical computer that's sort of lo-fi and janky, if you're running algorithms that are brain-like or probabilistic algorithms, you don't need as much precision. And so we characterized that, and now we've done projections for the next chip that we're taping out, you know, end of month right now is the schedule. And that's with TSMC. It takes 6 months to get back. But we're already planning scaling up our systems. And, you know, unlike quantum computing that kind of, you know, when I started in quantum computing almost 10 years ago, we had 3 qubits commercially. I was the first user of Rigetti. And then it was like 10, 20, 30. I think now they're hovering in like 100, 200, maybe 300. So, you know, this device technically there's a few hundred p-bits. We've only activated a few at a time for I/O purposes. So you have like 8 degrees of freedom in parallel here. Our next chip is 250,000 Pbits, and so that's what we're going for.

All right, that was a very clear, strong answer to my challenge question that I had for you. So, Haik, I mean, what does this look like once we reach the commercial scientific advantage that you were talking about? And if it goes as planned and we're living in this mixed world where we have the old traditional compute, we have thermodynamic compute from Gil, and we have quantum compute coming online. What does that look like, I guess, for you guys at BlueQubit?

This is one of the most fun exercises that we as a team at BlueQubit do as well, trying to understand, let's say you had like a million qubits, fully stable quantum computer, what could you do with it? And the possibilities are just crazy. The, The way we think about it, it's like trying to, you know, imagine it's like 1940s, you know, Turing is building his machine to break Enigma, the World War II thing. And now you have Turing in his spare time thinking where this machine is gonna lead to, and you try to predict like internet or TikTok or Zoom from just that one machine. And it's crazy how that one thing can lead to all these crazy technologies or AI. Try to predict AI from the Turing machine. We think a similar thing is going to happen over long term in quantum computing. If you think about like 50 or 100 years where it's going to lead to, I joke about it sometimes, but I wouldn't be, I wouldn't be surprised if in 100 years we have like teleportation or like, you know, like really crazy things. But that's in the long term. In the short term, what's going to be very useful and everyone is going to feel it when we have like scalable mainstream quantum computers. The first applications we think are going to happen in material science and in chemistry, meaning we'll be able to create better drugs, which means better, more cures for different diseases, maybe even longevity. Also going back to material science, I know it's just like an abstract term, material science. What does that mean? But The point is with quantum computers, you can model molecules at the basic level going off of first principles and building new materials, simulating them inside a quantum computer and then synthesizing them in the lab, doing million iterations in an hour instead of like waiting 6 months in a lab to do a single experiment that can lead to new types of batteries, fertilizers, catalysts that can change and revolutionize whole industries. So I want to end this with one phrase that I really like that I think applies to this scenario, to this case a lot. I think it goes by something like humans tend to underestimate things over long run and overestimate things over short run. I think with quantum computing, that's, that applies 100%. So whenever we try to kind of zoom in and understand what is it going to look like in a year or like in next 6 months, we're always off, but when you zoom out and see how far we've come and what's next 5 years or 10 years are going to look like, it looks very promising.

Gotcha. So on that point, we're running a little bit behind schedule, but we have time for one, one or two quick questions and quick answers from you guys, and then we'll move to the next panel.

When do you think we'll see the first kind of ubiquitous commercial application, and how will we know? Maybe you could make some predictions on like how that might hit kind of mainstream consumers or like mainstream people. I know you mentioned a couple of examples in materials and medicine. If you were to make maybe a prediction in some concrete application that consumers will see.

The revenue in quantum, a lot of different people look at it differently. One way to look at it is, so a lot of people in quantum say quantum advantage or quantum usefulness is when somebody is ready to pay for it. We think that's a terrible definition, but if you go with that definition, there are companies out there like IBM already having revenues in, in, I think they hit a billion dollars this year just by selling their quantum computers and time on their quantum computers. But we do not count that as actual value add because most of the times that they sell these computers to like research labs or basically forward-looking teams that are looking for applications and spending money for that. The actual value add we think is gonna happen again with one of the use cases that I mentioned, for example, material science or chemistry, molecular dynamics, things like that. And first is gonna be large businesses renting time on these quantum computers and solving these compute-heavy questions and problems. But the way customers and consumers are gonna see it, it's gonna be like a cheaper, just everyday materials or better batteries. For example, your Tesla car might have a better battery. Those kind of things will be a derivative of better quantum computers.

Okay, a round of applause for our compute panel. Great job. Amazing, fascinating stuff. Okay, so our next panel is our convergence panel. So I'd like to invite up Matthias and Andrew. Come on up. Matthias, why don't you go first? If you could just introduce yourself, tell us a little bit about yourself and your company.

Yeah, hey, my name is Matthias. I'm co-founder and CTO of Liquid AI. Come from the research background, like 10 years ago started into more efficient ways to do machine learning by looking into neuroscience, like what has been discovered in the neuroscience space. Met the co-founders during this research journey. And then about 3 years ago, we decided to start Liquid AI, and there we focus on on-device foundation models.

You guys met at MIT, right?

Yeah, exactly.

Like, we were all doing PhD work at that time.

PhD, postdocs, and Daniela is a professor there.

Right. That's right.

And yeah, since then focused on— yeah, so basically an MIT spinoff, 3 years old, and focus on on-device foundation models. So really, like, think of ChatGPT, but that runs locally in a car. And things like that.

So on device, uh, can or doesn't need to be connected to the internet. I think your unique value prop, your wedge in the market, is going on devices, uh, versus what we're used to with ChatGPT and Claude that run in the cloud and work to serve the user. So, okay, and we'll talk more about that. Andrew.

So similar academic origins, I would say. So 3 Laws was co-founded by myself, Tom, and Aaron, who's a Caltech professor. I did my PhD at Caltech studying safety. What does safety look like for this new class of emerging systems? Both from the hardware perspective, we're seeing a lot of interesting legged robots. We're seeing a lot of new interesting form factors. But also in terms of the software that governs their movement.

AI.

So these, you know, unexplainable models. How can you make sure that these things can be certified and deployed in the real world?

Awesome. Andrew, so give us an update on the current state of robotics from your standpoint.

Yeah, it's a really exciting time. We're seeing evolution.

Are we going to have billions of robots in the next few years?

The next decade. Few years is possible for sure.

In the next decade.

But I'm confident to say in the next decade. We're seeing a lot more general-purpose form factors in robotics. When I was first looking into robotic startups a couple of years ago, Everyone was doing one very specific vertical, building the world's best box taping robot. And now it's about a bunch of companies trying to tackle these general purpose machines, like humanoids, we're seeing a lot of hype in right now. So that's one big axis of change in robotics. And then the other one, of course, being the software architectures. I studied traditional classical control theory and motion planning. Now we're seeing a lot of foundational models. Kind of the robotics analog of large language models in terms of how can you build one autonomy stack that can handle a wide variety of tasks. So those are the big changes I'm seeing in robotics right now.

Gotcha. And the solution you guys had built, the universal safety layer for robotics, right? So kind of like the stack that we were talking about, we're going to have to— safety and trust was really important. And so what you guys were really early on was effectively— tell us more about that.

Yeah, sure. So, you know, of course you're going to have the intelligent software governing what the robot does. You also need a safety system in order to deploy these robots around people. And the problem became the autonomy or the decision-making, the brain, was getting so much more complex and capable and performant, but the safety systems are kind of stagnant. And so you end up getting limited in your performance by the bounds of your safety system, not the bounds of your hardware platform or your autonomy software. And so 3 Laws' mission is not necessarily to make industrial robots safer because there's already an extremely high bar of safety for industrial robots. It's how can we leverage all of these advancements we're seeing with new form factors and AI and actually get that performance gain instead of being restricted by the safety systems of today.

Andrew has some great videos of work that they've done with their clients where the robot before 3 Laws software runs into, sees a human, kind of stops, is thinking, and then carefully goes around it, but at scale. But that takes, let's say, 42 seconds. And then when they apply their algorithm, that takes 5 minutes to install. It's just like a mathematical layer that comes it improves the calculation and it does it effectively immediately or maybe with like a 2-second delay. It sees the person and goes. So, you know, tremendous improvement over the current process. That's a kind of better visual than me describing it, but hopefully that, that's helpful. Mateusz, from your perspective, what's— where are we at in the world of state of models? Obviously you guys don't compete directly against Anthropic and OpenAI, but you follow them and their work. They may want to come into your area one day. You may end up competing. But I guess tell us about where are you seeing models going and how that's impacting what you're doing at LIQUID.

The way we see it is like models will be everywhere, basically in the cloud, but also locally. Sorry, everywhere, like in the cloud hosted, like ChatGPT and so on, but also locally, especially for robotics. You really need to run on device. There's privacy and so on, obviously. But really, we have, for instance, one customer that deploys our model underground in some tunnel construction site. And there, they just don't have connectivity. This is one example of really where you need to really work on the lower-level stack, really to optimize the model to provide the same quality as you observe in the cloud on your device.

So in that customer's case, what are they doing? Is it like a mining— is a robot doing mining? What is the model doing for that customer?

Yes, like think of it as a worker there, you know, that works underground and like has basically heavy equipment. So they need a voice assistant. So they need kind of a voice agent that they control.

Okay, got it.

And our model is powering that.

Got it. So what is the market for what you're doing? What, you know, I guess we're all very familiar with the TAMs of consumers and businesses adopting their version of ChatGPT, Anthropic, When you say on-device, what products does that really mean and how big is that market?

Yeah, so we are in touch with like most smartphone manufacturers and also like a lot of the car manufacturing industry. So there's a lot of demand of really like pushing intelligence, what people are used to, like on ChatGPT and so on, really to bring it in all different devices because most of the silicon, you know, even though there's all of this data center buildouts.

So you're talking about phones, cars. Phones, cars, laptops, anything like that. Laptops, you'd mentioned TVs.

Exactly.

And you're, you're already working with customers across all these verticals, right?

Yes.

Any questions for Andrew or Mateus? So I have a company called Veterans Home Care and the government back in 2011 said by 2030, there's going to be a significant shortage, about 30% shortage of caregivers. COVID came along, accelerated that process, and we have a shortage. And, uh, I'm eager to hear what can be done, and obviously, humanoids, which I'd love to have, they're going to be a little farther out, and obviously, the best opportunity right now is transitioning. Do you have any suggestions?

Humanoids in the home is further out than any humanoid company will want to tell you. That's about the most complicated environment you can imagine, and it's incredibly hard to control these systems, even in industrial settings right now. To have them be robust and reliable, much less putting it in someone's home. They can trip on clothes, they can trip on anything. But what form factors that I'm really interested in, something called mobile manipulator. So mobile manipulators, you can imagine kind of humanoid arms, but on a wheeled mobile base. Of course, you know, there's issues with things like stairs, which is why humanoids are such a popular consideration for the home. But there's versions of those that can go upstairs. And it tackles the fundamental challenge behind humanoids, which is stability. Because you have a stable wheeled base, but you still have the dexterity of manipulation arms. And so those are the form factors I'm the most excited about and what I'm actually seeing deployed in the real world and adding real value. Because you can still do extremely complex manipulation tasks without the concern of stability in a home.

Paul's dying to ask a question there. It was— I was thinking about this for both panels, is what protection do you have, like China just knocking this off or stealing our tech? Is there protections? Because it looks like you're doing some amazing stuff.

You know, America innovates, China duplicates, Europe regulates.

I can start with that. Um, at least on the robotics side, uh, China creates some really, really impressive hardware at price point. It's really hard to beat the hardware that's coming out of China at representative price points of what they can do. The software is what causes a lot more pause. Not just because it's typically lower performing on the robotic side, but also because of security concerns that people have, and privacy concerns that people have. And so we haven't seen, we've seen a lot of Chinese robotics companies take this, what I'd call like the Unitree approach, of let us just sell you the platforms and you develop your own software on it. And so we're seeing that a lot with hardware coming out of China, but I think it's still very much the U.S. that's dominating the software that's running on the devices. There's also people pushing back to bring, you know, manufacturing of robots here in the U.S. And we are seeing companies come down in price points and, you know, start to compete with the price points there. But in terms of protecting the software edge, I think it's, you know, a big strength the U.S. has always had is our university system. We have people, you know, world-class talent from around the world coming to the U.S. to learn and to study. Being able to maintain those people, to employ those people, to keep those people motivated and working here, I think is the best way to maintain kind of a competitive edge if you're talking about a nationwide competitiveness.

Any other questions for Andrew and Matthias?

So for someone like me who owns a distribution and warehousing business, what are we looking at for, you know, bringing in some of those into our facility?

But maybe not as a humanoid, but what you talked about, which is the stable platform.

Yeah. So mobile manipulators, I think there's a lot of great manufacturers out there today and you can start automating workflows. I think what's really happened, you know, in terms of general shifts is we're entering the era of autonomy. We're leaving automation where you had to have, you know, an industrial manufacturing line. You know, someone comes in, installs robots with fixed software. I think we're really entering the point where there's, you know, commoditized robots that business owners and facilities owners can buy off the shelf and with relatively low pain, I would say, integrate them into a workflow. So, you know, pick and place applications, material handling. Like, this is stuff that, that we can do now with autonomy. And we can onboard new companies and new facilities much quicker than what we could do in the past.

Okay, awesome. Let's give them a round of applause. Thanks, guys. Thanks, Matthijs. Thanks, Andrew. All right, time for our next panel. Who cares about health and longevity? So why don't you guys just take a moment, introduce yourselves. Welcome, glad to have you. This should be exciting. Yeah, go ahead, Bruno, kick us off.

Thank you. Hello everyone, I'm Bruno, co-founder at AnyAI, and we are in longevity. But longevity is really dirty word these days. Everything is longevity. However, we don't know what aging is, what aging biology is what longevity is. How can we tackle something, how can we solve a problem that we cannot measure? So what we are working— we are working actually on measuring the aging itself. That is a process that we don't like, that is cause of all diseases, including death in the end. And we are trying to measure it with first-generation clocks like Horvath clocks, second-generation. But clock is one number, and biology is multifaceted. Biology is self-organizing system that is maintaining its coherence and we are measuring that coherence. We are doing a bunch of clinical trials, compressing everything into latent space and understanding, modeling one's biology so that we can understand what perfect health is, what, where is it decoherent, and how can we affect it. If we can measure it, then we can affect it. And long story short, that is what we do.

Yeah, we'll dive into that. Great intro.

Hey everyone, I'm Michael Landryg. My company is working on whole brain emulation. So this is scanning people at the cellular level to create a digital copy of them in a virtual world. I should say maybe a little bit about my background. I originally studied molecular biology, so that's why I'm probably here. But mostly I've done nanotech. So this is building high-speed electron microscopes for scanning DNA and on the hardware side of supercomputing. So moving bits quickly through a data center and between computers. So this kind of combines microscopes and supercomputing. The brain is incredibly complex. There's a lot of material you have to image. And you have to do something with that, with all those images to convert it back into a graph of where all the cells are and how they're connected to each other. Yeah, I can go further into detail about like what we do and why, but maybe you'll—

yeah, we'll get into that after we get the intros. Yeah, let's hear from Jeff.

Hey everybody, so, uh, Jeff Chen. I'm co-founder, CEO of Radical Science, uh, MD by training, a doctor gone rogue really trying to build a parallel system to the current healthcare system. And it's about proven precision wellness using foods, supplements, natural products. How do you get there? Uh, that's kind of what radical science is enabling. It's basically how do you take, uh, AI, how do you take citizen science, how do you take decentralized clinical trials, and how do you turn what currently is described as food supplement wellness products into legitimate medicines? And you know precise health effects for whom they work in. And because they're already FDA approved, all we need is the data layer to figure out who needs what, to actually make dramatic impacts on their health.

So maybe just give us an update. What's the state of the longevity market today? Is aging really a disease or is it part of life? And how are you each playing and capitalizing on that?

It's not even necessarily about living longer. It's just like, can you be really healthy and functional till the day that you, God forbid, pass? And that's a massive win. And right now, the ways that we measure longevity, they're a bit crude. Some people are using epigenetics. You hear words like biological clocks. Some people are doing it based on your functional outcome scores. For any of you that have used Function Health, you log in, your biological age always seems lower than your chronological age. But now there's some things like ARPA-H just funded an initiative to come up with an intrinsic capacity score, which is maybe potentially the most accurate. A comprehensive way to measure longevity. And it's going to be multifaceted. It's going to be based on a combination of all these things. What's your function? How many steps can you take a day? What do your vital signs look like? But also, what does your blood biomarker show? And yes, maybe some things around epigenetics as well. I know there's a lot of people who are also— there's on the pharma drug development side, there's some drugs pushing through to actually go after aging as a disease. That would be massive. Then once that path is paved, you're going to see a lot of investment and development in the field. And so it's an exciting time, but there's a lot of work that still needs to be done.

Yeah, there is a big battle in longevity field, and that is battle between healthspan and lifespan. Healthspan is politically correct. You can say it because then you can tackle it. You can go through FDA, through pharma. Everyone are fine with healthspan, but actually nobody cares about healthspan. We all want to stay young. If we are old and healthy, it doesn't mean much. We are still old. We don't want that. And that is the whole point behind aging is a disease. We don't want to lose our coherence. If we are losing our neurological function, we call it a disease. It's Alzheimer's. If you are losing all the other functions, we are calling it aging and it's fine. It's not fine. We have technology, we can tackle it. What we want, we want a system, we want an AI that understands us in real time and that can tell us adjust slightly and that we can grow with it. And that is actually what we are building. And that is a big challenge of the industry because if we actually do that, if we actually measure aging and coherence, that is the end of pharma as we know it. Because if we can stop disease before it's a disease, 10, 20, 30 years, and we know that we can detect it, then it's the end of pharma industry as we know it today.

So let's drill down on that. Yeah, round of applause. That was great comparisons to a lot of people. So continue where you left off. So that's what you're attacking. So what's the number one killer of all humans worldwide and how are you guys attacking that?

Cardiovascular diseases. And among first 5, we have cardiovascular disease. We know that we can see risk 20 years ahead. ApoB/ApoA ratio, so small lipid particles. And we can already see that only from your digital scan. This is really fancy. Now we are waiting for all the validations, external validations. But this will be first product because if we can catch number one cause of death globally 20 years before it happens, then you can steer it. How we are doing it is you, you mentioned all the epigenetic clocks, you mentioned everything that's happening. We are the only company doing everything. So if you now ask your ChatGPT what is the deepest ever profiling clinical trial who that was conducted, Mike Snyder did it. IPO, deepest ever. We have 4 times more that we just finished now in Sheba in Israel. And I know because Mike is on our board, so he is the second one now in the world. And the only one that is above Sheba that we just finished is Trial that some of Untapped LPs are part of.

That's right. So similar to you guys, so you found your product market fit with the Genesis product, but why don't you take us a step back and say, you know, what did you do that led to build the data and the insights and the systems? And then what is the Genesis product?

Essentially, the biggest problem that has really prevented pharma from scalably developing drugs or food companies from creating truly measurably better-for-you foods or dietary supplement companies from creating truly efficacious products is they all are missing the ability to generate rigorous clinical evidence on any efficient manner. So that's where everything gets stuck. We can do rat studies. We can do AI prediction models. But the second you want to actually put it into complex human beings living in a complex world, that's where everything comes to a grinding halt. So when we started the company, we said, OK, what's it going to take to create a universe of proven precision health treatments? You've got to solve the clinical evidence piece. So the initial go-to-market for us was, let's effectively reinvent how trials are done. And let's serve a sector that hasn't really had access to trials, which is the food, beverage, and supplement categories. So we basically took everything that you come to know. I used to run clinical trials at UCLA, and you basically— we took out every inefficiency, and we automated most of the steps with AI agents, and we brought them into the homes of the people in the stu— in the trials. So it's citizen science, and you're studying real people in real environments. And because the things that we're studying are already presumably safe, you now no longer need strict medical monitoring and hospitals to administer these. So now you can scale really rapidly. So we started serving medium and large size companies. And for the first time ever, they're now able to generate real health data on their products. So like our clients are, you know, Unilever and PepsiCo and P&G and all these big companies. And we're effectively running trials on their products. And that was kind of the first thing that we needed to do to master this system. And each year, at this point because we've scaled up, each year we're putting more Americans through placebo-controlled trials than all FDA drug trials combined.

So effectively what— there you go. Right? I think it's a big deal because what they're doing is taking all the supplements and all the claims that they make, that the manufacturers make, and they're running an FDA-level kind of the gold standard double-blind study on 10,000 participants that if this company's saying, hey, my ashwagandha is going to help you sleep, and the manufacturer wants to prove it, then they could actually run an FDA-style study to get the data around that. Now that they've done that enough times and they understand what works and what doesn't, that's led to Genesis. So Michael, let's go to you. So you and I talked, you have this kind of 5-step progression through the milestones that you guys are working on. So maybe can you kind of walk us through there? Let's start with the fruit fly that started all this.

You know, so what I mentioned is we're creating, we're creating, so right now, you know, these guys are trying to solve longevity and like the human thing. It is so hard. That was my first company I was trying to work on this problem too. It is so hard and I like please, please, please solve it because it's like it's required, it's needed. You know, there's 100,000 people that die every day of age-related diseases in the world. It's a gigantic problem, right? Every one of us is currently dying, just fairly slowly. So, you know, our goal is to create a digital replica of you. If we do it accurately enough, you might even consider it you, right? Right now we all run on biochemistry, like biochemical cells. If we can create a computational cell that acts exactly like every one of your cells in your body and do a one-to-one digital twin of you, that being, that digital person, will act and feel and think maybe like you. So we've already done it for a fruit fly. And if I had a video, I could show you the fly walking around and finding food and navigating an environment, obstacles, grooming, and avoiding, like, looming threats. So this is a scanned fly. This is not like an animation with AI, like, running it. This is— we have no idea what the fruit fly is intending to do. We're just simulating neurons and cells, right? And what we are is a bunch of cells that are interacting, right? And then we at some point have motor movement, and thoughts turn around in our heads. We see light from the environment. We hear sounds. So what is this useful for besides— well, I think it does solve the longevity thing. But before that point, that's the hardest part. That's our end goal. Before that, just scanning a brain and body gives you incredible information about brains and bodies. So we've never had a full scan of a person. The largest amount of tissue that's been scanned is a cubic millimeter from a human brain. The total amount of tissue that's been scanned in the world by everyone is about 2.2 cubic millimeters. So that's tiny, tiny, right? This is like not anywhere close to the size of a brain. So the two fruit flies that have been scanned like fits within this volume. So what do you do with that? Like you got a brain scan of a person, what can you do? So one of the most interesting things, for at least longevity, is helping solve brain diseases. This would be an incredible diagnostic tool. You can scan a person with Alzheimer's, you can understand which neurons are being affected. You can also make this a chemically accurate model, so you could introduce new drugs and see where do they bind in the brain, where do they bind in the tissue. That's for Alzheimer's, depression, schizophrenia, wiring diseases especially. So that's the nearest-term thing, meaning just a brain scan helps us learn incredible amount of just about brains. What else can you do with brain scan? You can make better AI. This also helps solve longevity in a kind of a roundabout way, in that intelligence helps solve problems. Right now, companies are spending maybe up to $1 billion per training run. So hopefully Gil can reduce this cost. It's a lot of money, right? But it's worth it because you've now made an artificial mind that can solve problems. Actually, how many people here use ChatGPT or Claude or Gemini? Okay, almost everyone. Great. So, you know, the value of AI is clear. Now, we could hire super expensive engineers and run very large training runs to train the next model, or we could just literally scan this thing that evolution invented over millions of years and find the algorithms that evolution has created. So really just discover those algorithms. Humans clearly still have an advantage in a bunch of areas. We're just more robust in many ways. And so yeah, there might be the next transformer in here.

But tell us where this is going to go. Once you do one, you've proved that you could do one whole brain, which is obviously a moonshot. But if we've done one, then it becomes a market. And where does it go from there?

We're aiming at high fidelity, sort of like perfect emulation of the organism as they were when they were scanned. So that's a cell-for-cell emulation. We think, I mean, we are made of cells, right? But we are not the atoms we're made of, meaning everyone in here is made of entirely different atoms than we were 10 years ago. We're just changed out. So why do we even call ourselves the same name? Why do you think they're the same person? Well, it's because the pattern is continuous. We're changed out little by little, but the pattern remains. So if we can record and emulate that pattern on different hardware, on a different substrate, just like when you upgrade your cell phone, right? Your old phone, it's cracked screen and it's dented. You've dropped it a few times. When you restore, you plug in your cable, you transfer your apps, you transfer your photos, and everything's there again, right? And now it runs faster and it has a better camera. So I think just like that, we will be able to transfer who we are, our memories, our habits, our mannerisms. I think we can transfer that to a digital substrate. And what can you do then?

I think the practical use cases that we talked about was say I'm a lawyer and I pay, I'm one of the early customers, I pay the big bucks, whatever the initial scans will be before cost comes down, I get my brain scanned, now I could have the virtual George lawyer go and practice law on my behalf, take 1,000 times as many clients as I can take, theoretically do the same quality of work, and so that's one use case.

Yeah, yeah, definitely. So this is the autonomous economy. Yeah, exactly. So you can make a copy of yourself with certain skills, and you could have them go out and do work and be an expert or thing. Probably manage other teams and agents and things like that.

And then the big one is death, right?

Yeah, and then of course, you don't have to die if you don't want to. So in some ways, it's like— Solves longevity. It solves all diseases and aging. And also gives you like a stake in the future, like a part of the future. You know, we see AI is like improving so quickly. This helps you not be left behind.

I told you guys, I braced you, right? I said put your seatbelt on. You know, Michael's gonna tell you something that you probably haven't heard. But, you know, I mean, I think a couple AGMs ago I shared with you guys the story about my dad, right? And I lost my dad when I was 1 and a half. And I spent a lot of nights crying and wishing that I could have some way, as a kid, right, that I could have some way of interacting with them. So, if this service was around, right, and my mom and dad could have got that subscription, he could have had his brain scan, and then when he died of what he died of, you know, then I could have had some semblance of a connection with them. So, I think the market will probably start as each of us thinks of our aging parents and, people that are on the deathbed that have fatal diseases. And then we'll— there's a lot of lines around ethics and things that we'll need to think about for sure. I want to go to the Agentic panel, and I'm sure you guys have some questions, and we'll have a lot of great dinner and drink conversations over libations, but I'll take just like one if anybody has a burning question. Yeah.

Well, Bruno, I gotta Question for you. And Jeff, if we can get beer into the rocks category, I'd appreciate that. But so Bruno, this idea of measuring your health, right? So like I got a full body MRI 10 years ago. I do my functional health twice a year. But health is something that happens every day. And I'm a big fan. So I like all the science behind what you're doing, but how you do it. I think is like super important in the fact that you've made it really easy to collect a huge amount of money or data every day that's actionable. If you could just talk about that for a minute.

Yes.

So we have the largest biobank in the world of coupled data. When I say largest, I mean deepest. So multiomics, genomics, proteomics, transcriptomics, metabolomics, microbiomics, et cetera. We are the only ones doing all of them plus digital modalities. We have scans of tongue, eyes, face, couple of questions that are dynamic, that are changing, and we are doing them simultaneously. We are the only company having those two simultaneously because you cannot reverse engineer from all the datasets out there. And then we are coupling them and creating and using, as Gil is always talking about, model of physics for biology, because this is state space that is dissipating heat and that is moving and we must measure it temporarily. And because of that, you do it with your phone daily, 1 minute, and then you have longitudinal. If you have longitudinal, you can see where it's moving, you can see its inertia. If you can predict its inertia, then you can affect it. You can affect it today. You see what's working in real time. You don't wait for 6 months to see if your trial that you're paying $50 million as a pharma is working. You see it right away. You can steer it. And then you have N-of-1 personalized medicine. That is the only way how we can go into longevity medicine, not averages, but going for each person every day and steering that biology.

On that, let's give them a round of applause for the panel. Thank you. Thanks, guys. All right, our Gentech crew, bring us home nicely. Thank you, guys.

Thanks.

Why don't we go start with you, Phil? Let's have you guys introduce yourselves and just tell us a little bit about your company.

Absolutely. Thanks for having me. Before I talk about myself, which I enjoy, trust me, I want to say thank you to all of you because people like you make this entire thing possible. LPs who fund VCs like this guy make this possible. So I wanted to take one second and say thank you. I've met over 100 investors in the last year building this company, and I told you before, I think he's doing something cool here. So thank you guys.

Thank you.

Appreciate that.

Now I can start. All right. My name is Philipp. I'm German originally. That's where the accent comes from. I'm the co-founder and CEO of Nexa AI. We're an applied AI lab for heavy industries. And the thing that I'm obsessed with is I grew up in the metal industry. I worked in rail. I worked in construction, top management consulting, always like Fortune 500 level. And yes, the whole world talks about AI, but the reality in most companies, mid-market to enterprise in the US, in Europe, in Japan, you name it, nothing changed over the last 4 years. I just came back from New Orleans from a conference and the main sentence is we just started using Microsoft Copilot. Nothing changed. My company, we're combining human and artificial intelligence to bring AI into those environments, into those companies. That's what we do. We started working on this 14 months ago. We work with companies like Siemens, Schaeffler, BNSF, some of the biggest companies in the world, backed by Andreessen Horowitz and other investors. We enjoy doing what we do. That's my intro. So we have time and don't run too long.

Yeah, thanks, Phil.

Hey, everyone. My name is Sarah Tang. I'm the co-founder of SluiceBox. My background has been in building and commercializing greenfield technology, previously done so alongside projects with NASA, AeroElectronics. Prior story was featured in a Harvard Business School case study about startup to scaleups. If you want to talk to me about that, that's fun. SluiceBox is really about bringing AI to the physical world. And if you look at it, AI is not just about software, right? It's actually driving one of the biggest buildouts in hardware industry since, since the age of time. And trillions of dollars are being spent in data center buildouts, semiconductors and chips, yet actually this entire very complex and advanced industry is still super archaic. I mean, it's built on spreadsheets, it's built on disconnected suppliers, and so what SluiceBox does is actually bring product intelligence autonomous agents into connecting that entire ecosystem to enable the hardware industry, the AI hardware industry to build and sell better products.

I'm Kasim. Surprised I'm not Dakota, but I'm here to represent. Hopefully I can be just as charismatic, but we're with Harper. I'm currently leading our GTM. So Harper is one of the first, you know, companies that are, you know, building an AI-native insurance brokerage. So to the outside world, We're selling you insurance, we're working with you, we have brokers that are there selling you insurance. However, our workflows are all AI-automated and we're able to move a lot quicker than the traditional insurance broker. We specifically focus on ENS, which is excess and surplus lines of insurance. Typically, the more complex, the more we like it, right? Because it's harder to place. but we're able to work in our markets and help our customers a lot. So that's a little bit about what we're doing there. A little bit about Dakota, I'll share a little bit about him since he's driving the vision here. He was a former investor at CodeTwo, worked at Carlyle and Goldman Sachs as well, was a founder prior as well, and the reason he built Harbor is actually his family owned a brokerage when he was growing up. And his— he got super close to the problems that people were facing. Nobody really likes getting insurance, right? It's not intuitive. It's not super easy, especially for commercial insurance. And he, in fact, had an uncle who had to shut down a trucking business, not because of, you know, operational failures, because getting the insurance for that business was so difficult. So when he was looking to build his new thing, it was so clear to do this. A little bit about me, I was a former investor, so I've been on the VC side before, but worked at Goldman Sachs as well in asset management and have also had some exposure to startups prior. So—

All right.

So let's open it up and just kind of go around, you know, maybe And by the way, I should have called this out earlier, but if you haven't caught on, we've been basically going up the AE stack, right? If you remember my stack diagram, we started with compute, and we've been working our way up, and now we're at the very top. We're at the autonomous experiences. So these are the ladies and gentlemen that are taking the power of everything underneath and taking it to enterprises and helping them save a bunch of money, do things faster, better. So, so you're the ones working at that level. So, you know, let's talk about what the customer sales process is like. How do you scale your go-to-market, right?

Yeah, I'll kick it off because I'm super passionate about this topic. I constantly challenge Silicon Valley that they're stuck in the SaaS era, which means you build a great product and then you just sell it like cookies. I think those times are over. I think good go-to-market, excellence in execution, delivery is just as important as the product and the technology, especially at that level where we operate. So the thesis that we have and the way we go to market is that, um, we give our customer the entire experience from first conversation to education to delivery. We have multiple revenue streams. We don't have only ARR. We have consulting services. We have forward-deployed engineering. We charge for every single engagement that we have with a customer, and our customers buy the work. They buy the outcome. And funny enough, while I was sitting back there, I, I read a brand new thesis from Sequoia Capital around services is the new software. On every dollar spent on software in the average enterprise, there's about 7 times more services spend on the service. Means you're spending, let's say, $10,000 on QuickBooks. Most likely you have an accountant for $100,000 that closes the books. These $100,000 are up for grabs for software companies right now.

But for AI-native software.

Absolutely. Yes. So, so the way to go in there for us is we're not selling only the software, we're selling the entire package. But we are running fully AI-native, just like you said it. We're doing consulting workshops, 2-day workshops with some of the largest industrial companies in the, in the country. And our AI is preparing the workshop. Our AI is recording the workshop and our AI is vibe coding in real time during the workshop, the prototype for the customer. All the human does is read the questions that the AI tells us to ask. So we get the recordings, we take the client to dinner. That's our job now.

But the AI is doing the entire value proposition around it.

And we just try to look good and build the relationship and wow the customer because we do in 2 days what McKinsey did in 4 months before, right? So if I summarize all of that, it's like, it's so much more than the product, especially in enterprise go-to-market. And it's about being AI native and creating an amazing product, an amazing technology, but the services around it make it possible to actually have the value from all the great work that happens in compute and infra and all of these layers.

Yeah, I love that you went there with it because I know this is also a topic that both of your companies are passionate about. So yeah, Sarah, you want to share on that?

Yeah, I think in the age of AI, autonomous agents is the service, right? That is the layer. Previously in SaaS 1.0, it's like you got to onboard somebody. The interface and UX is as it is. You have to go in and learn the system. But the whole point of autonomous agents for our sector is because we started in sustainability product data carbon in the semiconductor industry, what we're replacing is actually sustainability consultants that have been hired. Elmer actually hired Amazon over $1 million for somebody that worked on a project for 18 months and couldn't even gather all the data required to actually benchmark just one component or the server and the racks within the data centers. And so the way we're thinking about the space is like to your point, is like what is the outcome that customers actually need to drive? The quantifiable ROI of that outcome, and then actually how do you create and design an agent that does all of the tasks end to end and that can actually, you know, doesn't really need much onboarding, right? And this approach we've seen with our customers actually that, you know, despite being just shy of 2 years old, we have got onboarded into some of the largest companies in the world from Infineon to NVIDIA, but also this is way less onboarding time. So I think that's part of it.

Luckily, I've had the opportunity to build some context around how Harper is tackling this problem early because I sat on the other end as an investor once, and the fund I was at was actually one of Harper's first investors.

So are we.

Yeah, exactly. And so the thing that I think is really interesting The thesis was selling software to insurance companies or insurance brokers wouldn't work, right? You have to push for adoption. You have to really make sure everyone's using your tools, and you build insights naturally slow, like slow by doing that, right? So when Dakota and Tushar were really building this company, they said, "No, we should actually become the brokerage." There's a huge advantage in doing that because one, everyone here is going to use the tools, right? We have to, and to be able to move a lot faster. And then, you know, the second part is what does the customer really want? Insurance is something that's not— it has to be top of mind to our customers, right? They might need it for you know, a contract, right? They might need it for some type of requirement or legality reasons, right? The government's actually enforcing them to have insurance and also just risk. So when we looked at that problem, we said, okay, well, what does the end user want? And they want insurance cheaper, better, and faster. So as we build and as we look at how we actually solve these problems, it's what are the bottlenecks in these, you know, with the traditional insurance broker? And the bottlenecks were very simple. It's one, how quick could they move, right? And we can move very quick.

We're nimble.

We're a small team. But also as we scale, we can keep moving quicker because of the tools that we've built.

Respond quickly. Give them a quote quickly.

Exactly. Exactly.

—Bonded, or what am I trying to say? —Bound. Yeah, exactly.

—Quickly. —Exactly. We're able to move very quick. And for us, that's, you know, a clear advantage because we know the markets already. And we built out the tools to know that these markets already. And we've done that by actually operating the business in a real way and then building out the tools based on that framework.

So interestingly, it's working so good, right? And I know Hamlet over here is helping you guys. It's working so good that the agents can't keep up. You're actually hiring more humans to deal with all of it because people can go through the agentic process, but at some point they're wanting to talk to a human, or you're finding at some points it makes sense to switch it to a human. So talk to us, and maybe all of you guys can talk about, you know, are we still— earlier one of the things I talked about is this, you know, like human in the loop, human, you you know, on the loop and human off the loop, right? So basically, it seems like you guys were out of the loop completely. You— that's how you built it, but now you've put some humans back in. And, and yeah, so talk through that as you scale. And I know you guys have big plans to become a household name insurance company. Is it going to be all agentic? Is it going to require thousands of humans? Where is it going?

Yeah, I mean, innately we have to have humans in for this. One, we might have a customer that prefers talking to humans about some of these things. But to answer your question earlier and kind of tell you how powerful this is, the typical broker in America is only able to capture about $40,000 in revenue, where, you know, for us it's about $1 million. So we're able to move a lot quicker. We're able to have a lot of context on our customer as they come in, which is great.

So we're able to keep a lot of that context. Awesome. Sarah, so you guys started with helping electronics manufacturers, chip companies, and other electronics be able to understand what was the carbon impact of every component in their device, right? In that whatever kind of part or product, small or big it is, right? Everybody has to— had to comply still has to in Europe and stuff like that. Then that led to what, a few new unlocks, right? Some new go-to-markets.

Yeah. So I think, I think the future where this is headed to is like pairing autonomous agents with like human in the loop. Right. And so what we started with is actually solving the product carbon problem for semiconductor companies because they have to report up to hyperscalers. And the steps involved there were first, you know, collecting very disaggregated, messy data that doesn't exist in ERP systems. Then it's about enriching the data because actually in the bill of materials in this industry, believe it or not, 80% of the bill of materials has missing or erroneous data. And this is what's held in ERP systems and is being used to estimate transport costs and shipping costs and pricing the product and building the product. That's a pretty huge error rate. And then the next step is actually analyzing all of this messy and disaggregated data, of which we enrich from Bill of Materials, into something that's useful for the manufacturer for public reporting and to their customers. But going through this entire process, we've actually discovered that this created a massive unlock for us because we're gathering a lot of proprietary data about how products are being built within the semiconductor industry. So I guess coming full circle, I think when you talk about, you know, how autonomous agents are helpful. Obviously, this entire process in our industry is super complex around how you build AI hardware. So you can't actually totally solve it with AI agents just yet. So the way to do it today that's different from software, again, which is fixed UX/UI, you have to go in and click buttons, it's about pairing experts on our team into every part of this process. And as you build agents, you start unlocking, I guess, Removing the coordination tax of humans, the human API across all these different systems, right? And you start enabling them to start making their own decisions. One use case that we have actually recently just launched with Western Digital, where their market cap actually has 5x since we started working with them because of how booming this industry is, but is we actually created an autonomous agent for data collection for their procurement teams. Whereas previously they had 100 or so plus of their tier 1 suppliers where their procurement teams spent 30% of their entire week on just supplier follow-ups to ask for basic information like, oh, can you fill out this form, or this is a specific template that I need to collect around costs and other data and things like that. And it took them, the first time they did this, I think 2 years to get 20 to 30% of their suppliers to actually respond fully and get all the data. We launched our agent and actually within 5 weeks got 58% of the suppliers actually completed the exercise and 78% actually responded. And it's totally via email, they don't have to log into our system, it goes into their native workflow and totally follows how an expert consultant would run that entire process. In fact, the agent could even answer questions like, oh, what is this particular material? Why do I have to respond to this? Why is this important? you know, how do I, you know, and it can answer all of those questions. And in the span of 4 to 5 weeks, there was 377 emails that were exchanged. And the supplier team were so shocked by their responses that actually some of the suppliers wrote to us and was like, can we deploy this in our organization? Right, so I think in this new age of software, if you want to call it, you all of a sudden have network effects. It's great virality. Virality, right, so to all of a sudden, you know, have a multitude of of use cases that can extend.

Awesome.

Let's open it up to questions. Great presentation. Just a question. We talk about all the benefits of agents. What are the risks that you guys are seeing as you're rolling them out to industries, like either objections or some challenges that you're seeing? I think trust and accuracy because I think the threshold for that for agents is so much lower than a human. A human, you're willing to accept mistakes and you say, oh, it's okay. Like if you explain to me, I get it. Your insurance broker make a mistake, it's okay. I like you, right? But we've seen, and this is why we actually put so much investment into ensuring not just that the accuracy threshold is really rigorous because when we report on something, when the agent responds on something, if you have an extreme outlier, all of a sudden they start, like for our bill materials, like for the chips that we're working on, there's like 14,000 line items, right? And if one thing is matched incorrectly, all of a sudden the person on the other side is like, oh no, I have to look over all 14,000 lines and see if everything else is incorrect. So I think being able to not only have improved the accuracy, it's also impossible to have 100%, right? But not only improving it, but also have some sort of flagging system that's able to surface that, but also having human in the loop that says like, hey, we're here. If something happens, like we can quality check and do things in the backend, right? So being able to figure out what the right mix of like, having humans versus AI is still really important to build that trust.

Wrong expectations are a risk for AI agents. I think some of the companies we work with, they have this utopia picture in their head, even though their reality is completely like 2015. But they have LinkedIn, they see the news, and then you get to the point where they see an agent at work and it's 70% accurate in an industrial setting, and they're disappointed. So we do a lot of like expectation management at the beginning of like, hey, AI has to learn your industry-specific, company-specific knowledge graph, has to learn. It's not going to be perfect day one.

Give it time. That resonates really well, actually. That is so true because companies also— I didn't— I thought that was only with us. They come in, they're like, but this is an agent, supposed to be 100% accurate. We're like, no, it has to be trained on your specific data because there's nuances and contextual subjectivity. Just like humans, right? And you accept that with humans, but for some reason people expect perfection with agents and we're just not there yet. And so I think that's really, really true.

Yeah. Awesome. Al, you had a question? I'm just curious because I heard—

I've worked with McKinsey and BCG before and they've completely re-engineered their business selling outcomes. Building agents and going after these companies and just launching their entire business towards agent building and launching and teaching, you know, so it's a consulting business.

So how do you compete against those giants?

How do you sustain as these things become easier to do?

I think there's two things.

One is that technology is not going to be a moat anymore. Like you're not going to be— AI is not— it's going to be democratized everywhere, right? So data will be the moat going forward. And the second thing is it's not going to be generalized systems, right? So it's going to be highly specialized vertical solutions like insurance, like semiconductor areas. And the way that we're thinking about the space is building agents, like going really hard on enterprises. We're lining a suite of Fortune 500 companies. And getting deeply embedded into the workflows and systems of those companies. And not only tackling one division, we started out with sustainability and believe it or not, we're now expanding into engineering use cases, operational use cases, like procurement use cases. So getting deeply and widely embedded within those organizations. And so you can only do that if you get hyper-focalized on a specific vertical and get hyper-focalized on the particular data. And insights of that vertical. And so that's, that's a bet that we're making on in our space.

I think it's the same for you guys. McKinsey, professional services, management consulting. The big difference between these companies and Palantir today, Palantir is product-led services. So every single thing that would do on site with a customer, every single project somehow feeds back into the core platform, into the core product. And I think for you guys, that's true. For us, that's definitely true. Every single forward deployed project we do, feeds back into Nitro, which is our proprietary platform. And that is the one thing where I say we create data that makes our models better. The autonomous agents that work in one company will work in the next company. I think that's where these companies won't, won't be as good. That's where startups have the, the advantage, like AI-native companies. All right.

Awesome. Let's give them a round of applause. Thanks, guys. And there you have it. 4 founders, 10 companies, 5 layers of the stack, all pointing towards the same horizon. Agents, compute models, robotics, longevity, the building blocks of the autonomous economy straight from the people actually creating it. If this episode resonated with you, make sure to subscribe to the Gentic Podcast so you don't miss future conversations with the founders, investors, and operators building at the forefront of this shift. And if you're a founder building in this space, especially at pre-seed or seed, we'd love to hear from you early on. At Untapped Ventures, we back AE, autonomous economy, founders early with million-dollar first checks and work closely to help you build this future. You can learn more about Untapped Ventures, explore our thesis, and pitch us at untapped.ventures. And if you found this episode valuable, share it with someone who's thinking deeply about where the autonomous economy is heading. That's it for now. Thanks so much for listening. We'll catch you in the next episode.