How Science Should Actually Work

and cut it, so.

OK, cool, actually it's a good. It's it's worth asking. ⁓ Do you do one continuous run or do you do edits along the way just in case I want to restart or something like that?

So if you want to restart, just take a pause. And then we'll just go around. Yeah, we'll just go around. And I'll post-produce the edits, so don't worry about it.

in your edit way. OK, cool. Yeah, just making sure some people like to just do it all in one run. Some people like to edit, so whatever works best for you.

Yeah, yeah, we only have one hour, so let's just go, let's just win it. Okay, sounds good. Hello everyone. Welcome back to TwoSet AI. Today I'm sitting down with Jorge Colinge. He's the co-founder of Radical AI and who's creating a new version of science with AI. Jorge, welcome.

All right, let's do it. Thank you, Angelina. great to be here.

Yeah, thank you for joining us today. You know, I looked at a company on your website, Radical AI, great name. Can you explain what Radical AI is to maybe different audiences? For instance, like an elementary schooler or maybe your investors or engineers who you might hire to work for you.

Mm-hmm. Yeah, yeah, this is great question because there is ⁓ a lot of simplicity in the mission that we are going after, but there's also a lot of complexity in how we go about it. ⁓ The most important thing is to think of us as not just an AI for material science company or an AI for science company. ⁓ What we are doing is we are building an entirely new system that both discovers and manufactures the physical world in a way that is far faster than humans could ever do on their own. In the simplest of terms, think about the world around you. The computer we're having this conversation over, the phones that we use on a regular basis, the cars that we drive every day, all of these things are made of materials. But those materials can take engineers and scientists decades to create. At Radical, what we are doing is we are using machine learning to imagine new materials and then a robotic lab to rapidly both create and test those materials. We bring those two things together into a continuous feedback loop that intelligently learns how to iterate over time and in doing so can both create materials way faster than ever before and can actually create materials that help us solve the problems that will lead to the world that we are all imagining.

Wow, that's fascinating. Do you see yourself as AI company at all?

I think we do see ourselves as an AI company. It's certainly in the name. ⁓ But first and foremost, I think the main thing about us is that we are trying to create a new version of science. And AI is a tool in the process towards getting there. But ultimately, at the end of the day, ⁓ what we're talking about is not just better predictions of materials using AI or faster experimentation. using robotics. What we're talking about is compressing 10 to 20 years of materials development into a continuous compounding ⁓ cycle. And it is our belief that if we can do that, we can actually create a system that allows us to approach the hardest problems in the world. This is not about automating science using AI or automation. This is about actually fundamentally re-architecting the way science works. by using the tools that are currently available to us. It's our perspective that if radical is successful in that journey, we will dramatically alter the course of human development.

Hmm, sounds like you're the accelerator.

In many ways, we do view ourselves as an accelerator. We are trying to speed up a process that, in our opinion, is far too slow and has existed in the same way for far too long.

What is an example of the hardest problem that we could be solving with radical AI?

There's so many problems. I this is the beauty of material science. At the end of the day, like I was mentioning earlier, so many of the things that we create right now and so many of things that we want to create in the future are ultimately rooted in materials. And so there's no shortage of hard problems within the space of materials. You can think about nuclear fusion, for example. This is something that we've been working on for 70 plus years as humans. You can think about lithium ion batteries. Those came to market. the first time in the seventies and it took 40 years for EVs to reach mass scale. ⁓ The airliners that we fly in today, all the airplanes that we take today, those are made using materials that first ⁓ emerged in the 1970s and 60s. And so all of these problems are bottlenecked by the fact that materials take too long to develop.

What, at the end of the day, what are you selling?

We sell materials. At the end of the day, what we try to do as a business is not only creating new operating model for science, but really to create a new materials company. If you look at all of the big materials companies that are out there, whether it's Dow, DuPont, 3M, BASF, all of these have been around for 100, 150 years. They all make billions of dollars in revenue. And so they're not really incentivized to go after the next generation of materials challenges.

material.

They're not incentivized to take risks. They're not incentivized to try to create new novel discoveries that might unlock the future for humans. ⁓ Our perspective is that there needs to be a new company that from the very beginning decides it is going to be our objective to actually create and bring new materials to market so that we can actually go out and solve those big problems.

I'm curious, how many years do you think you'll make radical into the 100 year old DuPont again in this new age? But not 100 years.

I certainly hope that the company outlives me and goes on to be a hundred plus year old company. I hope that it also continues to innovate all along the way. But this is something that I think is existential towards the future of humanity. And I hope that it's a company that will continue to provide benefit to how we achieve technological improvements over time for a very, very, very long time.

One of the hardest problems that you mentioned about like this airplanes, right? I'm thinking like what's the most relatable things that I could imagine that can change how we live our daily life is maybe we can fly a lot faster with new materials to like cross the globe. Maybe going to Japan will be, you know, half a day instead of like, you know, 10, 12 hours.

Yeah, no, absolutely. mean, things like hypersonic flight and supersonic flight, those, you know, there's a lot of challenges to solve in unlocking those. But one of the core things happens to be at the material level. For example, if you want to fly at Mach 5 at Mach 6, then you need materials that can survive the intense environment that they're going to be put under. These are extreme conditions, very incredibly hot. There's a lot of oxidation in the environment. so materials tend to corrode and break down. ⁓ And we don't really have access to those right now. We struggle at the material level to come up with the alloys, the ceramics, the coatings that would enable something like hypersonic flight or even supersonic flight.

Yeah, I'm looking forward to that day that travel can be faster and I don't feel like I'm on a roller coaster. Right. I also learned that you have two co-founders. How did you guys meet? Can you take me back to that moment and what happened?

Yeah, absolutely. Absolutely. Yeah, so ⁓ I have known my co-founder Joseph, who's our CEO, for a couple of years now. We actually both work together ⁓ as venture capital investors. ⁓ He was mostly focused on deep tech. I was mostly focused on software. ⁓ And in investing in software companies, I was starting to invest in a lot of machine learning companies as well. And so one of the things that started to happen, this was probably 2000, 2022, 2023 or so, lots of machine learning or AI companies were pitching us and they were telling us what they were starting to create. eventually I actually became frustrated. I became frustrated because I believed at a very deep level that machine learning was this incredibly powerful technology that was gonna change the world. And yet I kept seeing companies that I did not think were actually going after the biggest, hardest of problems, right? These were companies that were doing calendar optimization or better email copy, all things that are quite useful and we use them all today and I like having those, don't get me wrong, but they were not in the same vein as what we were hearing in terms of the magnitude of impact that AI could have. And so there was a real disconnect for me when I was seeing all of these companies. And I really wanted to get to the root of it. I wanted to understand, am I wrong? Is the technology actually not that impressive or? Am I in fact correct that technology is impressive, but it is that people are picking off low hanging fruit, the easy problems. And so what I actually did is I went down the rabbit hole and I started reading research papers, research paper after research paper. probably read north of 200 research papers within the domain of machine learning because I wanted to get to a root understanding of that question. And I concluded that, yes, the technology is incredibly powerful. had so much promise. for humanity. And then it was immediately that I thought, OK, well, then we have to create companies that are actually leveraging it to its fullest. And it was at that moment that I turned to my co-founder, Joseph, who happened to sit next to me. And I started talking to him about what I was thinking through the research I had done. And we started exploring different areas where we thought there could be large impact on machine learning. ⁓ It's actually funny. What I turned to him and said is, why are we not solving cancer with machine learning? ⁓ It seems like we should be going after those types of problems. And he said, well, I don't know about cancer, but let's look into the different spaces. And material science was one that was a natural fit. First of all, Joseph is a material scientist. And so I was able to pepper him with questions. We were able to have a real dialogue around how could you use the technology to really make an impact there? And it was through all of that that we also learned about our third co-founder, Herd Seder, ⁓ who is a leading academic in the space of material science. More specifically in the application of machine learning and robotics and automation to material science and how you could accelerate that process and really create a new method for doing materials discovery by leveraging those tools. And, you know, the three of us ultimately concluded that we knew this was the way that science was going to develop over the next decades. Whether we started a company or not, it was inevitable. Science was moving in this direction. And We knew we had to give it a shot. We knew we had to try to create a company that was going to go after that objective. And that's how Radical AI was born.

I think you are still ahead of the time, to be honest. Like just by reading 2020 to 2023 papers, I would not think like, you know, this is going to be the automation. The automation will go all the way from, from algorithms to the labs.

Right, right. There's a lot of different bets that we've taken ⁓ in the company and some of those bets pan out, some of them don't. But I think one of the things that remains fundamentally true to our mission is this belief in what science will look like. And that comes from the very, very beginning of the company. And so it is the paramount reason for why Ragley AI exists.

In having a co-founding team, I'm sure you guys build tremendous trust over the several years you worked together. Was there a moment that you knew that, you know, this is the people I'm going to trust for maybe a decade or even longer and building my career on it?

Yeah. There are several moments that I would point to ⁓ where we found this alignment. ⁓ There was an instantaneous feeling of alignment. The very first time Joseph and I spoke with her, we just knew we had spoken, by the way, to probably 30 to 40 academics before we spoke to her. And we just knew as soon as we started to have the very beginnings of that conversation that he got what we were talking about, that he understood at a very deep and profound level how we were describing this new approach to science and that he also believed it was something that needed to exist. And so that was a very instantaneous thing where we knew, OK, we're all thinking about things in the exact same way, and we all want to build a company that is going after that objective. I think between Joseph and myself, there's always been ⁓ a good relationship in the sense that we balance each other out in so many, in so many great ways. ⁓ But I think this entire journey of building Radical and over the last couple of years, having experienced, you know, bringing a startup from just the two of us in ⁓ his living room to where we are now, where we have about 37 people on the team. And if we've raised about $65 million, that trust has been built up. It has been built up over a lot of pain, a lot of ⁓ celebration ⁓ as well. But

Yeah.

a journey that we are going on together along with the rest of the team and that builds inherent trust.

Have you ever felt truly vulnerable as a founder? Like when that happened? No, every day.

Every day. an entrepreneur, think is, ⁓ especially being an entrepreneur, trying to go after a problem as big as the one that radical AI is pursuing is one of the most painful endeavors I have ever undergone. And I think it's the same for most of the people on our team. But the amazing thing about it is that people do self-subscribe to that. And so people choose to go into this endeavor despite the fact that they know it's going to be painful, despite the fact that they know there will be failure all along the way. Everyone who joins Radical.ai is so committed to what we are trying to do, to the mission that we are going after, that all of that stuff effectively doesn't matter. We are here because we believe that there needs to be a new approach towards science. and that when we are successful in creating that new approach, we will change the way humans interact with the physical world.

Give me some examples of the pains that you're going through. Are you talking with anybody? who do you talk to if you feel vulnerable of the day?

Well, the number one person I talk to is my wife always. ⁓ She is my biggest supporter. Yeah, she's my biggest supporter. She has been committed to everything that we're doing here from the very beginning. ⁓ I could not be where I am without her. Joseph and I spend a lot of time ⁓ trusting in one another and being vulnerable and open about the types of things that we going and we do lean on each other all the time as well. ⁓ But ultimately, I would say it's the broader team.

health.

We all know what we are what we are trying to achieve and in various ways we are all going through similar difficult journeys and at the same time We are all firmly rooted in this shared belief that we are trying to do something that yes is incredibly difficult and painful but when we are successful You know it yeah, the world will be different exactly and

the world.

We don't have a choice is the honest truth. think, you know, when we founded the company and we realized that this was something that we wanted to do, it was certainly a big opportunity and we wanted to pursue it. But it was also one of those moments where you feel like you don't have a choice anymore because you know that you have the potential to impact things so dramatically. You cannot say no to that. You have an obligation almost at the purest of levels as a human being. to try to make an improvement for the rest of humanity. And I think that's what everyone here believes as well. And that's why they're here and committed to ⁓ our mission.

Is some of the challenges that you're facing, is it more technical? Is it like, you know, this is gonna, even though you're accelerating it, it might take longer than you expected.

Yeah, we do face a lot of technical challenges. I mean, when you're trying to create something that is a ⁓ re-architecting of, you know, the scientific process, there are a lot of things that people have just simply never done before. And so you are discovering how to engineer things in a new way. You are trying different things that the world has told you that that won't work. That's not the right way to do it. And so continually we do face technical challenges. Our team regularly overcomes them. And I think the reason why we are able to do that is because of the ethos with which we approach our work. ⁓ We approach our work with a commitment to what we are trying to do over the long run. So a commitment to the mission. B, an understanding that because there are so many unknowns, because there are so many things that no one has ever done before, there will be many setbacks and failures along the way. ⁓ But we're not afraid of those failures. We actually proceed forward through those failures. ⁓ And the final thing is this idea that you do need to lean into risk. You do need to lean into risk in order to achieve the ultimate reward. And so leaning into risk means making quicker decisions with less information. It means ⁓ having a gut belief in not just your ability to be successful, but also to correct course when you are not successful ⁓ as well. And those are attributes that I know, we look for in all of our team members and it's something that we believe is core to becoming a part of radical.

And can you give me some example of like, you know, the ⁓ things that you're the material, the new materials that you're developing and that's getting close that will have actually an impact on our day to day life. I'm really curious, like how far away away from those?

Yeah, Yeah, absolutely. So one of the areas I'll talk about that we've been spending a decent amount of time on is ⁓ a field called hyintropialoids. And hyintropialoids is one of the ways you might actually achieve hypersonic flight, for example. And so these are materials ⁓ that are designed to be able to sustain the intense and extreme conditions that ⁓ come with hypersonic or even supersonic flight as well. ⁓ And it's a very, very complex field. So complex in fact that the entirety of academia has only ever validated about 3500 or so HGA compositions. ⁓ At Radical in just the last couple of months, we have validated a few hundred HGA compositions. Academics have been working on HGAs for about 40 years now, and so it is just an order of magnitude improvement in the way we can approach discoveries for something like an HGA. ⁓ There's a bunch of other things that we're working on at the moment, but I think it's a good example of the types of things that we can do and the pace at which we can work when we're trying to solve those big important problems.

How far away are we from hypersonic plane?

That's a question I cannot answer, unfortunately. wish I did know the answer. But I can say that from a materials perspective, we are making very direct, forward, aggressive progress towards developing materials, ⁓ first at the laboratory scale, but eventually at larger scale as well, that will become a critical component in solving the hypersonic challenge.

What's the process like? you are sounds like you're developing faster. Would you call it AGA materials? Right? Can you put that into perspectives? If you have more as your end results and what's going to happen next? Do you send them to like the the bullings or the you know, the manufacturers? What's going to happen next?

Yeah. So again, one of the core objectives of the company is to take these materials into market, right? So that we can actually solve the end problems that are out there. And this really gets at one of the problems that we think is broken with the way scientific innovations are ultimately brought into market. They're actually very, very reliant on industry players in order to carry them all the way. And this is why materials get stuck. This is why we don't have a new alloy being used inside of our airlines. airliners. ⁓ This is why we don't have the batteries ⁓ in our EVs that we would like to have. There's a bunch of different problems that we should be solving right now, and we probably could be solving them from a scientific perspective. ⁓ But there are legacy reasons that prevent us from doing so. And so our perspective at Radical is that instead of working directly hand in hand with those industry players, who again, from our opinion, are living in the past. They are still operating in a different model. ⁓ We want to create a company that is fully verticalized. We want to both create, ⁓ process, scale up, and eventually manufacture and distribute materials into the end markets that we want to go after. And so over the long run, we don't plan to work ⁓ with a contract manufacturer, for example, or any kind of downstream partners. ⁓ Yes, in the meantime, we may have some of those relationships in place. But our ambition over the long run is to completely create a new from scratch materials company that is end to end integrated and at its heart is driven by technology. We think that that is the only way we'll be able to successfully take materials into end market.

You're saying that you will create actually the potentially the hypersonic plane material and then so to the manufacturing companies making a plane.

We do want to sell it to the end customers, whether it's an aerospace company, an automotive company, an energy company, that is our plan.

And you have a ⁓ completely automated lab for actually carrying out your calculations. So you calculate the potential material composition, and you actually are realizing it in your labs.

yeah, absolutely. Yeah, so our overall system right now uses machine learning to design the new materials. And then those designs are sent down into a robotic lab that will then execute the designs that the ML comes up with. We capture data all along the way throughout that entire process. We run through about nine different tools in the lab right now. And as we capture all of that information, our human scientists ⁓ augment it with additional context. They'll label the data. they'll add some additional commentary, things like that. And ultimately that augmented data set makes its way back up to our machine learning models so they can iterate and improve over time. And we've seen dramatic improvement in what our machine learning models can do when they are given that additional context out of the experimental setting. So the robotic lab is really not just about making materials faster than ever before. That is ⁓ one of the objectives of course, is to drive throughput. The key thing to think about is that the robotic lab gives you a systematic way to capture the totality of data that lives in the experimental setting. This is the richest data set you could ever think of when it comes to materials. And so we are capturing all of that and then using it for the way we design materials.

So is, when you have the material created in your labs, does your scientist actually label them that, okay, this is a good one, this is a bad one, so that you actually have the judgment at the end?

Right. The answer is yes and no. Our AI is able to make that judgment itself. But we do have human scientists do a couple of tasks. Number one, we have human scientists create benchmarks. So human-grade benchmarks to just simply understand, are we actually improving over the human baseline? And then, like I mentioned just now, we do have humans also ⁓ do some direct analysis of the materials that are coming out of the lab as well to provide additional context to the ML models. But for example, We have models that can do automated analysis of ⁓ microscopy. We have ⁓ automated ⁓ machine learning models that can help us understand, you what's the composition of the material that we're working with? And so all of these different things are part of our overall system. Humans are involved in some aspects of bolstering that, ⁓ but we do have automated analysis going on in our lab as well.

So you are having the machine learning models predicting the results already.

Yes, we have machine learning predicting designing materials as well as analyzing the things that they went out and made.

What are you predicting exactly? is this... Go ahead.

So there are many different ways to do prediction. This is actually one of the tricky questions within within materials and you know for our purposes we predict end properties so we try to design our materials with end properties in mind. For example, how hard is this material? What's the ductility of the material? ⁓ Those are the kinds of things that we try to create when it comes to our experiments and that is driven ultimately by AI and then we run the same analysis ⁓ using AI as well. Once the material has been made by the robotic lab and we have samples and we can run them through the tools, we get out all of the characterization data and we run models on top of that to determine, we tried to make a very hard material. Now let's analyze the results to see if it in fact is actually hard. And if not, what can we learn from that so that the next one we make is closer towards our target?

Mm-hmm. That's very interesting. Can I ask what's the input to those models?

Yeah, there are so many inputs. We scan hundreds of thousands of scientific publications, both open science as well as science ⁓ that we have acquired through partnership. ⁓ We do computational modeling as well. Our ⁓ AI agents can write code, and so they can ⁓ run a bunch of ⁓ in silico experiments or simulations, for example. ⁓ And then, of course, we have our previous experimental data. that we use as part of our feedback loop. So there's a multitude of different inputs that make their way into this design of experiments. But at the very heart of it is an agentic process that can leverage all those different tools to ultimately come up with the experiment that it thinks will yield the properties that are most promising.

Mm-hmm. When you are predicting properties, right? It does sound to me more like machine learning model rather than the large language model type of AI, right? It sounds more traditional machine learning. Is it more like a regression model?

to it. Yeah, that's a great question. just to take a couple steps back through what we mean by we use machine learning to design materials. Traditionally, the ML for materials world has been direct property prediction. So it's been using machine learning models that are trained on quantum mechanical data in order to make predictions about, at the atomic level, make predictions of the material. ⁓ We use those as part of our inputs into our approach towards designing materials, but that is not the only thing that we do. And so at the core is an agent and this agent uses those quantum mechanical methods as well as the literature that I talked about, as well as some proprietary databases, some code writing abilities and so many other things in order to reason about what could be a good experiment. And very directly, we describe them as hypotheses. So our agent will craft a hypothesis for an experiment that it wants to run in the lab. It will then send that hypothesis down into the lab. The robotic lab will then execute the experiment that was designed and then send all of the feedback back up to our ML side so it can run automated analysis on that information and ultimately feed that back in as yet another input into the agentic design of experiments for a better next hypothesis.

That's very interesting. it's not just predictive modeling. It's trying to predict some material, whether, let's say, it's stable or not stable. And then you have that information. Maybe this will come out as a stable material. I'm not a material scientist. I'm just making things up. And then you go back to the agents and then try to design. Let's try to make it and then see if it satisfies your criteria. Then you can go back, send that feedback, loop back for me.

Yeah. Yeah, exactly. That's exactly right. And so basically what the agent will do is it will say, I'm going to try to create a material that has these conditions, these properties. will be hardness of this measurement. will ⁓ have ⁓ a thermodynamic property or oxidation resistance at this level. And then it will use all of the tools that are available in order to craft that hypothesis, including understanding whether or not at the structural level, this is a stable material. ⁓ So all of that, like you said, is a part of the process.

Is there any human in the loop?

We do have human in the loop. We have human in the loop again to provide some amount of augmentation to the data that comes out of the experimental setting. And there are some tools in our lab that are still human driven as well.

And is this is this is not the traditional labs are doing this right? Is this the fastest possible feedback loop that you're creating?

Yes, this is by far the fastest possible feedback loop for material science. And we believe it is the ⁓ archetype for how science should be done as a whole. We are starting in material science, but we believe that this approach and this method really does represent a new standard for how science should be done at large.

Yeah, this, think you're an engineer, software engineer, right? So, so, so thinking of like how AI is self improving itself by having these, all these agentic, know, checks and self reflections and feedback loop. You're just bringing this to the physical world. Right. ⁓ That's why it is even harder than just, you know, the, AGI of the world.

Yes. Correct? Yeah, Yeah, absolutely. I think so, look, you know. ⁓ In my opinion, there is a direct path towards leveraging machine learning to improve the digital world, and I think that's why we've seen so much improvement in software and encoding abilities and things like that. But the physical world is messier. It's incredibly complex. ⁓ It is very difficult to apply machine learning into this domain. ⁓ Ultimately, I think that's why there's so much opportunity here, though, to be honest with you. The way I think about opportunity is where are the biggest bottlenecks, the limiting factors for how humanity can move forward, can take actual big leaps forward? ⁓ That if you can solve those bottlenecks, if you can address those problems, then you're actually solving a massive opportunity. ⁓ And so I don't think there's a bigger opportunity than applying machine learning into the physical world.

Do you think hallucinations still problem in this whole pipeline?

Yeah, we do face hallucinations. Some of the models that we built in-house hallucinate. From the agentic side, there's some hallucinations. But ⁓ the models are getting better every day. ⁓ And we have ⁓ various checks and things like that in place so that we're not overwhelmingly subject to hallucinations and other things.

Do you have your own model?

We do. We use models that are available. ⁓ We use generative models that are out there. We use large language models that are provided by the Frontier Labs. But then we also create our own models as well for various tasks. For example, we have models that live inside the lab to do some of the automated analysis. And then we also have models that we use on the design side as well. So models that help us design, again, at the quantum mechanical level in a way that isn't currently available. We also have generative models that we're using in-house as well for various tasks too. So it's a combination. We use available models that are out there and we create our own models when those are not available to us.

You know, all the frontier labs are racing to build like bigger models, right? And you have, you, I mean, there are different ways in the industry and people are building smaller models, a main model like yourself. Do you think which direction is the right one?

I think it ultimately just comes down to what problems are you trying to solve. And if you can start with the problem, you have a better chance at really understanding the right solution. ⁓ And so for us, we are focused entirely on a multitude of problems that exist within the scientific domain and more specifically within material science. And that helps us make decisions around what models to use. And so we get pretty good clarity around whether or not a publicly available model, either via API or through open source, is a good one for us to use, or if in fact we have to go out and create our own. But you really have to start with the problem first. What problem are you trying to solve? Why is it an important problem? And then you can layer on the right solution.

I really like your framing about, know, this is not a problem of building bigger models. It's really about the use case and this is the product mindset. It's really right. Yeah. So, so what do think is the, the most here is, is it, is it really the models you're the main models or the AI engines or the, the, the bigger LLMs or your self-driving.

Right. Right. No. Yeah, look at the end of the day, I think it does come down to data. Where can you accumulate a large, diverse enough data set that you can then leverage in order to create the models that will help you solve problems that you have first identified? ⁓ And that's a big motivation for both creating the robotic lab. As I mentioned, it is not just about throughput, but it is also about creating a systematic way for capturing the breadth of data that lives in the lab. But it is also a big motivator for us moving deeper into the verticalization, right? We believe that there's valuable information in processing. We believe that there's valuable information in scale up. We believe that there's valuable information in how a material is ultimately manufactured. And if you can capture all of that data, then you really have the ability to, from the very beginning, design materials that will not just give you interesting properties at lab scale, but will ultimately produce materials that can scale up and be manufactured to the degree that they actually solve end problems within industry. And so ultimately that is the moat, is if we can create a system that captures the data at every step of the way and uses that as part of this learning feedback loop so that it can continuously compound on itself. we actually do have a chance at taking a 10, 20 plus year timeline and compressing it down into one to two years. And that because of that, we will actually create the innovations that the world needs.

Sounds like you are building a really large data set that's very, very valuable.

That is our objective, yes.

What is the data that do you have that nobody else on earth has today or you're building?

Yeah, so at the moment we are focused on the lab. And so that is where we're capturing most of our data. We do create our own atomistic data sets. ⁓ But for the most part, where we are focused on rich differentiated data comes out of the lab. Over the long run, we do want to move into the other physical domains beyond the laboratory setting. And that's where we think even more rich data will come from. But as of right now, the focus is on the lab.

And how's your lab data different from like a traditional research lab for material science?

Yeah, so the main thing is if you do the comparative understanding of how science is done today, you can then get an understanding of how the data is so dramatically different. And so today, science is a very manual iterative process. A human scientist comes up with a hypothesis. They ultimately design an experiment that they go run in the lab. They read a few weeks for results. They publish a paper. And they repeat that cycle over and over again. And so In that process, there is tons of valuable information that gets lost. For example, when we publish a paper, we publish on what worked, despite the fact that there are 90, 95 % of iterations that did not work. We just don't talk about those in the paper. They're not spoken about in conferences. And so that is a huge amount of negative data that never makes its way out into the world. It lives in the mind of the scientist. Beyond that, the methods for which we capture data in the lab are imperfect and they're incredibly lossy. It is a human scientist taking notes. is a human scientist, hopefully understanding what those notes look like after they've waited a few weeks for results. And so there is just so much information that gets lost in the traditional fashion that when you compare it to our approach, which is fully automated, fully systematized with computers capturing every piece of information, that you can now understand why those datasets are so dramatically different and why the second one, the radical AI dataset, is so much more valuable in the world of machine learning.

That's a really good point. Can you speak about why our failure data is so valued?

It's one of the most important things that you could capture as a scientist. Again, so much of what we do in the lab is through iteration. is through the failures that we learn how to eventually come up with good experiment. ⁓ And so you need to be able to capture those things because those are the things that help you directionally get to where you want to go. You can better understand the materials you are creating and the experiments you are crafting if you can understand the negative results. that are leading you towards the ultimate positive signal.

sense. Yeah, it's like machine learning models, right? We need negative samples and then we need positive samples and then we know the classification, how that works, right?

Absolutely. And the machine learning models, by the way, also learn from the negative signal, right? They can transfer learn from some of that stuff. They can understand where not to go as well. And so this sort of broader distribution of data, this diversity of data is actually helpful at the machine learning level as well.

Yeah, that makes total sense. Do you feel that if there is any general lesson here for founders building AI companies here, should they be thinking less about, for instance, model architecture? And what's your suggestion here?

Again, I think it all comes down to the problems you're going after. ⁓ It's not for me to say that you shouldn't spend time on architecting a better model. ⁓ If that solves your problem, you should probably do that. I would say that I think there is an overwhelming number of problems that are rooted in insufficient data. And so I think the vast majority of founders would benefit from paying attention to data a lot more closely.

I grew up, I mean, I think we all grew up this way, being told to think long-term, right? Short-term thinking is a trap. You need to think about where do you want to go for your college? What career you want to do? If you're in finance, it's harder to go to tech. Technology is moving so fast now, so it feels like long-term keeps shrinking. Do you feel there's even a long-term anymore?

⁓ Certainly, think long term ⁓ is important. think long term is non-negotiable. ⁓ You need to decide what the future will look like. And then I think from that, you can then work backwards to what are the problems that I need to solve in order to create that future. ⁓ So I don't think there's a trade off between short term and long term. I think these two things operate on different clocks and you have to accept that those are two clocks. You know, on the one hand, you have to have an opinion and a vision for where you are taking things. And then on the other hand, you need to be able to relentlessly march towards that long term vision with rapid signal that you are actually making progress all along the way. So both of them are very, very important. You know, if you if you have long term vision, but no intermediate progress towards doing that, then it's a fantasy. And if you only have long-term vision and no intermediate results, then what are you even doing? And so I think it's incredibly important to have both of those things and not in a balance, but an understanding of how they relate to one another and an understanding of what you need to do in order to both execute in the short term and make sure that those executions stack up and compound towards that long-term vision that you are pursuing.

How far out are you thinking now? And can you share one of your longer term bets?

Thank you. We think pretty long term here. ⁓ We also focus on the day to day and we are very aggressive about making sure that we move with execution and with speed towards the problems that we think are most important because they prevent us from ultimately achieving that long term vision. ⁓ But for us, know, a good example of long term thinking is again this verticalization. ⁓ We are thinking about verticalizing the company and moving into manufacturing. You know, at a time where it costs hundreds of millions of dollars to set up manufacturing facilities, we are thinking about materials that go way beyond the material classes that we are focused on right now. And at our core, again, the core principle of radical AI is the redefinition of the scientific method. In my perspective, there isn't anything that is more long term than that. We are trying to truly recreate how we approach scientific innovation because we think that when we do so, we will create a world that is far better for humans to live in.

Mm. That's amazing. That's all I mean. Yeah. It's just hard for me to imagine. Like it's not a world that I feel I can imagine. But how about you tell me, like, let's say in five years, we're now thinking, I'm not asking to ask you a hundred years, five years, what do you think the world will change would have changed because of radical or

I think you, yeah, no, I mean, we can answer that question. I think you will see ⁓ the rate at which scientific innovation at the R &D level ⁓ moves will be dramatically faster than what it looks like today. We will simply be doing way more experiments and we will be learning from those experiments with much higher fidelity than we are today. ⁓ And because of that, we will be pushing aggressively forwards ⁓ into some of the harder, more difficult scientific objectives. Some of the things that we think about today and don't really have an understanding of how to actually approach them because we don't have the right tools in place. Over the next five to 10 years, as the radical AI approach towards doing science really starts to solidify itself, we really do believe that we are not only going to make it easier to solve the problems of today, but we will actually enable humanity to start to think about the problems of tomorrow as well.

Yeah, I would love to have that mental space and while thinking that I can travel around the world within half a day. I'm waiting for that news from your company. Please, please tweet it. Like, hey, you can start flying. Don't give me four hours then.

⁓ I you that we are working relentlessly to try to make that kind of stuff a reality here.

Yeah, and also cancer research is a big one as well. I mean, I don't know if you're going to be focusing on that, but if a company's like new sign, new way of doing science companies like following your style would probably accelerate that as well.

Absolutely. Yeah. I mean, again, I think this is a part of our ethos is we are focused on material science right now, but our core belief is more fundamental. It is that science at large should be done in a different fashion. And so if more companies can take that approach, if more companies can go pursue some of big heart problems that are out there using this new method of going after scientific achievement, we think that Broadly speaking, the world will benefit in so many ways. Beyond just materials, we think we will end up in a place where humans can do more with what we have available. We can live more fulfilled lives over the long run.

Where do you think is AI's biggest opportunity? It doesn't sound like it's the software side of things. What's your opinion on that?

No, mean, like I was saying earlier, I think there's a lot of great things that are happening ⁓ in software. And I think there's nothing wrong with all of the achievements that have come out of the software world. But ⁓ in the physical world, we're talking about the largest economic ⁓ cycles. We're talking about the resources that are critical to humanity's ⁓ ⁓ livelihood. ⁓ we are talking about the types of things that will either allow us to thrive as a species or falter. And when you think about things from that perspective, you have to believe that that is where all of the opportunity exists. And again, there's nothing wrong with what's happening in the digital world. There were a lot of achievements that are coming out of software. ⁓ but when we are thinking about, well, what's going to move the needle, for humans as a race, it has to be in the physical world.

I know you're also hiring. Can you tell me more about it? What roles are you hiring for? And tell me about your culture.

Yes. Yeah, absolutely. So culture, you know, is one of the things that we spend the most amount of time on ⁓ as found. It is it is paramount to our success. ⁓ And we tell everyone at Radical, ⁓ you are not here for a job, you are here for a mission. And so every single person that works at Radical is

I can talk. I can talk to you.

committed to that mission. Culture is ultimately the way we achieve the alignment of that culture, ⁓ of that mission, I should say. And so it's incredibly important to be within our culture. And what that means for us is having courage to go after the biggest and hardest problems, despite the fact that you might fail along the way. Having belief in your own abilities, having the aptitude to really pursue things that have never been pursued before. Those are the kinds of things that would make you successful here at Radical AI. If you want to work on the biggest artist problems, if you want to have influence over direction or how things will develop, not just within this company, but beyond this company as well, into humanity at large, then you should be working here because those are the kinds of things that we will offer you. ⁓ We're hiring across the board. We have a very interdisciplinary team. We have people that come from robotics. We have people that come from mechanical engineering. We have software engineers. We have machine learning people, of course. We have material scientists. And so we are hiring across the board for all of these different roles. But the central thing is, are you committed to the mission that we are going after? If you are committed to the mission that we are going after, then the culture will be one that you subscribe to as well.

I'm convinced that science should go this way. Like, you know, the faster, the better, right? We're going to solve bigger problems, poverty, maybe food shortage, right? Things like that. ⁓ But I've got to believe because of your new way of doing things, like even for material scientists, haven't worked in a lab like yours, or engineers, AI engineer, machine learning engineer, hasn't worked in material science. What draws them to radical?

Yeah. It is the culture, you know, 80 % of our team, maybe 70 to 80 % of our team does not come from a materials background. They were not materials beforehand. And so ⁓ why are they here? It's because they believe in what we are doing. They believe that there needs to be a new way of doing science. And they believe that radicals approach has the opportunity to actually craft that and that when it is created, you will actually make things happen and you will shape the world in a way. that previously did not exist. And so that is the big driver here. Yes, we're excited by material science. Yes, we think that there are massive innovations to unlock there. But at the end of the day, the root belief is, you think that there is a new way of doing this? And do you think that way will change the world? If the answer to that is yes, then you would love working here. ⁓ And by the way, I should mention also that You know as an engineer here at radical you will weekly see physical environments of that mission come to life. I'll give you an example a couple weeks ago we did a torch test on one of our materials which is a third party test to understand you know can our material survive under extreme temperature so 3000 degrees and in that test we we discovered that our materials are actually outperforming. industry standard materials that have been around since the 60s. And so that's an early test. There's a lot more to do beyond that, but it does act as a sign that, hey, this works. What we are doing can actually change how things happen. ⁓ There is nothing more satisfying, I think, to anyone, not just material scientists, than seeing real examples of the mission that you were pursuing actually materializing.

I'm curious for that experiment, what's the reaction ⁓ from your material scientists who built that test and your algorithm engineer who built that experiment from the machine learning models?

Everyone is very, very excited. There's not a single individual ⁓ who can claim success over that. It is a team endeavor. It involved every single person ⁓ in the company in order to have some of those early signs of success. And that's really how we treat all of our wins. is about our group of people coming together to really execute on the ground floor and stack those things up in order to achieve the long-term objective. ⁓ But everyone was very, very excited with the news.

Wow, those are there like, wow, this material can stand such high temperature. I cannot breathe. That's crazy. Yeah.

Yeah. Yeah. Exactly. It's pretty incredible to see the results and the team is very proud of them.

Let's say for an ⁓ ML engineer watching this and who's thinking about where to work, what problems would they possibly be working on if they joined Radical? Can you give some examples?

Sure, yeah, we have a lot of machine learning problems within the company. ⁓ There are some that are living inside of the lab. And so again, it's being able to take a bunch of X-ray diffraction data and then make sense of that in an automated fashion, really be able to drive that automated analysis over a data regime that has previously been underexplored. ⁓ And so there's real novelty in coming up with well, what are the right approaches and what's the architecture we should be using in order to solve that problem? ⁓ We also have a lot of agentic work going on as well. And that can involve everything from AI software engineering and how do you build these multi-agentic systems down to more applied AI work as well, thinking about post-training and things like that too. So there are a number of big problems that we are working on from a machine learning perspective, everything from predictive modeling to perception, computer vision, ⁓ generative modeling, all of those things are in play here at Radical. No, we are not. We are five days a week in person and we are pretty firm on that. That is something that we believe actually enables us to work more cohesively as a team to better understand the depth of the problems. ⁓ We think that it allows us to move at greater pace ⁓ as well. And when you're going after something that is

Are you a remote company or in person?

so ambitious and so big, you really do have to be able to move very, quickly and efficiently because there are so many steps that you need to take in order to be successful. And so it's really important to us to have everyone all working together, all pursuing the same objective in unison. And so we don't do any remote work. We are all in person, all in the same office.

That makes sense. I want to see the material being burnt as well, like live and not at home and missing that opportunity.

Yeah, exactly. No, no, absolutely. And all of our people, they all have lab codes. They can walk into the lab at any point in time, and they oftentimes do.

That's very cool. And have a traditional closing question. If you could recommend a book or any type of content to, say, when you were 20 year old, or yourself today, or maybe yourself 20 years from now when radical AI is a huge success and you define a new way of doing science, what would they be? No pressure, anyone.

No, no, no, this is an awesome question. I'll give you a very quick ⁓ personal background on myself. I am an avid reader. I read a lot. Or I should say I used to read a lot. ⁓ I had to unfortunately slow down the amount of reading I can do. I have two kids at home and a third one on the way, plus starting a startup. so something had to give. And unfortunately, reading is one of those things. But ⁓ before that, I used to aim for about 24 books a year. That was my internal metric that I always went after. ⁓ So I have a number of books to recommend. The number one book I always, always recommend, and I would recommend it to myself as a 20-year-old, is a book called Essentialism. I believe the author's name is Greg McCown. ⁓ And the premise of the book is effectively that if you want to be successful and if you want to achieve the things that you set out to do, then you really need to focus. You need to ⁓ really drill into what are the most important things and then go after those things very directly. ⁓ You cannot spend time with all the different things that are going around in the adjacent areas. And it's this sort of underlying principle that we oftentimes fail not because we're not trying hard enough, but because we're actually spreading all of that effort across way too many things. And we should be focused with what we do. so that is definitely a book that I've recommended to a number of people here at radical. It's one that I would recommend to anyone who's going after a really big, hard, important problems. ⁓ if I were to recommend a book to myself right now, it would probably be the Nvidia way. ⁓ it's, about how Jensen created Nvidia and about how the culture within Nvidia, lives to this very day. just think it's full of incredible anecdotes and it really helps you understand the amount of pain and struggle it takes to create a lasting enterprise, not just a business, but a true enterprise that's really been influential in the world. And finally, thinking about myself in 20 years, I love Thomas Paine. Thomas Paine is probably one of my favorite authors. And I would probably say something from him. Yeah, Thomas Paine, think, is just a ⁓

Mm-hmm.

It's just a great thinker. The Age of Reason is probably my favorite Thomas Paine book. And I just think it's a ⁓ bastion for independent thinking and something that I hope in 20 years I continue to have. It's something that I strive to have right now. So Thomas Paine, The Age of Reason would be what I recommend for myself in 20 years.

Mm-hmm. Thank you so much. Yeah. Thank you so much for telling with me today. mean, you had an incredible journey. went from founding engineer to VC and then building, I have to say one of the most ambitious deep tech companies I've seen. founders that talk to us. And one of the things you mentioned about being stay focused sounds like resonating a lot with a lot of my engineer audience. ⁓ And I mean, you could have probably a easier lifestyle, same software, right? And then, but you chose, you chose atoms over bits. And you're betting your career on like how, you know, how science should work. And I think that takes a lot of conviction and ⁓ courage. So thank you. really appreciate you spending the time with me and being so open about how you think. And also the books. Thank you so much.

Absolutely. Yes. Thank you for having me.

Yeah, don't hand up yet. I'll stop recording. Give me one second.

Okay. No problem.