Material Abundance: Radical AI’s Closed-Loop Lab Automates Scientific Discovery

Today Joseph Krause and Jorge Colindres of Radical AI join The Cognitive Revolution to discuss their ambitious mission to revolutionize materials science by combining AI-powered discovery engines with fully autonomous robotic laboratories, exploring how they're accelerating the development of breakthrough materials like high-entropy alloys and room temperature superconductors that could unlock hypersonic flight, interplanetary travel, and other transformative technologies. Check out our sponsors: Oracle Cloud Infrastructure, Shopify. Shownotes below brought to you by Notion AI Meeting Notes - try one month for free at ⁠⁠⁠⁠⁠⁠⁠⁠⁠https://⁠⁠notion.com/lp/nathan Material Development Challenges: Current material development faces three major obstacles: costs exceeding $100 million, 10-25 year commercialization timelines, and fragmentation between academic research and corporate R&D. Business Model Approach: Radical AI determined that selling software or licensing materials would be insufficient business models. Instead, they focus on selling materials at scale for high-impact industries. Focus on Processing IP: The real IP value in material science isn't just in composition patents but in the trade secrets of scaling production efficiently and cost-effectively. Public-Private Partnership: Radical AI advocates for government involvement in material science through public-private partnerships, particularly for critical applications in defense, energy, and advanced computing. China's Material Science Approach: China establishes entire manufacturing hubs around new material discoveries to scale production, demonstrating their prioritization of materials as foundational to innovation. Culture of Productive Failure: The company embraces daily failure as a necessary part of innovation, building on these failures to achieve breakthroughs. Sponsors: Oracle Cloud Infrastructure: Oracle Cloud Infrastructure (OCI) is the next-generation cloud that delivers better performance, faster speeds, and significantly lower costs, including up to 50% less for compute, 70% for storage, and 80% for networking. Run any workload, from infrastructure to AI, in a high-availability environment and try OCI for free with zero commitment at https://oracle.com/cognitive Shopify: Shopify powers millions of businesses worldwide, handling 10% of U. S. e-commerce. With hundreds of templates, AI tools for product descriptions, and seamless marketing campaign creation, it's like having a design studio and marketing team in one. Start your $1/month trial today at https://shopify.com/cognitive

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Nathan Labenz?

  Hello and welcome back to The Cognitive Revolution. Today, my guests are Joseph Kraus and Jorge Colindres, co-founders of Radical AI, a company that's building a materials flywheel, an integrated system combining frontier AI models with almost fully autonomous laboratories to dramatically accelerate the discovery and production of new materials. As someone who spent a year as an undergraduate research assistant weighing out small amounts of fine powders and running very tedious experiments, I have long dreamed of the day when science would start to be automated. That day, I'm pleased to say, is now arriving. Radical AI's lab can run 100 experiments per day, roughly doubling both what I did

• 0:44Nathan Labenz?

  in a full year as an undergrad and what Joseph did in a full year at the Army Research Lab. And they're using that unprecedented throughput to tackle some of humanity's most important materials challenges. The upside here is truly enormous. As Joseph and Jorge explain, advanced materials are the foundation of nearly every major modern industry, from semiconductors and aerospace to energy and defense. Yet, developing new materials remains painfully slow and expensive, typically requiring over $100 million and more than a decade to go from discovery to commercialization. This creates a so-called valley of death between academic research, which focuses on fundamental understanding, and corporate R&D, which for the most part focuses on incremental improvements.

• 1:29Nathan Labenz?

  In that gap lie potentially transformative materials that could enable futuristic technology, such as floating trains, interplanetary travel, and hypersonic flight, but which nobody has successfully brought to market. Joseph and Jorge walk us through their entire closed-loop system, how they use property-driven optimization to navigate the vast space of possible materials, how their AI engine incorporates multimodal data from papers to microscopic images, how their robotic lab conducts synthesis and characterization, why they believe that capturing experimental data at scale is the key unlock, and crucially, how they're vertically integrating all the way through to manufacturing.

• 2:11Nathan Labenz?

  Along the way, we also discuss why material science has become more expensive over time, the LK99 saga and what it reveals about the challenges of synthesis and reproducibility, how the radical AI system learns from both successes and failures, ultimately building toward a sort of scientific intuition, and their recent Air Force contract to develop high entropy alloys for hypersonic applications. What struck me most about this conversation is the scale of radical AI's ambition. This is not a company trying to make paint a bit more durable. On the contrary, they are trying to create the materials that will define entirely new industries. One quick note before we get started. For those who want more on the machine learning of material science, my previous episodes with Tim Dignan and Jonathan Godwin

• 2:56Nathan Labenz?

  of Orbital Materials could be great compliments to this conversation. Today, we focus primarily on the discovery engine that radical AI is building at a system level, but those earlier episodes do go much deeper into the technical details of how AI models learn to simulate molecular dynamics and ultimately predict material properties. Of course, while the two companies could naively be viewed as competitors, I think this conversation makes it abundantly clear that there is functionally unlimited opportunity for better living through material science, and it's certainly my hope that both companies will go on to massive success. With that, I hope you enjoy this conversation about accelerating materials discovery with a vertically integrated combination of AI and automation with Joseph Kraus

• 3:40Nathan Labenz?

  and Jorge Colindres, co-founders of Radical AI. Joseph Kraus and Jorge Colindres, co-founders of Radical AI. Welcome to the cognitive revolution.

• 3:51Speaker 01

  Thank you. Thank you for having us. Shy to be here.

• 3:53Nathan Labenz?

  Yeah, likewise. So you guys are doing some really interesting frontier work in the application of AI to material science, and there's a lot to unpack. We were just joking beforehand that automating science is something that, as an undergrad research assistant, I dreamed of for many hours as I was sitting there doing the tedious work of weighing out small amounts of fine powder. And that is, I was told at the time that that was very difficult to automate and probably wouldn't happen in my lifetime. So the fact that it is starting to happen as, you know, obviously part of a bigger vision gets me excited right off the bat. For starters, you know, our audience is very interested in AI by definition,

• 4:35Nathan Labenz?

  probably mostly doesn't know a ton about material science broadly. So maybe you could kind of just set us like a baseline. Why is material science hard? You know, what's going on? One of the observations, you know, from kind of doing my homework on the company is that the cost of material science breakthroughs seems to be going up. So what's the kind of lay of the land today that you guys are entering into with a new paradigm?

• 5:01Speaker 01

  Yeah, well, luckily for you, Nathan, your dreams are about to become, if not already a reality in today's world. And so it's going to be an incredibly exciting, really next five to 10 years. And we think and fundamentally believe as a company, science will never look the same as it's going to look over the next decade. And so we're very excited to be, hope, leading that charge and building that. You know, if you think about why materials are so exciting, and I'm sure we'll get into this in the founding story of the company, but they really impact so many different industries in the world. Automotive and aerospace, manufacturing and defense, climate, energy, semiconductors, electronics, the most important industries in the world are all a direct result from materials R and D.

• 5:44Speaker 01

  But we always see really three problems. First, cost, as you called out, north of a hundred million to develop a new system from that computational framework through to really scaling that process. Second is time. Takes a long time to do a novel discovery. On average, we see typically 10 plus years to really push a novel discovery into commercialization. And sometimes it can be much longer, spanning 2025. And then three, fragmentation, probably one of the biggest problems that contributes both to the cost and time that I just alluded to. In material science, we always describe it as a company as two different paradigms.

• 6:25Speaker 01

  On one end, you kind of have this very fundamental approach to science, what we call the academic pursuit of science. It drives your understanding of science, very fundamental in nature, and usually not focused on commercial application. There can be some tie in, of course, but that's not the objective of academic research. The objective is to drive your understanding of what you're studying. And then on the other paradigm is really commercialized R and D or corporate R and D. And they're entirely focused on commercialization, which is great, entirely focused on optimization as well. And so they are driving to one, two, five percent performance gains so that they can change their margin and report that to Wall Street.

• 7:09Speaker 01

  They are not focused on discovering something like an R type superconductor, which I'm sure we'll get into. Uh, and this is the problem. In the middle is this wide open white space to do this novel fundamental discovery, but entirely focused on commercial applications. And then of course, that's exactly what we at Radical AI are trying to do. So that's where material sits today and how we view the landscape and the opportunity that's sitting before us.

• 7:35Nathan Labenz?

  Was it always that way? I mean, the obvious analogy that comes to mind for this sort of discovery is drug discovery. And, you know, broadly there, the story, I guess, seems to be kind of two parts that people typically tell. One is like, for one thing, it used to be easier because there was lower hanging fruit. And maybe that low hanging fruit has sort of been picked. And then the other side is like, well, clinical trials are really expensive. And so, you know, even once you have something that's really great, you still got a long way to go. Uh, before you can actually get it to market. And, you know, rightly or wrongly, obviously we want to be somewhat safe with that stuff at least. But safe to say that the costs have grown dramatically. Is it kind of similar in material science? Would you tell a low hanging fruit depletion story?

• 8:18Nathan Labenz?

  And is there something analogous to the clinical trial process that makes things harder than it would naively seem like it should be?

• 8:26Speaker 01

  Yeah, absolutely. There are very strong similarities and then of course, very core differences as well. You know, one important thing about materials is that we don't go into the human body. Well, most don't go into the human body. And so there is a big delineation there. There are some that actually do, and there are classes of materials that of course have to be bio-friendly. But there's a lot of similarities on search space, the size of the search space. Now, things like proteins or other things in bio are bigger, but materials are quite literally impossible to move through as well. So they're both large. And then there is this challenge on going the full scale, what we call vertically integrating across. So actually going from the novel discovery and pushing that all the way through to producing that.

• 9:10Speaker 01

  We have different steps in that process from bio or drug discovery versus materials, but they both equally take around the same amount of time. Whereas bio might be going through clinical trials, materials are going through scale up and processing. And so how do you actually take this small reaction that you run in the lab and figure out how to make tons of that material so it can be used in commercial application or industry? Again, different use cases, different processes to do that approach, but you still get this high cost, you still get this long timeline, and you still get this fragmentation, candidly, where different people are doing different parts of that process.

• 9:50Speaker 01

  Just like in drug discovery, you might have a company focused on discovery of novel drugs, and then a big pharma really thinking about, I'll pay for the clinical trial, I'll go through that process. And so you put all this process together, whether you're going through clinical trials or you're worrying about optimizing the microstructure for a novel alloy, both of these are time consuming, they are incredibly expensive to operate and go through, and they lead to this really value gap that doesn't allow a lot of discoveries to proliferate into real commercial application. And so there's a strong corollary between drug discovery and materials, and the differences, we think, are actually less explored inside materials today. It's the exact opportunity for a company like Radical.

• 10:35Nathan Labenz?

  It's interesting because I would think that we'd be getting better at this sort of scale up and process over time, right? Whereas on the drug discovery side, the clinical trial, you can sort of tell yourself a story of like, well, you know, we're becoming ever more safetyist, ever more bureaucratic, you know, everybody's got even more boxes to check on all the steps of that process. And again, maybe, you know, to some extent for good reason. But on the material side, it seems like we should be getting better. Like, is there a warp speed? You know, if there was a sufficiently great breakthrough or sufficiently great market need? Do we have the ability to be like, we're going to do this in a year? Or is there just some reason that that kind of focus and energy can't be brought to bear on the material scale up in the same way, you know, it was for COVID vaccines?

• 11:20Speaker 01

  Whether there's an answer to that question or not, I can assure you that Jorge makes the entire team aware that it needs to be done sub a year and our technology is going much faster than that. No, to your point, we, this is where we think autonomous experimentation is incredibly important. And we'll get into this in deep detail. Jorge will walk through why experimental results are so important to the AI framework and why our flywheel is so differentiated. But it's not just about changing the way we do discovery from an AI lens. Meaning, how do we recommend more things that are actually more attuned to the applications we're trying to solve for? It is those things, but we actually can just do more experimentation faster with self-driving labs today.

• 12:05Speaker 01

  And as we're doing that high throughput experimentation, we are not just running something in high throughput repeatedly over and over. We are using an active learning approach that every run that we do is incorporating the learnings from the last one, just like a human scientist would do today. And so in the lab that we are building at Radical AI, this full self-driving lab, we are targeting 100 experiments per day right now. I probably did 50 experiments per year when I was at the Army Research Lab as a materials scientist. And in some extreme cases, directed programs from agencies in the government, you might push up four or 500 experiments a year, like the MACH program, for example.

• 12:47Speaker 01

  But that's a week of research for us at Radical. And so it's this compounding effect of speeding up experimentation and then this huge unlock of the experimental data. And Jorge, I can kick that to you and kind of how relevant that becomes on the other side of the engine.

• 13:05Speaker 02

  Yeah, and the experimental data is really the big unlock, we think, to all of the big materials problems.

• 13:12Speaker 02

  But really, it goes even just beyond experiments. It actually, to your question, I think goes all the way down into scale up, into processing, and then ultimately into manufacturing. And so. To answer your question around where's this big challenge, it is quite literally in scale up. You start with at the experimental scale, a material that's the size of a quarter. And then you're trying to create a 400 pound piece of material from that thing. And at the same time, retain all of the properties, even though you're changing it in so many different ways. That's actually really, really hard to do and a different kind of challenge than what exists in drugs, for example. And so a big part of our belief is that we want to be able to capture experimental data at the research level, right? At that quarter sized level, we call them buttons.

• 13:55Speaker 02

  And then we want to be able to thread that all the way through as we continue to scale up the material to make sure that we are retaining the properties that we want. And if we're not, we're understanding why did we lose those properties? What did we introduce environmentally, chemically, or so on and so forth? To change those properties and what can we do about it in order to not change them in the future? That's a big part of the unlock that we want to go after.

• 14:17Nathan Labenz?

  So do we then have sort of the equivalent of orphan drugs in the materials world? We have a bunch of orphan materials that sort of haven't been created at this small scale, but just have not been ever scaled up for all these kind of just very practical difficulties?

• 14:34Speaker 02

  I'll let Joseph answer that as the scientist.

• 14:37Speaker 01

  Yes, I would say that's probably a bulk of the work that we do. And there's different lenses of that, right? I mean, if you look back to some of the biggest materials that have come out, they just might have taken a longer time, like a transistor, right? And the use of silicon in semiconductors, obviously one of the most talked about, most impactful, most used material discoveries today. It also took a very long time for that to really push through and continues to exponentially increase today, right? With the node sizes that we're pushing into in semiconductors. And then on the complete flip side of that, you have something like an R-top superconductor. It's not a new idea. People have been talking about it for decades.

• 15:20Speaker 01

  It is one of the more exciting material discoveries you could probably ever make. The applications of that are Really incredible, but we struggle to understand the synthesis process of that. In the most recent thought, our top superconductor, LK99, we failed to be able to make it experimentally. And that was even just going from theory and one experiment to replication. Now take exactly what Jorge said and not only be able to reproduce it at, you know, a quarter size, but then go make, in that case, millions of tons because the world would want that. At large scale, that is unbelievably complex. And so there are a lot of these things that, from everything from computational discovery

• 16:05Speaker 01

  to just making it in the lab repeatedly, go to die. And that's not even stepping through the long process of manufacturing, scale up, and then really putting it into industry. What product does that material end up in? How is it used? It's a very complex process. And to Jorge's point, when you can capture all that information and use it to design better materials, better compositions, better structures from those compositions, we think we can really 10, 100, in some cases, 300, 400x the speed of discovery today.

• 16:38Speaker 02

  This customer aspect of it is huge, you know, in our minds. And it really does get back to that white space that Joseph was talking about a couple of minutes ago. And honestly, one of the prime examples that we always point to is Gorilla Glass. There's a lot of people don't know, actually originated as a product for Corning in the 50s. And so Corning developed something called ChemCor. That was in the 50s, maybe the early 60s, and it sat on a shelf for half a century. And it wasn't until about 2007, 2006, when a guy named Steve Jobs walked over to the offices in upstate New York and said, hey, I'm developing a phone. It's going to completely change the world. All of our phones right now have a plastic screen. And I'm Apple. I care about design. I don't want that. I want something that feels better, that looks better.

• 17:23Speaker 02

  It just has these properties that I'm looking for. And Corning said, you know what? I've got a product and it's been sitting on the shelf for a while. And eventually they took ChemCor and they turned it into Gorilla Glass. And that sits in every Apple product that they've created over the last, you know, two decades. And it also sits in the products of Samsung, LG. I think it's probably the biggest revenue driver for Corning at this point. And so the reason we love this story is because it really alludes to this sort of concept of Number one, we don't really understand the materials we make a lot of the time. We don't really know what their good properties are. And number two, we actually need a customer to tell us what they're looking for in order to map up those properties to customer expectations and where value can exist. So for us, a big part of it is, again, we want to bring in those customer expectations.

• 18:06Speaker 02

  We want to bring in those value drivers very early into the process so that as we're designing materials, as we're experimenting and doing R and D, and then ultimately scaling these up and manufacturing them, we're always keeping the customer in mind and we're always satisfying their needs. That's a really important part for us.

• 18:22Speaker 03

  Hey, we'll continue our interview in a moment after a word from our sponsors.

• 18:27Nathan Labenz?

  I'm very eager to get into the flywheel, but just a couple more kind of foundational understanding questions before we do. One, on the LK99 saga, it almost sounded like you were saying that that was maybe real, but we just don't know how to reproduce it. Like, I had sort of come to the conclusion that it was a mistake, like that is some measurement error that they never actually had anything. How would you describe the field's best understanding of where that kind of landed?

• 18:57Speaker 01

  No, I think you're correct. I think, you know, the most recent thing I read was accidental creation of an impurity copper sulfide, I think, was one of the things that played out from that. However, there is this notion around, we didn't even know how to reproduce it. Things like the temperature ranges, as well as even like, what were we making it in the container? Because of course, that's going to affect the experiment or results. We're unknown, we're undisclosed. And so, it really gets into the importance of synthesis, how challenging synthesis is. And I think the best example is to think about a cook in a kitchen, right? If I give two people the same ingredients to make a loaf of bread, and one scores the

• 19:39Speaker 01

  bread and lets the starter rise and puts the right temperature and bake time, they make a really good loaf of bread at the end of it. And if someone rushes the process in 30 minutes at 1,000 degrees to speed it up, well, they're not going to make a loaf of bread. It's going to be a rock and it's going to be pretty bad to eat. And so, same exact ingredients, same exact process, I'm going to heat it, I'm going to score, right? But even just changing the time variable there and the temperature, entirely different output from that. So whether LK99 is real or not... I don't know. I don't think I have enough experimental results that we've run to determine that. What I do know is that the proposed method of synthesis and the lacking details

• 20:23Speaker 01

  from that method don't quantify as a new discovery because a new discovery needs to be reproduced. You don't have a new material until you can make it in a lab and you don't have a new commercial material until you can take what you made in a lab and scale that up. And so I think we have not done that yet in RTAP superconductors. Whether that material exists or not, I think time will tell.

• 20:45Nathan Labenz?

  Fascinating. And has the field just kind of moved on from that for the moment? Like, we're all just doing other stuff now? I mean, it seems like we would want to run that to ground collectively, right?

• 20:55Speaker 01

  The field might have moved on, but I can tell you that we think about that problem quite frequently. And when we think about the vision for Radical AI, look, we have a beachhead that we can get into. We have to, otherwise you'll boil the ocean. With material science, it impacts every industry. As an early-stage startup, you cannot do that. You must be laser-focused and really drive towards execution, then scale. But when we think about what our technology can really do, our materials flywheel is attempting to remove materials as the biggest bottleneck to our most important industries. We want to build a world where you go from a human-driven process to an AI-anatomie-driven process. And this approach allows us to create what we call enabling materials. We actually create

• 21:40Speaker 01

  the industries that we sell into. We don't just optimize the current industries today. Just like Jobs created the touch glass for, you know, smartphone and other products today, that's the exact type of applications that we want to enable in the long run, as in North Star for the company. And clearly, RTAP superconductors with their immense applications are somewhere situated on that roadmap.

• 22:08Speaker 02

  Yeah, one of the things you should understand about the company and why we founded the company is it's not because we thought there was a big opportunity in materials. I mean, that's definitely true and it's a big part of what we're going after, but we we seriously wanted to create a company that was going to move humanity's steps forward. We want to live in a world where we can ride on a floating train. We want to live in a world where we can actually travel to a different planet. We want to live in a world where you can do some of the things that we think should be basic. Like being able to drive a Tesla from New York to Boston without having to recharge it. Like, those are the kinds of things that we feel like they should be available to us and within our generation. And so that's the whole point of Radical AI. We want to make those things happen.

• 22:52Speaker 02

  But we can't because materials are constantly getting in the way. So that's the real driver for the company and a superconductor is probably from a materials perspective, the epitome of an enabler. It will create so many different things. And we like to think about the second and third order effects of what will come out of the materials we create. In the sense that we can't actually predict what will come out of a superconductor. We won't know what sorts of things it will allow. We won't know what sort of products will come out of it. We won't actually know what the world will look like when we have room temperature superconductors. But what we do know is that it will change the world and it will make us able to do things and approach technology in an unprecedented fashion. That's the whole point of Radical AI.

• 23:35Nathan Labenz?

  Could you maybe give a little bit more of a range of like?

• 23:40Nathan Labenz?

  And I'm conscious that you just said, don't know exactly how the world will look, given certain breakthroughs. Obviously, it is, you know, nobody would have immediately classic example from like the smartphone glass, right? It's like people didn't immediately know we were going to have Uber and DoorDash as soon as we could touch the screen. So I, you know, certainly I get that it's going to take time for these things to play out. I get the sense that, you know, obviously, okay, room temperature superconductor is one. How many of these sort of canonical things do we have where we're like, if we could just get this property that we don't have today, it would be like a massive step change unlock. I guess I have kind of a couple of layers to this question. What is like, what is the range of those things? You know, what specific examples, you know, are kind of forefront in mind? A layer down is like how?

• 24:26Nathan Labenz?

  How precisely do those get articulated today? Is it like, you know, this thing needs to be 4% lighter and then we cross some threshold? Or is it like lighter is better in general? And then I always think back to the Elon mass begets mass. I'm sure he didn't coin that, but that, you know, I heard him kind of go on about that. So I'm also kind of interested in like general mental models. If people just want to think like about the material world, about them and around them in a more sophisticated way. What other things are there that are along the lines of mass begets mass? Where you can just develop your intuition for like why materials matter and how they're getting better translates to material progress? So a lot there, but take your time.

• 25:07Speaker 01

  So there are a lot of ways to think about stepwise innovation, right? There are a couple of classes of materials that I would say everyone is attuned to. RTAP, superconductors, high entropy alloys, ultra-high-temperature ceramics, CMCs, semiconductor-based materials. These are, people have been thinking about these industries and the novel materials that will drive fundamental impact in all of them, if you discover. And it really, if you notice from that answer, comes down to the end application. And we think this is so important. Jorge alluded to this a little bit earlier. We don't make materials in a vacuum at Radical AI.

• 25:48Speaker 01

  Because if we do, then we actually don't know the problems that we're trying to solve. And if you're an academic research group or a national lab or a nonprofit, that could be a great approach to do that. We want to make materials that leave a fundamental impact on the human race by directly solving a huge problem that exists in the industry today or helps enable a new industry in the future. And so if you're going to take that perspective, a lot of the things that we think we can do, like the examples I just gave, are grounded in, well, what do you want to come from that? I'll give you a perfect example of the versatility to your second question. How versatile are these materials? So something like a high-entropy alloy, which is a core area of focus for Radical AI today.

• 26:30Speaker 01

  These materials have really incredible performance across mechanical properties, thermal properties, and then really their ability to blend them together. Meaning you get a really high strength ratio while still being effective at high temperature. It's not one or the other, and that's important. And these materials, these high entropy alloys, can be used in something like a hypersonic application where you are flying at speeds above Mach 5. Very high temperature ranges, very high pressure, dealing with things like oxidation inside the atmosphere that are really challenging for a material to solve. We have struggled to solve that problem as a nation today. And then in the exact same material

• 27:15Speaker 01

  can actually be used in a nuclear reactor, nuclear fusion reactor, that is actually getting bombarded with radiation from this power source. And our current materials today, like tungsten, can't withstand that radiation and erosion over long periods of time. And so these high entropy alloys are really one of the options. Amorphous metal being another option, that's a subcategory of a high entropy alloy, that can try to solve these problems. So the versatility is very big because the properties are so desirable. And these are again what we call these enabling technologies, right? These are unlocking not a 5% improvement game, right? Or even a 50% improvement game.

• 27:59Speaker 01

  These are 100x better, 50x better than what exists today. And this is what we get really excited about. And to go kind of to your third question and think about the way that these can, I guess you could say extrapolate out and get very defined. Every industry in the world can define a materials problem. If you go into the athletics industry and clothing industry, well, they're looking for novel materials to wick away sweat and reduce odor. And there are silver nanoparticles inside Lululemon workout shirts today. So that's a real application and they're trying to optimize there. And if you go to the exact other side of the dimension and look at advanced hypersonic aircraft or space travel,

• 28:41Speaker 01

  they're looking for intense advanced materials that can block extreme heat, block radiation, withstand immense pressure, all while doing that in an atmosphere that is quite corrosive to the material system. There's different problems, different defining mechanisms from them, but really important to be able to straddle all those constraints. And so in summary, the real core factor that you have to think about with optimization is what are the most important properties? And when we're defining materials today, rarely.

• 29:16Speaker 01

  Is it one property? There are probably a few examples you could think about if you wanted to. That would be a single property. We always get asked from customers. Yeah, I need it to have this mechanical performance in this environment for temperature and pressure. And at this cost with this supply chain. For an example, inside the defense or aerospace industry. And so you're not just optimizing on thermal expansion. You're not just optimizing on resistance to oxidation. You're actually optimizing on what does that cost to have that solution? And then what is the supply chain to be able to employ that solution with where these raw elements are coming from? And so we are always seeing this optimization across a multitude of different properties.

• 29:58Speaker 01

  And because of that, there become so many different combinations, so many different industries, so many different applications in those industries that these materials, like a high entropy alloy, can be useful. And so in our opinion, there is an endless amount of problems to solve inside of material science. And you can go very specific, like optimizing the silver nanoparticles in a workout shirt. And you can go very large, like building radiation blocking shields. So when we land on the surface of Mars, that doesn't become a problem for communications and electronics. Everything in between there is very, very dense. And why we think it's one of the biggest opportunities and markets in the world.

• 30:40Speaker 03

  Hey, we'll continue our interview in a moment after a word from our sponsors.

• 30:45Nathan Labenz?

  So maybe that's a great Jovanov point to talk about the flywheel. If a customer comes to you with one of these problems, we want to, you know, make a Mach 5 aircraft. And to do that, it's going to have to be, you know, no heavier or maybe even lighter and stronger and able to handle heat and maybe able to handle radiation. Who knows, maybe this thing's going out into the outer reaches of space. I guess, how, take me through the loop that you guys get into to work your way up to actually achieving that.

• 31:17Nathan Labenz?

  Maybe starting also with just like, how often is there that Steve Jobs opportunity where something maybe does exist out there? How would you even go and search? And then if it does exist, like where does that, you know, allow you to shortcut into the flywheel versus if no such thing exists and you have to come up with your hypotheses of, you know, never before seen materials and work from that point?

• 31:38Speaker 01

  Yeah, so the flywheel consists of, at a high level, two different buckets. You have an AI engine on one end and then this fully robotic self-driving lab on the other, and those work together. I'll let Jorge start on the AI front, what that looks like. It gets into a lot of how we do property-driven optimization directly to your question, and we can talk about the experiments after that.

• 32:01Speaker 02

  Yeah, so when you think of the way scientists kind of do science today, there's really a lot of different things that come into the picture. You're looking at images out of a microscope, you're reading a bunch of papers, you're using your intuition from previous experiments, you may be doing some computational modeling. That's really all of the tools in the toolset that a scientist uses to eventually come to a confidence level where they want to go out and run an experiment. And so we try to figure out how to create an AI engine that incorporates all of that, all of the different modalities from a data perspective that come into play there, all of the different architectures that we need to use at a modeling level to come into play there. And then this sort of scientific intuition aspect is another part of it that we try to capture.

• 32:46Speaker 02

  So we bring all of those things together into one piece of technology that we then kind of throw at these problems. And the way we do it is, as Joseph said, by doing what we call property-driven optimization, which is we start with the properties that we're looking for. Those typically come from a customer. So it's the things that you mentioned, right? I would like to use these sets of elements. I would like to have it hit this oxidation level. I would like to have it hit this softility or this strength. And then we give all of that to the AI engine and we prompt it and we give it some input as well, and we let it go to work. And we let it design experiments that we think are going to be successful against those properties. That's when it then hands it off to the robotic lab, which I'll let Joseph talk about.

• 33:29Speaker 01

  Yeah, and this is where, to Jorge's point, is really important. I love the way he said, Kind of how do you use the tools at your disposal? What is an experimental scientist's job? Well, you come up with a bunch of hypotheses, excuse me. You think about the ways you're going to test them, like Jorge just mentioned, all of that's happening on the AI side. And then you go and run experiments. You make things in the lab. You test the things that you made in the lab and you look at all the results. And that is all happening inside our self-driving lab today. So our lab will receive through our operating system, you know, a, a design of experiment approach. Here's the composition. Here's how we're going to make it. You know, here's what we're looking for. And it will start executing on said experiment.

• 34:14Speaker 01

  And so we will make a material. That material will come out. It'll go through material characterization. We do a bunch of different characterization techniques to understand what did we make? And really, what does it look like? Which is important to microstructure and then therefore mechanical properties. And then we'll actually run property-driven tests. How strong is the material? What is the stress strain curve of the material? How does it handle under creep or crack propagation? How hard is the material? And we will do those property tests inside the lab as well. When we collect all of this data from both the material characterization and the property characterization, we actually use machine learning to analyze all of this data in real time. So we are taking this information in.

• 34:59Speaker 01

  We train kind of these sub-modes, which Jorge can talk about, to learn how to analyze, just like a scientist would. We start to build this intuition. Start to build this understanding of if I'm optimizing for strength, then I need to pay attention to grain growth and microstructure, uh, formation. I'm going to actually look at that evolve in a scanning electron microscope or in a tensile test that I actually do on said alloy. And so as we analyze all this information, we pull out the insight from that and then feed that as active learning approach back to the AI engine, where you can think about this scientist takes back the results and says, oh, that was what the outcome was of experiment number one. Let me now run experiment number two,

• 35:45Speaker 01

  updating or tweaking that process based on the results that I just saw. And if you think about what I just described to you, that's what a human scientist does, not some insane new changing the way we do.

• 35:59Speaker 01

  The scientific method, but we are changing the way we kind of extract information from the scientific method. When I ran that process, I did all of that in a serial-based approach. I looked at one XRD scan, one SEM scan, one tensile test, thought about those, and then went back and showed the next name. And I might have remembered there was a paper I read, I might look into a patent database, or I might not. When our AI scientist looks at all that information, it is analyzing and comparing those results to millions of XRD scans. And it might be pulling from publications, it might have done past experimental results in our lab, it might have context on other experimental data sets that exist today, you

• 36:44Speaker 01

  know, and then using all of that to to make a new informed hypothesis. That operates on a dimension candidly the human brain just cannot do. And this is where the power of the flywheel becomes so impactful. It's not just an AI scientist or an AI engine, as where I described, and it's not just a self-driving lab, as I just described. It's the combination of these techniques together that build this lucrative approach to not only developing and designing, but then making and testing novel materials. That is how we can actually take a step forward in discovery and in the future, you're scaling these materials to larger systems.

• 37:22Nathan Labenz?

  Um, so yeah, I guess I want to dig into this on a bunch of different levels. Maybe for starters, like when you're developing the AIs, you kind of mentioned like sub models. I'd be interested to understand the range of models that you have. Like we're seeing increasingly stuff posted from like GPT-5 on Twitter that's like, you know, this thing helped me with think through a frontier science problem. So I wonder to what degree those sorts of things are in this flow. And then obviously there's going to be more specialized models. And there I can imagine a lot of different, you know, approaches. Obviously, we've seen many things that are sort of graph neural network structures that have some sort of radius, you know, around particular atoms.

• 38:06Nathan Labenz?

  What are they connected to and kind of model things in a more like literally physical way? I've also seen interesting structures where it's sort of a bunch of different things. And then you kind of mask out different tokens and ask the model to sort of fill in, right? So you could have, you know, here's the chemical composition and here's certain properties. And like, here's how we synthesize it and mask those things out and kind of train the model to fill in those details. So yeah, can you outline like the sort of family of different kinds of models that you're using? And like what the inputs and outputs are to those?

• 38:39Speaker 02

  Yeah, so we use a bunch of different models. We use G Ns, like you mentioned, to do atomistic modeling. We have generative models. We use language models, including the G P T family of models to do everything from understanding and extracting information out of scientific literature, even generating recommendations. We use computer vision in the lab. We use a bunch of different things in a bunch of different ways. And obviously the inputs and outputs of all those models are slightly different depending on the use case and the model itself. But I think the core thing for us is that there are so many different tasks and things that have to be done from a material science perspective, that we actually have to use different architectures and different models in different ways. But I'm happy to go into more detail on any of those.

• 39:19Nathan Labenz?

  Yeah, please do. I mean, I'm not even sure exactly what you would consider to be the most core one or a few tasks. I mean, obviously, you know, anything, there's, at least in my experience, there's always sort of one or a very small number of tasks where you're like, if we could get this working super well, we know we'd like be successful from there.

• 39:36Speaker 02

  Yeah.

• 39:37Nathan Labenz?

  And then there's a lot of auxiliary things that are sort of, you're more confident it's going to work and it's more kind of efficiency and ironing out details. What are the really core ones that are sort of the biggest challenges? What are the inputs and outputs of those?

• 39:48Speaker 02

  Yeah, so the one where we're most confident it will work is on the computational side. So that's using something like a G N. In the material science world, we call them M L I Ps. There's a lot of M L I Ps. They're all fairly successful in different ways and they work in different manners. We know that that's a core part of the strategy. We use them, a lot of people use them, have been using them for a long time. But where we would want to place our bets, where we really think there's going to be a big differentiator is actually on the language side. We think that really what we're trying to do is understand the vast space of scientific knowledge that's out there and then reason about that. Incorporate different modalities and different tools in order to really come about good conclusions so that we can generate hypotheses, design new experiments, send those

• 40:33Speaker 02

  down into the lab, learn from the success and failures of those experiments and iterate and adapt over time in order to improve the recommendations that we sent down there. So we do think that at the core, language models is really what's going to drive a lot of this scientific intuition, this reasoning about what makes a good experiment, what makes a not so great experiment over the long run. Where we think the other models will come into play is sort of bolster those models. You can imagine a GNN being tapped by an LLM, for example, in order to run a computational workload because it just needs to reaffirm some uncertainty or fill in a gap around some knowledge that it doesn't quite have, even though it's scanned thousands of papers. The same thing could be said for some analysis coming out of the lab, right? We may want to look at a bunch of SCM images and do some computer vision in order to

• 41:18Speaker 02

  update our intuition, update our understanding of what kinds of experiments we're running in the lab and how are those performing at this sort of microstructure level. So that's sort of how we use all the different models. But again, if we were to say, there's one thing we have to nail, it's this sort of replication of scientific intuition and the designing of what will ultimately become experiment.

• 41:39Nathan Labenz?

  As an aside, what's the best anyone has ever been at the human intuition on this? Like, what is the sort of ceiling of human ability? People are very obviously enamored with John von Neumann. Is there a John von Neumann? You know, I also remember this story of, I don't know if this is even true, but like supposedly the structure of benzene came to someone originally in a dream as they envisioned like a snake eating its tail or something.

• 42:03Speaker 02

  Yeah.

• 42:04Nathan Labenz?

  And they were like, oh, it's a ring structure. Okay, we've, you know, we've moved a lot past that and we've got lots more concrete knowledge to, you know, to base our intuitions on. But like, are there people in the world that are just like an obvious cut above everybody else in terms of intuiting, like what material is going to have what property or how to make it successfully when nobody else knows how to, how to do it? What does that human landscape look like?

• 42:30Speaker 02

  Yeah, it's a great question. The answer, I think, is yes. And one of the best examples I could point to is actually our third co-founder, Herb Seider. He wouldn't say this, but I think we could go out and say this. He's probably one of the world's leading material scientists in the top three, if not the very best. And when you speak to him, you just get this feeling of this guy just knows stuff that the rest of us don't know. He can just understand things and connect dots that are difficult for the average person, even a great scientist. And so, absolutely, there are people who have this intuition. But the funny thing about Herb is when you talk to him and you're sort of really probe at where does this intuition come from? He breaks it down pretty simply for you.

• 43:14Speaker 02

  He says, "Look, I've been doing this for 40 years. And so, I have run a bunch of experiments over my career. Or I have had a bunch of students run those experiments and I have learned alongside them. Or I have read papers from all of my colleagues and the experiments that they have run." And so, this scientific intuition, yes, I'm sure there's a part of it that is innate and there are some people who are just talented at this stuff and they can become better than the rest of us. But a lot of it is actually learned. It is learned from going about and doing these things and updating your understanding of what works, what doesn't work. And that really is that scientific intuition that we're trying to capture. Joseph, you probably have some thoughts on this as well.

• 43:54Speaker 01

  Well said. There's so much info in the scientific processes scientists will use that is from learned experience. Whether that's at the beginning of your career or even taking something you've worked on for 20, 30, 40 years and applying that to a new industry. How do they like to test that material? What properties do they care about? Is there a different way that they need to consider manufacturing? You know, I had spent the better part of my undergraduate and early PhD days on 2D materials for optoelectronic sensors. And then I ended up getting a fellowship at the Army Research Lab and went there and was working on normorphic computing. And to your boy, Von Neumann, I was working on getting rid of that architecture inside semiconductors. But...

• 44:36Speaker 01

  The the testing requirements, the saturation current, the on off ratio, the conductivity, the resistance, the memory states that we were building, all those were learned things. And after a year at being an ARL, now it's better than probably 99% of other scientists in thinking around how to design new semiconductor materials that were directly at norm or for computing, that were directly using ion gated channels. There was a group of people who were experts in that as well, all from intuition and experience. But exactly to Jorge's point, it comes from this learned process through experimental research, and that is very, very hard to capture.

• 45:16Speaker 01

  Actually, we don't capture of the year of work I did at ARL, 90% of the stuff I did didn't work, quote unquote, as those negative results we always talk about. But there is no recollection of that. There is no capture of that. There's no like data set that exists of here are all the E beam and and PVD and ALD experiments Joseph ran over the last 12 months. There was just one resulting material which we pushed out and shared with people. And so if you can't capture that, and the only place it exists is up here, and there's actually like no framework in place today to build this intuition, except with exactly what Jorge just really walked through. And that's what gets us so excited.

• 45:57Speaker 01

  So to Jorge's point around herd, you know, that's why, one, he is such an important part of what we're doing and why when we met him, said, we have to build this company together. But two, has this really unique ability to how do you replicate that? How do you build scientific intelligence? Intelligence that is not just general from like a, I know the English language and I know what science is, and I know the periodic table and electron orbitals and the way we actually make new elements or compositions. But how do I go deeper than that and build this intuition, this scientific intelligence around, "Yeah, I'm going to use these four elements specifically because I am targeting an X property,

• 46:37Speaker 01

  and that property requires..." And I just know that. How do you teach an AI that? It's quite complex to do.

• 46:45Speaker 02

  Yeah, and I think there's sort of two layers of this. There's like the easy aspect of it, which is, okay, yeah, science is really hard to do. It's super slow. It's very manual. We don't record it. And so we should automate that, and we should learn how to do that with AI. That's not a big leap for people to take. That's an obvious thing, and we obviously do that. I think the bigger leap here, really the one that we're after, is that is this sort of understanding that there is an exoscale set of data that's out there, but is virtually impossible to capture, at least it has been in the past. And the only way to get at it is through a really, really inefficient process, which is being a 40-year academic. And that's why we are bottlenecked by all of these different things.

• 47:29Speaker 02

  And that's why we don't have materials that solve some of these problems, right? Because not only are you stuck in doing it over a four, five-decade time span, but in that four to five-decade time span, you can really only focus on one area, maybe two areas. And so you're just limited by the amount of things that you can do on top of it being an inefficient process. And yet, like Joseph talked about, we know the data is out there. And so a big part of our belief is if we can build a system that can actually tap into that data and then figure out how to use it in a good and efficient way, we can couple that with all of the advancements that we've had in machine learning to really drive at this problem that takes 40 to 50 years and shrink that down so that it's sub a decade.

• 48:11Speaker 02

  And of course, that will translate to solving materials problems through sub a decade as well. That's a big part of our belief.

• 48:19Nathan Labenz?

  For the people that have this intuition, how sort of language grounded is it? How much can they articulate it versus how much does it live in a sort of fundamentally different space than language? Because I could imagine somebody might say, I can't really tell you why I think this is going to work. I just think that if we do it this way, it might work. Or you could have, you know, obviously a very mechanistic and clearly articulatable theory of why it's going to work. In my brief chemistry experience, it was honestly a little bit more of a former from what, I mean, and I had a limited time there and a limited view. But the one reaction that I worked on developing for a year, we never really had a clear mechanism,

• 49:01Nathan Labenz?

  like the sort of, it was a catalyzed palladium catalyzed hydrocarbon oxidation oxidation at room temperature or slightly warmer, but like mild conditions, whatever. Anyway, that, who cares? But the point was, we never really knew what was going on at the reaction center. We had kind of vague sense of that. And so it was a very kind of, yeah, maybe we'll try this, maybe we'll try that, you know, but without like a, without any precise articulation.

• 49:25Speaker 01

  Yeah.

• 49:26Nathan Labenz?

  How common is that? And what does that imply for whether like GPT-5 is going to be able to sort of, you know, chain of thought its way into making these things happen versus whether there's just kind of different modalities of data that have to be, you know, collected and learned from?

• 49:42Speaker 01

  Yeah, you took the, one of the answers there, which is modalities of data. It's very important and we'll talk about, J C can go into deep detail around where that comes into AI engine and why it's very, very important. Language is an important piece. It's not just language, as you just alluded to. You know, scientists learn a lot from the different data streams that they're pulling inside their experimental process. There's this one other thing that is a form of data, but is not maybe what you call structured data from an experimental tool or from an output in the lab. And that is this.

• 50:18Speaker 01

  I guess intuition, for lack of a better word, or experience you have in running experiments. I'll give you a perfect example. When I was working on these 2D semiconductor materials, we would always look for the transparency of the film. And the technical reason why that we later came to figure out was they were monolayer 2D TMD materials, not multilayer 2D TMD materials, and that drove way better conductivity, etc. We didn't know that upfront. And so you could see this in an optical microscope. We weren't looking for this in an optical microscope. We weren't taking pictures of transparency. And hmm, you know, what were the changes in transparency? And so we kind of were racking our brain around like, why are we getting such better conductivity

• 51:03Speaker 01

  in these materials? They're the same material and we're making them the same way. And it actually came from an expert in the 2D material field who said, oh, those are monolayer films. And we were like, what? Like, we know what monolayer films are. I worked on monolayer films for three years, but never in that setting have we thought about looking for monolayer based films. And so this randomness, this accidental procedure that you hear all about in famous science stories is really important to capture this negative result based process. And that is almost like a separate form of data where we deeply believe it's not just making the things that work. It's not just using the intuition from multimodal data

• 51:47Speaker 01

  sets that actually allow you to predict new things. It is looking at exploring and capturing the one-off, random, unexpected outcomes that might have a key insight in that a human scientist will look right past, but an AI engine would never look past. There's a very important piece of what is, I guess you could call intuition as well. But on the multimodal data, that is also equally as important. And there's a huge kind of approach we take on that side.

• 52:17Speaker 02

  Yeah, the multimodal aspect of it is just enormous because it gets at this idea that the data is out there. It's just in so many different ways and forms and shapes and it's so unorganized. And if you can bring a little bit of order to it and a little bit of cleanliness to it, then you can start to really enrich your representations and really get to better predictions and generations over time. But I do want to go back to what Joseph talked about because this is, I think, something that's actually quite hard to do with machine learning, right? When you have models that effectively learn out of a distribution of data, it becomes very, very challenging to have it do things that are out of distribution, which is effectively what Joseph is talking about. How do you go about this randomness when the entire purpose of the model is to learn

• 53:01Speaker 02

  from a distribution where those random things are generally just ignored by the model, right? And they're sort of said like, now that's out of distribution and I'm not going to fear in that direction. And so we think about being very purposeful with that. We think about saying, look, in science, it's a little bit different. We have to do the sort of normal thing that we always do when we have to learn from a distribution of data. And if we can improve that data over time and better enrich the modalities of data that come into that, we'll be good. But that's not entirely true. We have to also design systems that are effectively prized towards novelty and surprise because that is really where the biggest discoveries in science come from. They come from some weird corner of the chemical universe that everyone thought was kind of ugly and they didn't want to look at. And low and behold, here's this amazing scientific discovery.

• 53:45Speaker 02

  Or pure accident sometimes. And so we have to design systems that not only learn really well from good distributions of data, but also figure out how to basically go hunch hunting, right? How to develop hunches and then say, you know what, I'm going to go out and pursue that thing. I'm actually going to go deeper here. And there's a bunch of different ways we think about doing that. You could do some sort of randomization or diversification, which we do through Bayesian approaches and Monte Carlo and stuff like that you can do. You can do noise sampling. Active learning is a huge part of this for us. But then there's another side of it, which is, again, the underlying design of the system itself really has to be geared towards capturing scientific intuition at the end of the day

• 54:30Speaker 02

  because that is such an important part of it. And so it's what would drive a scientist to pursue something that is a little bit weird, but just like Joseph said, at 10 or 20x or even 100x the scale. That's really, really important to us. And so. We have a bunch of techniques that we use for that as well, reinforcement learning and things like that. Human in the loop is a big part of it as well in the early days for us in sort of capturing this un-capturable data like Joseph was talking about too. So there's a bunch of different techniques that we think are really, really important that go beyond just the normal, hey, train a model on a good distribution of data.

• 55:07Nathan Labenz?

  Yeah, that's really fascinating. I mean, that's also basically the core problem in all of AI, I would say in some sense, right? Is like, if for better or worse, you know, are we going to take these things past the point where they modeled the past? And into the point where, you know, into the realm where they can sort of figure out the future. Yeah, I mean, I'd love to hear more about that in kind of any way you want to go deeper on it. But one, to try to map a little bit of what you have experimented with and learned onto the fully general problem. How do you see like base models, so to speak, versus like RLHF versus just pure success signal RL?

• 55:47Nathan Labenz?

  Kind of fitting into that because it seems like people are, that middle ground of RLHF is pretty good for the human assistant. And there's part of me that's concerned with like big picture, let's say, AI safety questions. It's like, man, as much as we can get the things to like imitate us, I'd really like that. Because I do worry about, you know, if we just go super hard on RL and give nothing but reward signal. What are these things going to learn and what, you know, reward hacks might we find ourselves encountering? But the flip side, I would guess is that if you just sort of imitate, you know, humans too much, or just get the thumbs up from the human evaluator too much, you maybe sand down a lot of these rough edges. That certainly seems to be what we see from the chatbots, right?

• 56:28Nathan Labenz?

  People are like, the base model is more creative. I like that more in some ways, especially if I want to get, you know, kind of outside the box ideas. The RLHF kind of gets boring and then the RL again gets interesting, but sometimes in like kind of scary ways, depending on exactly, you know, what sort of problem you've put in front of it. How would you compare and contrast your learning against that stylized story from the?

• 56:51Speaker 02

  Yeah.

• 56:51Nathan Labenz?

  The chatbot space.

• 56:52Speaker 02

  You know, I think a lot of that is definitely true. For us, a big part of it is, is do you have good evals? At the end of the day, if you can develop good evals, then that helps you go about this process. Because a lot of it is undiscovered and a lot of it is true research on the ML side. And trying to understand how to bring these things together and move them in a direction that you want to over time with, like you said, not forgetting about safety and some of the other things that come into play as well, which obviously in material science and in chemistry is an important thing. So evals is hugely important and for us, there honestly is no better eval than the lab. I mean, that is as ground truth as it gets, right? If you're looking for a good benchmark, a scientific lab is pretty good as a benchmark. It's a really good way to understand, well, did my model actually successfully do what

• 57:37Speaker 02

  it was supposed to do? And did it do so in a way that didn't come at the expense or the sacrifice of some of these other key things that I really wanted to hold? So that's a really, really important part of it for us. And it brings us back to, again, some of these active learning loops that we were talking about earlier. And how do we smartly incorporate that so that we can update these models? Not just kind of like at the weight level, because it needs to go beyond that. It needs to be a much more dynamic process than, oh, we're just going to retrain the model. There really needs to be a sort of like surrogate or agentic process that sits on top of it and says, yeah, I'm going to use my model to kind of get things going and hopefully we're going to improve that model over the long run. But in the meantime, I need to be able to be adaptive and responsive to the new learnings that are coming back out of, again, what is a best-in-class benchmark or a best-in-class

• 58:21Speaker 02

  eval, AKA the scientific lab. So that's a big part of it for us. And I think the other thing that I would bring up is that we do truly believe in trying to understand how these models work. We're making a ton of investments in mechanistic interpretability, for example, because we think that's going to drive a lot of performance and a lot of gain. And the correct steering of these models towards the objectives that we have. And again, if you think about what Radical AI is trying to do and the core purpose of the company, it's to solve really big materials problems for people. And so if our models are just not going in the direction we need them to. We need to be able to move them in that direction. If they're doing or against the expectations or the wishes of an end application or an end customer, then we need to be able to curtail those things, too.

• 59:04Speaker 02

  And so it goes a lot further for us than just, well, give it to the model and let it do its thing. And, you know, it's okay if it's a black box and we'll just learn how to improve that black box over time. We really need to be able to use it as a tool over the long run if we want to solve some of these big, hard problems.

• 59:21Speaker 01

  Especially going back to one of the things that you brought up earlier, like this is the nexus. So this is the one of the big problems in AI. You know, coupling on top of that, that we don't have historical data sets to tap into. Like even in the English language, you can like digitalize, you know, old books, right? And bring them in and learn about how Shakespeare wrote and, and kind of build intuition or, or knowledge off of that. We don't have Newton's notebook. And even if we did, there'd be scratches of all the things that they tried to do. There wouldn't be a written text that we can directly learn and, and transcribe. And so coupling on top of that problem is this kind of idea that not only do we have to know where they need to go,

• 1:00:06Speaker 01

  but like, what can we bring in so that they know the different directions that they can go? Cause we don't have a different text of the way we write books today and the way that Shakespeare wrote books back in the day. So if you don't have that intuition, you don't even have that understanding. It becomes really hard to think about how to index that. And I think you've seen this in software, right? Like where a lot of the best software is private. So a lot of the coding agents use what's open source to train on. And that's gotten really good because there is an open source community that pushes really good software. And so it can learn what really good software looks like and then utilize that in an effective way. There is no open source science community. There's open source science and publication.

• 1:00:49Speaker 01

  I mean that there is no multimillion data point experimental set of, oh, these are all the scientists who have subscribed into, have opted into publishing their lab notebooks. We don't even have that, let alone structured data, like pulling a GitHub repo and being able to learn from, okay, here's how professional engineers write code, here's how it's organized. We can't do that. So if you layer that on top of exactly what Jorge just talked about, it is really important to know which way our models are going so that we can get in there and kind of like bring a scientist in to trying to drive them to better discoveries. Cause exactly like Jorge said, our objective is better discoveries that fundamentally impact

• 1:01:31Speaker 01

  humanity. It's not just about, you know, producing or recommending new things that are marginally an improvement. That's not North Star for the company.

• 1:01:40Nathan Labenz?

  So I definitely want to get into the mechanistic side in a second. Shout out to Eric, CEO of Goodfire, who originally introduced us. But maybe first on data. You've said a couple of times, like data is the limiting factor. And you know, the big three, the holy trinity, right, of AI progress is data, compute and algorithms. I was struck in looking at some of the specialized models by just how small they are, like 30 some million parameters in some cases, like, you know, orders of magnitude, smaller four, four orders of magnitude, you know, smaller than some of the real big ones, even five, maybe. That's a lot of orders of magnitude. So that would, you know, reflect, I guess, obviously that data must be scarce. Otherwise, I would assume you'd have a bigger model.

• 1:02:24Nathan Labenz?

  Where, where does that future data come from? I mean, you said sort of like, you know, we don't have the history. I was kind of wondering, I think of an analogy a little bit also to the computer use agents. I've remarked many times and still nobody has come forward to pay me to watch me use my computer. So I'm like, why is nobody just offering me a hundred bucks a month or whatever to install some observer on my computer? And I kind of wonder, you know, and there could be different exponent, but it's too messy or whatever. It's not, it's maybe not worth the trouble. So sources of data I had in mind for you guys were, one, literally going and paying grad students to let you like, you know, digitize their notebooks. I don't know what the going rate for that would be. Grad students, you know, would probably be interested in the money, I would think. I don't know if

• 1:03:07Nathan Labenz?

  there's, you know, IP concerns there or whatever. So, you know, has that happened? Why not? Then there's just the straight simulation data, which I understand is at the core of a lot of sort of at least filling in gaps for some of these like small kind of force prediction. What's the next time step in the evolution of a collection of atoms kind of model? That would be one way you could like invest heavily in compute and get data, but I'm not sure what gaps that may leave or like why that where that falls short. And then obviously the lab is kind of the the final stop. I mean, I'm even missing categories. What do you think the portfolio is of contribution from those different types of data going forward?

• 1:03:47Nathan Labenz?

  Are you ultimately kind of like, our progress is going to be rate limited by how many of these autonomous factories? We can build and how many experiments we can run per day? Or is there any other way to convert like financial resources into progress? That doesn't go through that sort of factory build out.

• 1:04:05Speaker 02

  Yeah, I mean, data is always going to be the limiting factor. Certainly in our field where we're so data limited, but I would, I would even argue generally speaking across all areas of machine learning, data is going to be your biggest driver of success. And if you can figure out a way to get over the hurdle on the data side, then you're going to be in a good, in a good spot. Yeah, improvements in models will help and certainly access to compute and being able to do more with compute is a big part of it, too. But if you asked anyone to choose between each of those three, they would almost universally say, yeah, give me data. Give me data at the end of the day, and the same would be said for us. Joseph can probably answer the mix of data and where we think that will come into play for us.

• 1:04:46Speaker 01

  Yeah. It's even harder than just a mix of data, too, because it's all unstructured. It's not even labeled, actually. Like, if we could get access to every lab notebook in North America today, that would be great. We'll take it, by the way, if anyone off that we will, we will take it because to Harry's exact point, we'll always take data. And then we probably have to go hire an entire team to sit down with everyone who wrote those notes and understand what they wrote and what is it in reference to. I mean, even generalized, simple things like running an X-ray diffraction tool. Like the peaks mean things and people might interpret them differently for their research than we interpret them. When we're doing amorphous based materials, we're using XRD as a

• 1:05:31Speaker 01

  Confirmation that there is no crystal structure. We're actually not even looking for relevant data. So if you gave that to someone who had never worked in amorphous materials, they'd probably be like, I don't, there's no peaks here. I don't, there's, I don't get it. Right? Like, like, this is irrelevant to me. I can't train a model on learning XRD. It's like, but you can train one on learning amorphous behavior. Because there isn't this, you know, of course, crystal structure. So maybe that's like a very technical example, but it's very important. It's like, even if we had access to all of this, it's not labeled, it's not organized, it's not structured. It's interpreted in wildly different ways based on the discipline that you're working within. And most of it is like only a piece of the data.

• 1:06:14Speaker 01

  And for example, I used to always write down the timestamp of halfway through my reaction. Cause it didn't matter when I started and I would always wind down temperature. I would like turn the temp knob off at halfway. And then I couldn't control the cooling. So however long it took to cool to the end of experiment was irrelevant. So I gave you a timestamp, one minute, one minute and six seconds, one minute and four seconds, 58 seconds. What does that mean for us doing the same experiments in our lab? Now, those might be relevant metrics. They might not. If I sit there and tell you, oh, that's the halfway point and I cut temperature exactly at that point. Well, now, yes, of course, that data becomes more relevant to what we're trying to do. But so this just this, I, this challenge in materials is not only is it multimodal,

• 1:07:00Speaker 01

  not only is it not captured anywhere, not only is it really important to, as Jorge said, evals, ground truth, benchmarks. We don't even have it structured and annotated to be able to use it, even if all three of those things I just mentioned existed. And so this data problem, I think, will actually spur up an entire industry. We have a lot of ideas on the way to get around that, which we can dive into. And from a scientist's side of understanding the data output I produce as a material scientist, that in and of itself is quite a big challenge inside and out.

• 1:07:35Nathan Labenz?

  Yeah, on the computer use side, I think maybe one of the reasons that it hasn't been done is sort of analogous to what you're describing, which is like, at any given time, it's not immediately clear what I'm trying to even do as I'm using my computer. So there would have to be a sort of, or probably would need to be, you could guess, obviously, but you would be greatly benefited if you could also capture the sort of, you know, internal monologue that is going on when the person's like

• 1:08:01Speaker 01

  Exactly.

• 1:08:02Nathan Labenz?

  Going tab to tab or whatever, like, what is it that you're even trying to do here? So I take the point that that is generally, and I can think back to my own, you know, not super well documented or super tidy or thorough notebook notes myself to

• 1:08:15Speaker 01

  And how do you judge quality there, right? If my grandmother records her use of a computer and you record her, I hope they're gonna be different. I think they're gonna be different. How do you delineate that? What sets the bar of not great computer user and great computer user? Now bringing that into a subjective field of science, was I a good scientist or bad? Well, my boss and the people I used to work with think I was great. Other scientists in the field might think I was terrible. I would have not tried those experiments. I would have went in a way different direction, would have never used Malibden, would have used tungsten instead, or whatever the property is. So who sets that bar? Who determines how it moves?

• 1:08:56Speaker 01

  How do you even begin to build it when there are so many different fields inside material science? So every single bar or benchmark is different. There's like these deep problems that come up in even just annotating the data that are really hard to answer. And like in the computer example, I don't know how I would answer that. I would generally start to say, well, someone who's very proficient at moving around and understands where all their applications are and, you know, can get their tasks done quickly. But what if someone goes to email and and opens a new email, pops it out into a new tab, goes to somewhere, writes it in a notes app, copies that and then pasted in Gmail's email box? It's like, okay, well, I wouldn't do that.

• 1:09:40Speaker 01

  Right? Like, that's not the most efficient way to write an email. So how do you delineate that? There's probably someone who would argue it is. Oh, I make too many errors when I send emails. I have to write it in notes with grammatically installed and check it first. Who's right? So this is so much ambiguity inside classifying, or I guess just like labeling the data too. It just becomes really an exponential problem.

• 1:10:02Speaker 02

  Yeah, I think I think what we're hitting on is that it's pretty abundantly clear that the highest value data comes out of an experimental setting, out of a lab. And yet, it's also where the biggest challenge lies, right? Just getting access to the data and doing it in a way that's orchestrated and clean is really, really challenging, not to mention expensive. And that's why we just, we don't have the data. We have not seen people really build these data sets out because it is so hard to do. And actually, if you flip to the other side of the equation, the computational side, well, that's actually a lot more straightforward. And we've seen major, major push to develop data sets there. But our big thought is, look, that only solves like 10% of the problem. You're not there. And it's because of this intuition again.

• 1:10:46Speaker 02

  Well, we know that the highest value is in the lab anyways. It's just really hard to get to. And so, yes, computational data is important and it's helpful to us and we're going to do that too. But we're kind of missing the real big picture here, which is let's figure out the lab. Let's figure out how to get experimental data and incorporate it alongside the computational stuff that we're already getting pretty good at. And that is really going to push things forward. That that's really what we're trying to do here.

• 1:11:11Speaker 01

  And if you do do that, worth adding on, you actually control the problems I just brought up. It's no longer unstructured. It no longer lacks context. It's no longer unlabeled. All of those things become a reality. And so not only are we building the data set, but we actually can use the data set. And those two things are incredibly effective, in our opinion, at really developing new materials, to Jorge's point.

• 1:11:33Nathan Labenz?

  Yeah, that's really interesting. You mentioned earlier, the search space is vast. Could you maybe help me develop my intuition for the vastness of the search space? Like, how many, you know, different dimensions? Are there sort of big branches in it that are kind of, you know, just critical to understand forks in the road as you're going down a thought process, or ultimately, obviously, a material development process? And then like, how much of that, in terms of, you know, as you build out the capability of the lab automation, how much of that space can you effectively explore today? Are there parts of it that are like, we can't automate that yet? That's still kind of, you know, because we can't automate plumbing in my 100-year-old house, right? So, and again, I remember doing chemistry stuff, like,

• 1:12:18Nathan Labenz?

  one of the things I observed in my grad student mentor was like, this dude just knows how to do certain things, you know, in a very intuitive way that I don't know, and that certainly at that time, no robot had. So as you think about pushing the frontier of what you can automate experimentally, some of it is just horizontal build-out at some point, right? To do more throughput. But presumably, there's also frontiers of like, certain kinds of experiments or reactions, or what have you, that, you know, are just not yet automatable, that become automatable? So again, take as long as you want. How would you describe the vastness of this space? And what parts of it can we explore? Or can't we explore? What, you know, what frontiers do you think you'll be pushing the automation forward

• 1:13:02Nathan Labenz?

  to actually be able to get into and properly explore?

• 1:13:06Speaker 01

  Yes, it is an incredibly vast search space. It is impossible, it is improbable to think you're going to test every single potential combination. We always refer to, when we're talking about the company, to Eddington's number, I believe it's called, 10 to the 80 different observable atoms in the universe. And that is the number we use for the potential combinations. If you go into something like alloys specifically, which is of course, our area. You know, there are 60 or so, give or take, elements on the periodic table that you're going to use in novel alloy development. And so if you just, you know, and when we talk about high entropy alloys, we're doing five or six element systems. They're called multi-component alloy systems. If you just take five elements and, you know, use that,

• 1:13:50Speaker 01

  it's like five and around 5 million and change, and potential combinations that you can develop. And that's if they're all equally balanced. So meaning 20, 20, 20, 20, 20. Uh, uh, uh, percentages inside that alloy. No alloy, to my knowledge, is, is that equally balanced. And so within each individual element, of course, you have a zero to one hundred percent. And then even that many decimal places you want to go into optimized. And so it is massive. The way we always describe it, it is a single grain of sand on an incredibly large beach. A beach that's the size of the earth.

• 1:14:29Speaker 01

  That's what you're really trying to identify um and find and looking for a new alloy or, or a new material. Now there are spaces that can be a little bit bigger than that. I think proteins are one to call out that everyone tracks. Everyone thinks about how big they are. I think it's something like ten to the hundred and thirty, I think, um, that the, the search space is there. That's probably just a single grain across multiple earths or multiple universes even maybe that's huge. We think that's a huge advantage for materials. It's actually more likely that we'll find hopefully something in our, in our beach versus the bigger beach. But it is really big. And so when we think about, okay, how do we take that big number? And then we actually start testing things and actually get to, okay, how do you drive towards discovering now?

• 1:15:12Speaker 01

  Well, the first is indexing the search space. And all of that sits within the AI engine. And how do we actually do property driven optimization to weed down and filter out? Well, the best materials that we want to try. And then when we think about the challenge on the experimental side, it's really the material tools themselves. It is not robotics or different automation systems like our rover truck that we build. That is our materials handling system. It's the tooling themselves that we need to conduct scientific experiments. Those companies have been around for a very long time. You know, some twenty five, some fifty, some seventy five, a hundred years.

• 1:15:54Speaker 01

  They have been human operated for decades. So they're built entirely around a human scientist with no automation in consideration. And the ability to then drive those tools with software are, in some cases, nonexistent. Or in the cases that is existent, there's some type of interface, very, very limited in the actuation you can drive. You might be able to like, get outputs from the tool, but you can't control. You can't control the turbo pump, you can't control the heat source, you can't control the vacuum chamber and like the pressure that's inside the chamber. Therefore, rendering an effective tool from an automation standpoint.

• 1:16:38Speaker 01

  So a lot of our time on the automation front is spent on how do we take existing tooling infrastructure? And make them autonomous, make the tool itself be able to operate in an autonomous fashion. We do a lot of custom development ourselves. We have multiple tools in the lab that are entirely custom designed around the material tool to make it automated. We work with companies. We have multiple ongoing conversations with companies where we want to build a custom tool together. So that's taking that outside automation and bringing it into the real engineering and design of the tool. And then there are some tools that we think will probably be pushed out from its lack of automation

• 1:17:19Speaker 01

  or from the ground up by companies because the way that they were built was just entirely human. And we have examples in our lab today where they're built in a very, very mechanical way, reading humans flip switches and move knobs and turn dials. Those things we think will be replaced over the next 10 years in these new tools that come out. And so from an experimental setting, a lot of our effort is spent in taking the existing tools today and making them automated. Remember, Radical is not a material tooling company. We don't want to sell XRDs, we want to sell materials. And so we work with then companies and vendors and do a lot of that work ourselves. So that's how we do it on the experimental side.

• 1:18:03Speaker 01

  On the AI side, the search question is challenging because we do not do a combinatorial-based approach. We really don't think combinatorial-based science has driven enough value. Again, we move towards this property-driven optimization that has this active learning component and is using a bunch of different tools and generating there. Jorge can kind of go into the details if you want there, but we really kind of take a novel approach, in our opinion, to generating new systems to drive.

• 1:18:31Speaker 02

  The big aspect of it is definitely the active learning side of it, but you know, even in this sort of universe of, hey, we have dozens of elements that we might want to use and how do we go about doing that, some of that just goes away, right? Because, for example, if I really, really care about weight, I'm not going to use a little bit. So you could just remove a little bit. So some of these things do just start to naturally shrink down. You still end up with a very, very, very large number. And this is why you need to have multiple machine learning models that can solve multiple problems kind of all along that trajectory so that by the time you get to the active learning loop, you're kind of sending really good stuff down into the lab. And yes, you're learning how to make them in the right way, but hopefully the things that you send down to the lab are already good to begin with.

• 1:19:18Nathan Labenz?

  One more brief follow-up on that, and then there's so many different interesting directions to go. Are there things right now where like you have to either patch with a human? It sounds like, I know like Emerald Cloud Lab, from what I understand, has certain aspects of the process where it's like a person is told like, go to this machine, pick up this file, move it to this other machine, put it in, and they're almost sort of Chinese rooming it to a degree with a person who doesn't necessarily even have to understand what's going on, but it's just following like very point-level instructions?

• 1:19:48Speaker 01

  Yep.

• 1:19:49Nathan Labenz?

  Are there things like that that you are kind of bridging missing tools? And are there still sort of certain kinds of things that you just can't get to that a purely artisanal process like could do that are just like outside the bounds of what you can bring into the closed loop as it exists today?

• 1:20:09Speaker 01

  The answer to that is yes today, meaning there are some things we don't have fully automated. So we do have scientists in the lab who help orchestrate experiments. But the answer is also no in that while they end at that solution, that's our, that's our starting point. We are actually beginning our learning process from there. And of all the tools we have in the lab, everyone starts at the place I just described. Human scientists coming in, showing you how you operate at running it and moves to fully autonomous, no human in the way. Inside running that tool specifically. And so for all of our tools, we will reach that place. And if a tool cannot, or we think the value is not there to go through, you know, making it automated.

• 1:20:54Speaker 01

  We will go work with someone else to custom build a tool specifically to our needs. Now, I don't mean custom build the science part of the tool. Actually, we usually don't touch the science part of the tool. How the tool operates, we don't get into that. We are touching the loading chamber, the control of the vacuum, the control of when the laser turns on and when it turns off. The kind of shuffling of samples and prevention of contamination if there's multiple samples inside a chamber. All mechanically driven, optimization focused, or I guess you could say autonomy focused functions of the tool. We are not redesigning how to do X-ray diffraction spectroscopy.

• 1:21:39Speaker 01

  That's not, we don't do that as a company. We don't want to do that. Again, we're not a material tools company. We are material discovery and then therefore material manufacturing company. And that's what we want to move towards. The way that we can get there kind of can look different ways and we are going through that process with all of our tools today. But no, we don't have every single tool we've ever bought, fully autonomous today. There are some tools that are going through that custom design process right now and retrofitting process right now. That aren't just fully automated yet.

• 1:22:09Nathan Labenz?

  Cool. On, I heard you mentioned the idea of like, hopefully what you send down to the lab is good. Couple of questions on that. One, is there a sort of inference time scaling law that you can tap into where you're sort of like? If we run 10 times more experiments or in Silico 10 times more, and I guess a lot of different ways to think about it, we have this desired property, we have this sort of constraint around inputs, generate me a bunch of protocols, or generate me like the extra, what am I going to dope it with, or whatever, right, like you could have like a lot of different guesses. Is there a sort of known relationship between 10x more of

• 1:22:48Nathan Labenz?

  the use of the models, 10x more compute translates to a certain higher hit rate? And how does that relate to the sort of loss landscape, if you will, of materials? I'm not sure if my intuition should be, if I come up with a material that has like, pretty good, but not quite properties, does that mean I'm close, where I can like, explore the very local space around that particular thing, and expect that there's going to be something better that I can kind of gradient descent my way into? Or I could also imagine that there could be just like, a ton of weird discontinuities in the material space, such that, I might take one step in a particular direction and

• 1:23:31Nathan Labenz?

  be like, in a totally different property regime.

• 1:23:34Speaker 02

  Yeah, I mean, there's always this explore-exploit conversation that's coming into play, regardless of what stage you're at. You could be at the purely atomistic level, where you're again talking about trillions and trillions of different atomic systems that you might want to look at, and you're going to want to explore-exploit there, right? You're going to want to explore the vastness of it, and then when you find something that feels pretty good, you're going to want to narrow down and go deep there. And as you kind of go down the ladder, and eventually get yourself into the lab, it's going to be the same question over and over again. And a big part of it is, at each layer, you need to do that slightly differently. Because the types of questions you're trying to answer in order to understand where you should be exploiting, are actually just different. What you're trying to answer at the atomistic level is very, very different, in terms of,

• 1:24:17Speaker 02

  well, what should I be synthesizing in a lab? Are they related? 100%. They are absolutely related, but they are a different set of questions, and you need to have different techniques in order to answer those robustly and soundly. And so that's just something that we're constantly doing over and over again. In terms of inference time, and all of that kind of stuff.

• 1:24:38Speaker 02

  I mean, generally, the way I like to think about it is if you actually just forget radically, ah, you forget machine learning, you forget automation, you just say, hey, as a scientist, would it make sense for me to think about a bunch of things kind of in my head, come up with a bunch of hypotheses, narrow those down in my own head, maybe on some scratch paper, maybe do a little bit of modeling, and then walk into the lab? Or should I just say, you know what, I'm going straight into the lab? Which one is likely to yield the best result? Well, it's probably the former, and I think that same thing holds and applies. Again, you do have to always keep in mind that science doesn't always work in a logical, rational way. There is just a natural degree of randomness and counter intuition that has to come into play, and you cannot forget that.

• 1:25:22Speaker 02

  We have to design every single system with this idea in mind. We have to abide by the general rules, but also understand that we need to be able to break them too. And so, everything is designed around that core thought process. But, um, yeah, I mean, to answer your question, if we can offload some of these things in simulation, that naturally makes a lot of sense, and we will do that. You just have to do it in different ways at different stages along the process.

• 1:25:45Nathan Labenz?

  Can you talk a little bit more about the active learning? I mean, we've got these sort of like few-shot prompting type things. Again, there's, you know, test time, training. There's a bunch of different paradigms where people are trying to push this, but it seems like, again, this is one of the core questions is like, can we really get them over the hump to where they can take on this new data that is genuinely out of distribution and like do an effective update on that?

• 1:26:13Nathan Labenz?

  What more can you tell us about like how you're getting that to work?

• 1:26:18Speaker 02

  Yeah, we do active learning, again, kind of in every, every state. So at the atomistic level, we do active learning, and that's different active learning than what we do down in the lab. But we do do it all the time to improve our recommendations, to improve the likelihood that as we kind of move into the next phase, it's better and better. So at the atomistic level, we might be doing some active learning around uncertainty. And so, we might be trying to generate these recommendations around forces. And if we're not feeling pretty good, we might run some DFT and then kind of update our understanding of how we're making our predictions at the atomistic level. In the lab, that's a totally different thing, but we're still doing active learning. We might run an experiment and realize, ah, you know what? Even though we scanned a bunch of literature, even though we've done all of these calculations,

• 1:27:01Speaker 02

  we didn't actually realize that the pressure and temperature is just a little bit different today. And so we actually need to update it for that next experimental run. And so we do do active learning all of the time, and that's mostly because science is not necessarily predictable to a perfect T. It actually does change in a bunch of different ways. And it kind of comes back to this core thing around two main aspects, which is number one, we just don't have the data. It's out there, but we don't really have it. And the more of it we can get and the more of it we can get in a good fashion, the more likely we are to move successfully through that process. But then number two, never forget that there always has to be some degree of randomness that we have to lean into in order to open ourselves up to the greater discoveries. And so we do do active learning kind of all along the way.

• 1:27:43Speaker 02

  We think it's just an important part of the process that, you know, naturally happens in science. And if we're trying to replicate and improve and upscale the way that we do science, we have to include that into it as well.

• 1:27:56Nathan Labenz?

  Yeah, that's the core stuff these days. So we've, we've made it pretty far in terms of learning the, the past distribution. So this, how to take a surprise on board and really do the right thing in response to that. That's, yeah, it's central in, in all areas. In terms of like interpretability, um, you know, a bit earlier, Joseph, you kind of told the story of the, you know, the grizzled vet who was like, oh, that's a, you know, that's a mono layer film you've got there. That's why that's happening. To sort of tell a little hackneyed story there, it seemed like you had, you know, gradient dissented your way into a heuristic. I guess, like, if I understand correctly, you had sort of realized without a theory

• 1:28:40Nathan Labenz?

  that, you know, when they look more transparent, that seems good.

• 1:28:44Speaker 02

  Right.

• 1:28:44Nathan Labenz?

  And it seems like a lot of models today, certainly thinking at the language model level at least, kind of operate in this sort of heuristic zone where they've like learned the patterns, but they haven't really grocked the underlying laws. And then, you know, you've got this veteran that comes in and says, oh, well, I've grocked that situation. And I can give you not just a, you know, I can actually tell you why your intuition has, has any value with a, you know. Something that's rooted in the actual physics of the question. I basically think that like a lot of that is going on in models as far as I can tell. And I always think back to the original Grokking paper from OpenAI where you know, you get heuristics really fast. You can memorize the, the training set really fast.

• 1:29:28Nathan Labenz?

  It takes like orders of magnitude longer in that original experiment to actually do the Grokking and really implement an algorithm that does, you know, the modular division. And then it took interpretability to even know what's going on. You know, is it still just guessing or does it have some actual like right way that we can trust to solve this problem? Can we be confident that it will actually work for like all unseen cases? So where are you in the, you know, I guess, where, where do you think the models are? And to what degree have you been able to like do the interpretability at this point to have a sense for where the models are on these like Grokking curves? Are we, how, how much of the time are you getting stuff out that's still like heuristic and not really reflecting any real understanding versus, you know, are there things that you

• 1:30:13Nathan Labenz?

  can point to and say, like demonstrably like we see real understanding here that, you know, is principled in any particular model in any particular area?

• 1:30:23Speaker 02

  So I wouldn't say that we are in a place where we can demonstrably point to a very clear and definitive sense of understanding. What I would say is that depending on the models that we're working with, you see varying levels of this. And it's almost like in the inverse where you see sort of signs of memorization. You see signs of obviously hallucination and that kind of stuff as well. And so I think the fact that you could kind of see the inverse means that you know, there should be a way to get to, to the opposite end of it as well. And this is why we are so bullish on interpretability, really playing a big role here is because we think that there is a, a realistic world where, where these models can actually develop

• 1:31:06Speaker 02

  a rich representation of a materials problem and that we can then therefore sort of like learn how to exploit that model and steer it in the right direction. But I wouldn't say that we have any proof or anything like that just yet.

• 1:31:20Speaker 01

  And the one thing I would add on from a scientist perspective is, and this is one of the reasons we got excited about working with Eric and the Goodfire team is, is this potential ability to try to unearth things like this is.

• 1:31:34Speaker 01

  Oh, I don't know if we know. Actually, I do know that we don't know what I know that we know though. Yes, of what we might want to look at to help impact our discovery. Like the monolayer film on an optical microscope and seeing the transparency of that. If you had asked me before, I would never have said to even look for that. And so like out of the machine learning world, I got a base level, I got a scientific level. We, I don't even think we fully know like which areas we can look into, what heuristics we can try to capture. That might be impactful and in some cases might not.

• 1:32:15Speaker 01

  Like there, there's probably a world where the transparency of something does not matter to the end resulting property that you're going after. So there's the for it's not effective, but maybe it's transferable. Right. And maybe there is something that you can actually learn from that process. That comes into a different heuristic that you use. And I think at a scientific level, as a human scientist. This is why good fire and interpretability in this concept to me is so exciting. This ability for AI to maybe teach you what those heuristics are for lack of, for a very simplistic phrase. And actually use those to drive different results that you would never be able to insinuate. Now, there are numbers, there are weights and biases in a way.

• 1:33:00Speaker 01

  That's really cool to me because you, you remove the human bias, the human perspective from that. And you, and it's just math and, and on the underlying models. That's a really cool concept. I'm not sure we've had too much of that ever in science per se, maybe much more in theory. Where you can kind of have this optionality and do different type of simulations, but certainly limited and experimental. And I think that's a really, really cool perspective to bring into it from purely the scientific perspective. Not deep in the machine learning weeds and what does it mean and has it impact models? From a scientist perspective, what is the output of those models? What does interpretability give me?

• 1:33:42Speaker 01

  I just think that's such a cool concept that is very infant in its exploration. And we believe, as Horace said, will be a really important field for science at large over the next two to five years.

• 1:33:56Nathan Labenz?

  Have there been any like move thirty sevens in the material science domain? And or, you know, the alternative being like, there's a lot of just like doing kind of grinding work a lot faster and, you know, at a scale that just couldn't happen otherwise. I guess, have there been, would you be confident if one, like, would you immediately know it if there was one? Um, and if there haven't been any yet, you know, is that a sort of two to five year timeline in your mind? As you just scale out data, eventually you think you'll get there? Like, what's the the outlook from material science move thirty seven?

• 1:34:31Speaker 01

  Great question. I think we're moving towards that. I'm not sure if there was a earth shattering one yet. Me like, like in our taps you've conducted, like, just to use an example we talked about earlier, where everyone in the world is going to realize and then feel the direct result of that discovery that comes out of a lab. I can't see that at thirty move thirty seven there yet. There probably are examples of materials that are being optimized, that are being discovered in the computational realm, in the AI realm. I think it's very lacking still on the experimental side, hence why we think our flywheel is imperative in doing materials discovery. But I think one cool thing, we always tell the AlphaGo story. We think it's a really good story about indexing information. How do you bring that to science?

• 1:35:15Speaker 01

  No, it had learned Go games, right? I forget the exact millions of Go games, whatever the exact number was, and made this one in 10,000 move. And then the subsequent versions of that AI didn't need to learn Go games, right? It actually made similar moves without having done millions of of learnings in Go. I'm excited for us to get to that place as well. And I think we are building the technology to do the first version of AlphaGo today. That infamous move of thirty seven, everyone watching, AI made a mistake. It was a one in 10,000 move. We didn't see it coming. I think we'll have something like that. I think Radical AI will have something like that in the next 12 to 24 months. If we follow and hit our objectives in our roadmap. I think once those become.

• 1:36:01Speaker 01

  Hard for the course, become more normal. Then moving into the, we didn't look at experimental data, or we used experimental data from alloys and start making discovery in room temperature superconductors. That's, that's what we think is gonna happen. We're gonna amass the largest experimental data set ever built in human history. That's what we think. And so if we do that, we are very, very excited about the learnings that come from that. And then the other areas that we can make move thirty sevens in that don't need millions of trained examples exactly in that specific material space. To us, that is the holy grail of material science. That is removing materials as the biggest bottleneck to our most important industries.

• 1:36:46Speaker 01

  That is enabling a world that is not limited by material, but only limited by imagination and the laws of physics. That's what we want to build and we are so excited to watch that proliferate and candidly very excited to be leading the charge on the intersection of AI in discovery and fully robotic self-driving labs to supercharge and inform that discovery. We think that flywheel together will really craft a world that, as we always say to the team, you can't imagine today, the amount of Ubers and Airbnbs and Facebooks and even things like Nvidia's that come from this world that we can enable.

• 1:37:30Speaker 01

  We think that will be one of the most important companies to ever exist because of that. And that's the world that we're excited to live in.

• 1:37:37Nathan Labenz?

  In many podcasts, I would say that's a great place to end it, but in this one, I have a couple more questions because that is an inspiring vision and we can link to the careers page in the, in the show notes if you'd like. How about the future of the integration of modalities? This is kind of a hobby horse question of mine, but you know, I do, to be very real about it, lower impact of work, but you know, I started this company, Waymark, we helped small businesses create video content for local advertising. We had like basically a pre-AI version of this and then we've tried to infuse AI into everything that we do. And one of the biggest things that I've observed is that the integration of language and visual understanding has been

• 1:38:22Nathan Labenz?

  A real step change when you no longer had to go through a language bottleneck to say, "Okay, like I have this image, I want to change it to this other image," or "I just want an image." You know, but if I have to say all that in 100-word prompt, even if I train a language model, which I've done to, you know, prompt effectively, it's just such a lossy thing, right? And what has made some of the more recent models just go viral and just amaze people so much more is that they clearly have this joint understanding where now I can turn my image into, you know, anime, whatever, but it still retains the meanness of it while layering on this other kind of coat of paint, whatever, right?

• 1:39:04Nathan Labenz?

  But that depth of understanding where it knows what you're asking for in language, but it also really clearly gets the image itself, you know, at a pixel level, that has seemingly created a major step change. What I understand today about the system that you guys have built is like, there's a lot of models that are doing a lot of kind of subparts. Do you think that this kind of goes the Tesla route eventually? Or the Gemini 2.5 flash image out, whatever, route where it's like, there's one joint embedding space and kind of language and a sort of intuitive physics to potentially like even Grot physics of materials all exist together?

• 1:39:45Nathan Labenz?

  Or is that like too fanciful to expect at this point?

• 1:39:50Speaker 02

  I don't think it's fanciful at all. It's probably going too far to say that we'll only ever need that one model to do everything. But I mean, internally, we are developing something like what you described. And so it will be natively multimodal nature. We'll have encoders for all of the different modalities that we want to work with. And we will run it through this model and we'll be able to have things output from it that are quite useful to us. So, no, I don't think it's fanciful at all to say that we'll have something like that. In fact, I think, you know, in the next six months, we'll be using that internally here at Radical AI. And certainly within the next six to 12 months, we will likely publish and release something along those lines.

• 1:40:32Nathan Labenz?

  The interpretability of that is going to start to get super interesting. That is one of Eric's prized things. We have been talking about it for a while. If we could work on something, something like that. Yeah, cool. And that sounds like it's not super far off. How do you think about the IP associated with this? I mean, you're going to accumulate the world's biggest experimental dataset, as you said, and then I don't know if you meant like open source necessarily, the model that you're creating there when you talk about publish it, but and I also just don't know a lot about like the IP that exists in the materials science space. So how do you guys think about what data and what value you know, you want to kind of keep internal to the company versus what you can share?

• 1:41:15Nathan Labenz?

  Again, I don't even know if you can you, is it possible to like patent, you know, an alloy, especially something amorphous that would you be patenting the process? That creates that? I mean, you can't really patent something that's amorphous, right? And has no like repeating structure. I would think that would be hard. Maybe the process to create it. So yeah, I guess what is the sort of outlook for what is going to be core IP not shared and what can be shared? How do we get the how do we diffuse the benefits of this while still giving you guys the ROI needed to like scale out the the autonomous systems to actually scale up the the flywheel, the dataset and eventually get those move 37s?

• 1:41:54Speaker 01

  Yeah, I mean, it's an awesome question. And when we started the company, we spent a lot of time thinking, okay, what, what do we want to build? And not just what do we want to build? What is needed if we are trying to solve the material science problem? Not what's the easy low hanging fruit? Not what's a business model that we know VCs will get behind? What is needed to build a company that does the vision you've heard us articulate today? And what we netted at was, well, if you just sell software. You're never going to make real materials. And so that can't be it. So we're not going to do that. And there's been a bunch of companies who have tried and failed doing that. If you try to license every material you make, that can be a business model.

• 1:42:39Speaker 01

  But the problem is then you need to be creating new discoveries multiple a year for decades to keep up with and continue to drive revenue. Because of course, reverse engineering and expiration of patents becomes a serious problem when you're just licensing. And so then we looked at what the largest material companies do in the world. What is 3M, Dow, Applied Materials, BASF, Materion, et cetera? What do they do? They sell materials at scale. That's what their business models are. Except they're doing it in areas that frankly, to be completely blunt, don't impact society. Look, I understand the need to have a new Sharon Williams pink color every year. My wife is a massive fan. I know, I'm a customer.

• 1:43:25Speaker 01

  But that's not why we're building Radically AI. Honestly, they can continue to handle that. We are trying to make a species multi-planetary. We're trying to drive lossless nuclear energy. We're trying to rethink transportation with floating bullet trains. Like, we're trying to create a world that we want to live in. And so if all of that is true, then not only do you need to sell materials at scale, but you need to push the materials that you're selling at scale into enabling the industries and creating the industries of the future. And so when we do that, our focus is on the IP of making it. And in material science, yes, you get patents on composition.

• 1:44:06Speaker 01

  But it's really, the IP is really around the trade secret on how do you make it at scale? And for the lowest cost at scale. It's all, we call it processing in material science, all in that process. Where all of the real meat of it is. You know, if I tell you a high entropy alloy that we have today, like the exact composition. You're not really gonna be able to make it at scale per se, right? And so this is a really important piece of this. And so if that is the business model, then that changes the way we think about the AI models. And I'll let Jorge articulate that specifically.

• 1:44:39Speaker 02

  No, I think Joseph hit it on the head. I mean, the machine learning models will change. They will. They're gonna get better and better over time. We're gonna create some of them. Other people will create some of them as well. We may release some of them. Some of those we probably won't release, to be entirely honest with you. But the one thing that Radical AI will never not do is go after the hardest problems that we think are gonna be needle movers for the world. At the end of the day, that's the whole point of the company. And so that will stay with us through and through.

• 1:45:06Nathan Labenz?

  Love it. You wanna talk a little bit about the, I know we're just about at time here, and I really appreciate all the time you've shared with us. Just yesterday, you put out a news item about working with the US Air Force. You wanna tell a little bit about that?

• 1:45:19Speaker 01

  Yeah, Superx very, very excited about this. It is in high entropy alloys for hypersonic applications. We are conducting a lot of really hot, not just high-throughput experimentation, but pushing into how to test and optimize these materials for hypersonic applications. Materials, again, are one of the biggest bottlenecks in the most important industries. And in the hypersonic industry, China, Russia, both field hypersonic systems today, they do that because they never stop investing in material science. Actually, it's worth mentioning, China, when they make a new discovery in a material, will set up an entire manufacturing hub around said material just to figure out how to scale it.

• 1:46:01Speaker 01

  It is an unbelievable emphasis on putting materials at the bedrock of all the most important innovations. And we have struggled with that in the US and other places. And we are trying to drive that nexus back through. And so we spent a lot of time in DC, a lot of time with the Office of Science Technology Policy, the National Science Foundation, Department of Defense, Department of Energy, even going to the Hill and telling people in government, look, public-private partnership and a focus on materials is imperative. Whether you care about hypersonics, you care about nuclear fusion, or you care about the data sets that are gonna go inside the ML models to generate a bunch of other new materials. That needs to be driven from a government perspective. And you need a lot rely on private enterprise to do exactly that, which is where we fit in.

• 1:46:46Speaker 01

  The direct to phase two. Is in our opinion, kind of our belief in trying to push this forward. And separately to that whole plan I just mentioned, really an ability to execute on the first material system that we've chosen. You know, we had a thesis that high entropy alloys can be very impactful for hypersonic technology. They're very hard to discover. It's a very large search space. There aren't a lot of experimental results. And our work with the Air Force is directly around trying to solve that problem and really generate these new materials for them to then go look at exploring putting into hypersonic systems. So that is what we've been working on and you can find a lot more information in our press release and more information on that. But we're really excited about that first project, which is again, specific to the

• 1:47:29Speaker 01

  Air Force and hypersonics, but not just stopping there. We have a lot of other things in the works with the US government that we are trying to push forward so that we can identify, look, materials are not a subsector of industry or innovation. Materials are at the cornerstone of innovation, whether it's the stone age, to the bronze age, to the silicon age. Every modern technological wave has come from novel material advancement and the future will be exactly the same. And that's kind of the perspective that we're trying to push inside government and inside policy and inside why materials are really important as a nation and really as a society, as the human race.

• 1:48:09Nathan Labenz?

  That's great. That could be the right place to end on. Anything we didn't touch on that you would want to make sure people know about before we go?

• 1:48:18Speaker 01

  One thing we call out, so we are really intense about our culture at Radical AI because we're trying to solve some of the hardest problems in the world. And so you have to have people that have really ridiculous curiosity, this question, this ability to go to first principle and ask why, but are also unbelievably relentless. We fail every single day and we will continue to fail purposefully every single day for the next decade and beyond that. And that's actually where the unlocks come from. It's kind of building on those failures, of course, to reach success, as crochet it is. So if there are people out there listening who really have an, you know,

• 1:49:03Speaker 01

  an appetite to challenge everything down to the laws of physics and want to do that in a way that is trying to tackle our most important problems, recognizing that there's going to be a lot of failures along the way. But in the chance that you succeed, you leave a fundamental stamp on the human race of impact and innovation, then please reach out. We are actively recruiting right now and we are looking for not only the best people in what they do, but people that are willing to dedicate their life's work to a mission. If you're looking for a job, this is not the place for you. You should not apply to Radical AI. You will not fit in here. If you are looking to work on a mission, then please reach out. We'd love to talk to you. And Jorge, anything you want to add on top?

• 1:49:48Speaker 02

  No, I think that was perfectly well said. Nothing to add.

• 1:49:52Nathan Labenz?

  Amazing. Well, this has been an excellent conversation. I've learned a lot from it and I think people are going to find it fascinating. So obviously I wish you guys nothing but success in terms of breakthrough after breakthrough and look forward to changing material reality around us as we unlock one thing after another. Joseph Kraus and Jorge Colindres, co-founders of Radical AI. Thank you for being part of the cognitive revolution.

• 1:50:15Speaker 02

  Thank you.

• 1:50:17Nathan Labenz?

  If you're finding value in the show, we'd appreciate it if you'd take a moment to share it with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts where experts talk technology, business, economics, geopolitics, culture, and more, which is now a part of A16Z. We're produced by AI Podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at ai podcast dot I N G.

• 1:51:02Nathan Labenz?

  And finally, I encourage you to take a moment to check out our new and improved show notes, which were created automatically by Notion's AI Meeting Notes. AI Meeting Notes captures every detail and breaks down complex concepts, so no idea gets lost. And because AI Meeting Notes lives right in Notion, everything you capture, whether that's meetings, podcasts, interviews, or conversations, lives exactly where you plan, build, and get things done. No switching, no slowdown. Check out Notion's AI Meeting Notes if you want perfect notes that write themselves. And head to the link in our show notes to try Notion's AI Meeting Notes free for 30 days.