SPEAKER_05
None of you are here to see me but I do have to tell you where you are. Welcome to South Park Commons New York. Put your hand up in the audience if you are part of South Park Commons. After the event if you want to learn more you should talk to them. Quick note on SPC I run our New York office with an amazing team Suzanne and Arian. We've been around in the Bay for 10 years but in New York for a couple now. We're a space for deep technical exploration so we try to bring together the smartest people who want to solve the hardest problems. We want to bring them into our space. We don't take any equity. We don't charge. We just want really smart people to come together and hopefully we get to invest in a bunch of you and accelerate building businesses that might one day look like these guys. So if you want to learn more about SPC if you're curious about what starting something or doing full-time technical exploration could look like I'm always around. And with that said thank you so much all of you for coming and I'll hand it over to Eric to kick it off.
SPEAKER_05
Hey I'm Eric Newcomer. I'm the author of the newsletter Newcomer and host the Cerebral Valley AI Summit. Thrilled man this is quite the crowd and people are eager so I will not talk too long. At some point I'll open it up. We definitely want people to engage. I think these are three companies truly at the cutting edge so excited to get right into it. Do you just want to introduce yourselves and give the brief what your company does?
SPEAKER_05
Yeah sure. Hey everyone good to be here. I'm Joseph Krauss. I'm one of the co-founders and the CEO of Radical AI and what we do is build AGI for science. So we are discovering new materials with an AI scientist and a fully robotic lab that can actually impact some of the most important industries in the world. So that's us. PIM, general intuition. We build general agents for environments that require deep spatial temporal reasoning. So think things like robots or moving around in simulation.
SPEAKER_05
I'm Rob Cochran. I'm a co-founder and CEO of Fauna Robotics. We're building safe reliable and fun robots for everyone. Our first product Sprout is a three and a half foot tall humanoid robot that we've made available as a development platform. [SPEAKER_08] All right. We'll get right into it. Rob, Amazon has bought your company. Why did you sell? Tell us the story. Let's go.
SPEAKER_05
[SPEAKER_08] Yeah so this news came out last week obviously much longer in the works. I worked at Amazon previously several years ago in AWS and had a bunch of relationships there since starting the company. In many conversations we discussed partnership and different things we could do together. I think ultimately they really liked our vision for how robots would one day work and live alongside people and wanted to support that in a big way. And the chance to get to keep doing this and doing it in New York and with a lot of support from a great partner felt very exciting.
SPEAKER_05
[SPEAKER_08] And so you're going to lean in on the form factor you're doing which is a cute research robot. Give us the articulation of why we're not going straight to humanoid robotics.
SPEAKER_05
[SPEAKER_08] So though we have started with a two arm two leg robot named Sprout. Cuteness is part of the approach. Cuteness approachability something engaging something people actually want to be around. I think the reality of robotics today is that they don't work perfectly. And so one it has to be safe. So if you have a hundred and fifty or two hundred pound machine that falls over you or falls into a child that's it. So you obviously can't deploy that kind of risk into real environments. And I think there are still many innovations required to make robots work very effectively in homes. And so having a platform that is experimentally deployable in a lot of situations is very useful. And then making it something that people want to be around so they're a little forgiving of some of its faults.
SPEAKER_05
[SPEAKER_09] All right. Pim you come out of the video game world right. Is it you were running the largest runescape server at 14. That's a correct fact. I think I do all my prep with Claude now which is dangerous but it's so good. [SPEAKER_07] That's accurate. All right. And then you had a company that had access to tons of video game clips. You go out you start pitching them to foundation models and you figure out actually this data is pretty valuable. Maybe I should build my own company off of it. Is that a fair origin story or what is the sort of yeah. What was the idea behind general intuition.
SPEAKER_05
[SPEAKER_03] Yeah. So I was too young when the mobile revolution happened and I think I come from consumer perspective. So I looked a lot at what platform shifts do. Usually the biggest companies are built right after these things called platform shifts happen. Which is when there is a large unsettling of the status quo and how things are built or how networks come together. And in video games Discord was a large platform shift and Metal came on top of Discord largely. [SPEAKER_08] Metal is one of the largest video game recording applications.
SPEAKER_05
[SPEAKER_08] And so I was too young for mobile and I saw when the LLMs happened. I was already fairly obsessed with reading the papers and trying to understand how this works. And then I decided to go all in. And then indeed we just started looking at how the data could be useful. Like originally we had some ideas of taking gameplay clips and using them to create evaluations of the best players. And can you evaluate language models on how they would do certain things.
SPEAKER_05
[SPEAKER_03] And then we actually realized that the data set was so general that games are actually a very good representation of digital information. I.e. text which robotics lacks in the physical world. And then in the physical world you have spatial and temporal dynamics where things unfold over longer periods of time. Which text and pictures lack. And so we sort of realized that we could probably we had so much of this data that we could pre-train entire foundation models on it. [SPEAKER_08] And we've been able to do that very well. [SPEAKER_08] You're building agents for real world. [SPEAKER_03] Yeah.
SPEAKER_05
[SPEAKER_03] But the world models are built from video games. Is that right? Like what is your view on whether the successful world models. Because that's such a buzzy term right now. Will be based on mapping this room or based on video games just because there's so much volume. Obviously we saw similar debates with self-driving cars. [SPEAKER_03] Yeah. [SPEAKER_03] The phrase is so broad now it's very difficult to give a succinct answer. But I'll try.
SPEAKER_05
[SPEAKER_08] So what sort of world models is actions and frames. Right. So you have an action and then you have an observation. I need a next frame and then the next frame can be generated off the action. That's kind of the original term I guess. Or the next state. I'm using the word frame in this case because it's more familiar. But the next state of the environment is generated off the action that you take. That sequencing of data just doesn't exist in the real world in a.
SPEAKER_08
[SPEAKER_03] Yeah. [SPEAKER_03] The phrase is so broad now it's very difficult to give a succinct answer. [SPEAKER_03] But I'll try.
SPEAKER_08
So what sort of world models is actions and frames. Right. So you have an action and then you have an observation. I need a next frame and then the next frame can be generated off the action. That's the original term I guess. Or the next state. I'm using the word frame in this case because it's more familiar.
SPEAKER_09
[SPEAKER_08] But the next state of the environment is generated off the action that you take. [SPEAKER_08] That sequencing of data just doesn't exist in the real world in a timely accurate way. [SPEAKER_08] And what's important is that if you want your models to properly react to the actions you take. [SPEAKER_08] Then you want to be able to separate out the action from the state. [SPEAKER_07] Right. [SPEAKER_07] So the action from the world. [SPEAKER_06] As humans we don't have that. [SPEAKER_07] And if you want to label it off video.
SPEAKER_07
Even the best humans. Right. They will not get that precise within a hundred millisecond range. If you want your models on very little data to learn these causal relationships.
SPEAKER_03
[SPEAKER_07] Then you actually want that to be ground truth. [SPEAKER_07] Therefore actually getting it from video games and software. [SPEAKER_07] There's a faster way. [SPEAKER_07] You actually control how the frames and the actions are being put together.
SPEAKER_08
[SPEAKER_07] And therefore we just found that you could pre-train on that.
[SPEAKER_07] And then do a lot more afterwards. [SPEAKER_07] So we found this to be in our opinion a great way of doing it. [SPEAKER_07] And there's other companies that will try other ways. [SPEAKER_07] We'll see how it plays out. [SPEAKER_07] All right. [SPEAKER_07] Joseph. [SPEAKER_07] Radical AI. [SPEAKER_08] You're in some ways what people think with the frontier. [SPEAKER_08] Find new materials. [SPEAKER_08] And you have what Raytheon is a big backer.
SPEAKER_08
So this is outside. There were serious companies thinking about this.
SPEAKER_03
[SPEAKER_08] Yeah. [SPEAKER_08] What material. [SPEAKER_08] What's your approach. [SPEAKER_07] I guess defining materials. [SPEAKER_07] And then what is materials even the word you use. [SPEAKER_07] And what is then the first material you're hoping to stumble upon. [SPEAKER_07] Yeah. [SPEAKER_07] I think it's worth maybe taking one step back. [SPEAKER_07] Right. [SPEAKER_07] And I think the thing I always talk about.
SPEAKER_08
[SPEAKER_07] And you might have heard me say this before. [SPEAKER_07] Is what do materials mean? [SPEAKER_07] Why are they important? [SPEAKER_07] Why are we working on this technology? [SPEAKER_07] And the reason is. [SPEAKER_07] If you really paint the picture. [SPEAKER_07] Right. [SPEAKER_07] You think about how many industries are impacted by material science. [SPEAKER_07] Automotive, aerospace, manufacturing and defense, climate, energy, semiconductors, electronics. [SPEAKER_07] Most important industries in the world. [SPEAKER_07] All are direct result from materials R&D. [SPEAKER_07] But when we were paying attention and we can get into why we started the company.
SPEAKER_07
It just takes way too long. Way too much money and way too much fragmentation to solve them.
SPEAKER_06
[SPEAKER_07] And so the big thesis behind the company is this connection between.
SPEAKER_07
An AI scientist who has a ridiculous level of intelligence. I mean. [SPEAKER_08] Better than a PhD scientist would have today. And we're a couple months into training. On real experimental data. And then a fully robotic. Self-driving lab. And that means that we can actually run experiments. [SPEAKER_08] Synthesize materials. [SPEAKER_08] Characterize materials. [SPEAKER_08] And then test materials in high throughput. [SPEAKER_08] And we capture all that information. [SPEAKER_08] None of that information is captured today. [SPEAKER_08] And so you build this data set that does not exist in the experimental lens. [SPEAKER_08] Has really important information.
SPEAKER_07
[SPEAKER_08] It's the actual real world impact of materials and the way they perform. [SPEAKER_08] And then can actually be used to push those materials to end applications.
SPEAKER_08
So Raytheon's an investor. NVIDIA is an investor. Any is an investor on the energy side. We do a lot of work in nuclear fusion. And so you have all these companies. Materials is huge all over the world. [SPEAKER_09] Where do you start? [SPEAKER_09] It's anything could be useful. [SPEAKER_09] That's right.
SPEAKER_07
[SPEAKER_09] And so where we thought about was where is an area that really has so much opportunity to show why the technology is relevant. [SPEAKER_09] And in the space that we work which is called structural metals. [SPEAKER_09] There are 10 to the 40 different potential combinations. [SPEAKER_09] So I want you to envision this. [SPEAKER_09] That's a hundred trillion Earths of different potential combinations that you can synthesize. [SPEAKER_09] That'll take 7 million years if any of you try to do that. [SPEAKER_09] That's where AI is really good at weeding. [SPEAKER_09] And self-driving labs are really good at making. [SPEAKER_09] And so that's why we chose that area.
SPEAKER_07
[SPEAKER_09] Aerospace defense and energy are the areas that we're working in today. [SPEAKER_09] And then moving into semiconductors as a future area which we can talk about. [SPEAKER_09] Is there a very specific sort of? [SPEAKER_09] There is. [SPEAKER_09] It's called a high entropy alloy. [SPEAKER_09] So it's a material that has 5 to 7 different elements in it. [SPEAKER_09] And they're all about equally atomic. [SPEAKER_09] And it has really high entropy in the system. [SPEAKER_09] And so what it can do is have incredible properties at crazy conditions. [SPEAKER_09] You're flying through space at Mach 10.
SPEAKER_07
[SPEAKER_09] And then it gets ridiculously hot, ridiculous pressure, and it's really oxidative. [SPEAKER_09] It's really corrosive. [SPEAKER_09] These materials withstand all of that. [SPEAKER_09] So they don't feel the effects of those conditions. [SPEAKER_09] They're pretty exotic in that way. [SPEAKER_09] How much?
SPEAKER_08
[SPEAKER_09] This is for everybody.
SPEAKER_07
[SPEAKER_09] I want to understand how much was ChatGPT and what's happened with the language models? [SPEAKER_09] An inflection point for what you're trying to do. [SPEAKER_09] I know some people came after that and everything. [SPEAKER_09] But was that a core technology to what you're doing? [SPEAKER_09] Or do you see the actual technological leap as being something totally different? [SPEAKER_09] Yeah, I mean for robotics it's definitely a huge innovation and step forward for planning.
SPEAKER_08
[SPEAKER_09] So if you think about it, hey robot grab me a snack from the fridge. [SPEAKER_09] So they're pretty exotic in that way. [SPEAKER_09] How much? [SPEAKER_09] This is for everybody. [SPEAKER_09] I want to understand how much was ChatGPT and what's happened with the language models? [SPEAKER_09] An inflection point for what you're trying to do. [SPEAKER_09] I know some people came after that and everything. [SPEAKER_09] But was that a core technology to what you're doing? [SPEAKER_09] Or do you see the actual technological leap as being something totally different?
SPEAKER_08
[SPEAKER_09] Yeah, I mean for robotics it's definitely a huge innovation and step forward for planning. [SPEAKER_09] So if you think about hey robot grab me a snack from the fridge. [SPEAKER_09] The ability to plan that task out flexibly and accommodate, oh a guest at the fridge. [SPEAKER_09] And the drink that your favorite drink's not there so can it pick another one. [SPEAKER_09] That planning and replanning was just not possible through formally possible structured planning routines. [SPEAKER_09] But for us, I mean there have been many innovations.
SPEAKER_09
The availability of embedded GPUs that can run on device in real time. [SPEAKER_08] Actuators that you can buy off the shelf to animate the robot's body. [SPEAKER_08] And then for us, and my partner spent time at DeepMind doing work in simulation. [SPEAKER_08] And the availability of physically realistic simulated environments where you could train robot behaviors and have those behaviors actually reflected on a real robot was not possible and not a way that people developed robot behaviors, five or ten years ago. And is now quite possible and with accelerated compute can be done at scale.
SPEAKER_09
And that's I think driven a real set of innovations more from a robotic control perspective. And is AI but it's not language models. Yeah, obviously you're the next generation of language models. How would you draw the connection between what you're doing in the work that's happening in language models today? I think that the last point that he made was quite interesting on the simulation piece. I want to explain that a little bit better because I think it will make everybody understand why it's important. So previously you would have to hire an engineer who goes and codes up in a simulated environment. And then you do it in an engine.
SPEAKER_09
That engine is probably roughly modeling Newtonian physics or something. And you don't know precisely how much because it made some tradeoffs in some places to go faster. So then you need to check how far off they are. [SPEAKER_08] And then now what you can do is you can create verifiers where you look at the outputs of the engine. And you can have an LLM put together a bunch of environments, for example, and pick ones that work well. And so I think the role of the engineer in a lot of parts of the stack has moved from the person actually controlling the number of output to verifying the output and using that as a way of.
SPEAKER_09
It's a massive force multiplier that the size of the teams that you need to do what you needed to do before is massively less. I think LLMs are in many ways just a way of—a simplest way that I describe it is all the time that collectively as humans we spend searching the Internet, reading things on Stack Overflow and typing it. All that's compressed right into a very easy sequence. And so now, I think it's many exponentials of speed ups and it's funny. The work that we all do, which has in common, we saw what happened in text.
SPEAKER_09
And you see, also looking at the scaling laws and all the techniques being developed for scaling or pre-training, all these things are extremely important. They're all quite general. A lot of the models, a lot of model architecture is a lot of the pieces you can directly transfer. And so, all super fundamental work. I actually think people don't realize how similar it all is under the hood. Yeah, yeah, I agree with that last point entirely. [SPEAKER_08] I think there's so much similarity that comes over and even things like you learn about training. [SPEAKER_08] You know, ChatGPT learned the English language or no, of course not.
SPEAKER_09
[SPEAKER_08] It learned how to represent that and then use that when it produces text. [SPEAKER_08] And I think we think about similar things in material science that we learn from the language field. [SPEAKER_08] We also use it. And I think that's important. [SPEAKER_08] You know, we have this head honcho who runs our science. [SPEAKER_08] He's the guy in charge. He's the new PI of the lab. [SPEAKER_03] And that's an LLM—it's from OpenAI and Claude. [SPEAKER_03] We fine tune that to understand materials. But he doesn't. [SPEAKER_03] That agent. Do you give him a name internally? [SPEAKER_03] He's Rad. So that agent calls the tools that he needs.
SPEAKER_09
[SPEAKER_03] And it's really good at understanding, OK, I want to go into simulation. [SPEAKER_03] I'm going to run quantum chemistry. I want to run labs. I'm going to talk to my APIs in my lab.
SPEAKER_08
[SPEAKER_03] The type of PI where it's, I don't need to explain this to you. Just trust me. [SPEAKER_03] I'm an expert. It depends. [SPEAKER_03] It's how much did your humans come around and figure out why it says go this way. [SPEAKER_03] So we do that a lot today and it's really important to label the data—a common industry phrase meaning how does it know if an experiment is good or bad?
SPEAKER_09
[SPEAKER_03] Well, a human scientist tells the agent. So we'll come in. The agent will analyze the result from the lab. [SPEAKER_03] Our human scientist will analyze the result from the lab. Then we'll compare them and you build this training data set of OK, I understand what the human scientist is going for. [SPEAKER_03] So the humans are definitely in the loop today and it's really important. [SPEAKER_03] But there's a bunch of other technologies that we custom build. Diffusion models are a perfect example. [SPEAKER_03] There's not a real good diffusion model in materials today. We have to build that from scratch.
SPEAKER_09
[SPEAKER_08] We would never build an LLM from scratch. So I think they're important in the learnings and they're important in how we can use them in our workflows today. [SPEAKER_08] In five years, what's the world look like? That's the fundamental question. You know, robotics—I mean, there's the sort of will humanoids happen that quickly or not? [SPEAKER_07] Both on the sort of bullish and where you're bearish, how is the world more different than we expect in five years?
SPEAKER_09
[SPEAKER_07] Where do you think? Oh, I don't know. I don't know if we're going to make a leap in your domains or any area that would be interesting to people building here?
SPEAKER_09
[SPEAKER_07] Yeah, I mean, from a robotics perspective, I think there's definitely still a lot of work to be done to make robots very effective across the broader set of tasks that we'd like to eventually have them do in the home. [SPEAKER_07] That's part of why we started with the development platform. So AI labs in academia and corporate research labs are able to build and extend and develop some of these capabilities. [SPEAKER_07] What's your view on where you're bullish and where you're bearish, and how is the world going to be more different than we expect in five years?
SPEAKER_09
[SPEAKER_07] I don't know if we're going to make a leap in your domains or any area that would be interesting to people building here.
SPEAKER_08
[SPEAKER_07] From a robotics perspective, I think there's definitely still a lot of work to be done to make robots very effective across the broader set of tasks that we'd like to eventually have them do in the home. That's part of why we started with the development platform. So AI labs in academia and corporate research labs are able to build and extend and develop some of these capabilities.
SPEAKER_09
[SPEAKER_07] I think some of what we're hearing here is that there's a lot that still needs to be done to ground robotic simulation in physical reality and to be able to make forward projections that are realistic and understand physics of the world.
SPEAKER_09
[SPEAKER_07] You don't believe that AI is going to hit some sort of self-improvement point and the rate of development is going to change so much over the next five years? I think we are on a very steep path where we're figuring out a lot very quickly. And we should see robots very broadly in the world. We're building and shipping robots today. I think we are on a very steep path where innovation still needs to happen with people. [SPEAKER_07] What's one specific thing that maybe in five years you'd hope to see? [SPEAKER_07] Or Pim, you can be our futurist.
SPEAKER_09
[SPEAKER_07] I'm going to steal this. I'm stealing this from Andre Karpathy. He said it in a talk privately once and I really liked what he said. There's broadly three stages: bits to bits, so anything digital to anything digital where there's lots of information available. Bits to atoms and atoms to bits as an interface layer, and then atoms to atoms. Those are the three stages of AGI. And they all have very different problems. [SPEAKER_07] Bits to bits is in hyperdrive. We're flying. There's no limits in sight.
SPEAKER_09
[SPEAKER_07] Bits to atoms and atoms to bits is largely constrained by measurement systems that we still need to build. And I think it's a very interesting space.
SPEAKER_09
[SPEAKER_07] A lot of the work you're doing, I'm sure, could be seen as starting as bits to atoms to bits and ending up in atoms to atoms. And then eventually atoms to atoms where you have completely ungenerated physical systems modifying all their physical systems. And really, I just look at it as a number of combinations of things that are possible in the world and the pace at which we produce them just goes up so fast that you just start seeing a lot more things a lot quicker. I think the same thing you're seeing with web apps being developed on the Internet. All your friends can develop apps now. You start seeing that with lots of different things. I think that's how it plays out.
SPEAKER_09
[SPEAKER_07] Who are your first customers, or what do your customers look like? [SPEAKER_07] So at the moment our customers are any robot that can be controlled using a game controller. What we do is we build world models and also policies or agents that are trained inside and evaluated inside the world models. And what world models are great at is simulating stochastic environments that would be very hard to simulate in traditional physics engines or where you'd have to do a lot of manual coding. And so you evaluate a robot not by how it does when you drive down a straight road but by how it does when something unexpected happens.
SPEAKER_08
[SPEAKER_09] And so that's why you need lots of stochasticity and unexpected things to happen. That's why world models are great. [SPEAKER_09] So our customers are typically robotics customers that you can control using game controllers because a lot of the data obviously is gaming actions. That's the general solve which is really cool. So you don't need to solve for precise motor torque predictions. You can just solve at the input of the game controller which every robotics company already ships with. [SPEAKER_09] And then on the humanoid comments, I think I'm less optimistic on humanoids, not home robots and actual.
SPEAKER_08
[SPEAKER_09] So I view AI as largely also an interface problem, which is funny as a person who builds models. I think it's largely an interface problem. [SPEAKER_09] We didn't start paying attention until ChatGPT, which was largely just an interface. And so the same thing with home robots. It's interesting, but largely the reason why humanoid was picked as an embodiment is because the assumption was that's the only place where we can go. We had large-scale data because humans recording themselves from the egocentric point of view. And also at the end of the day, you don't want heavy, bulky, thousands of dollar things that require energy if you don't need them.
SPEAKER_08
In many cases, if you can control the devices themselves, it's probably a better solve. And so I think in the home they'll be big because there are lots of very random tasks. I think in factories they'll be big, but I don't think it's the everywhere thing that we're imagining. I think they're just hard to manufacture. I think you'll start seeing specialized systems that people buy much sooner than you see very general remote robots as the general thing everywhere.
SPEAKER_03
[SPEAKER_09] Going back to this bits to bits, bits to atoms framework, which piece is the hardest for you right now and where are you in that process? [SPEAKER_09] Is this still bits to bits idea generation? [SPEAKER_09] Yeah.
SPEAKER_03
[SPEAKER_09] I like the framing we call it in our industry concurrent engineering. And so a lot of the systems that we build today, whether that's a fighter jet all the way through to a humanoid, those technologies do not design, particularly the materials, but other systems in concert with them. So in aerospace, when we design a new turbine or we design a new plane, we go back 30, 40, 50, 60, sometimes a hundred years to pull the material to use. And that's true in the automotive industry today. A lot of the cars we're driving use 1907, 1910 Ford Model T alloys. It's crazy to think about how we're at AGI and we're using 1910 alloys. So this is the atoms to bits, bits to atoms thing. We view this world where if you're designing a new plane or a new car, you should be designing the material at the same time that directly meets the requirements and constraints that your product is going to have.
SPEAKER_03
[SPEAKER_08] We're close to that. We can't do that at scale yet, but our scientists can talk to our agent. I want a material that does this. That agent will go run a bunch of calculations, figure out what an example that is, send it to the lab, and then make it. So you're really going from voice to end material and it's not perfect yet. We're still building that technology, but if you get that at a loop where anyone can do that on our platform, now every automotive, every aerospace, every mechanical engineer in the world has materials customization available to them.
SPEAKER_03
[SPEAKER_08] We're close to that we can't do that at scale yet but our scientists can talk to our agent I want a material that does this. [SPEAKER_08] That agent will go run a bunch of calculations figure out what an example that is send it to the lab and then make it. [SPEAKER_08] So you're really going from voice to end material and it's not perfect yet. [SPEAKER_08] We're still building that technology but if you get that at a loop where anyone can do that on our platform like now every automotive every aerospace every mechanical engineer in the world has materials customization available to them. [SPEAKER_08] And how do you monetize that.
SPEAKER_03
[SPEAKER_08] So for us our end goal is to actually sell materials. [SPEAKER_08] So what we do is we're capturing the largest experimental data set in the world because we're running experiments. And when you have that data set we have a deep belief that that is the fundamental underlying technology you need to make these models relevant. When you can put this whole flywheel together you can start to target things like R-TAP superconductors for example or high entropy ceramics or new battery cathode materials. We want to sell those materials to the industries of the future.
SPEAKER_08
[SPEAKER_03] We always like to say we don't focus on optimization based technologies. There's a bunch of companies that optimize existing products.
SPEAKER_07
[SPEAKER_08] We focus on enabling based technologies technologies that are not here today and will be here tomorrow because of radical AI. [SPEAKER_03] You will then own the patent. [SPEAKER_03] So we'll make them we'll discover the material test it and then actually manufacture and scale it. [SPEAKER_03] We have a lot of people here who maybe some of them want to know what company they should start with space they should go into. [SPEAKER_08] You want the company you wish you could be doing besides this right. Just areas where you think there's opportunity.
SPEAKER_07
Obviously Rob I mean you were in the control labs like the human machine interface is that an area you're bullish on or just spaces you think people should be running at right now. Yeah so our previous company control labs which we sold to Meta was a wearable neurotech device. So you wear it around your wrist picked up a signal known as electromyography interface to a computer and let you have very high bandwidth control over a computer interface. Whether that be VR, AR glasses it is now part of the AR glasses that Meta ships.
SPEAKER_07
And I think in general these I mean there are two trajectories one towards highly immersive technologies and then maybe more towards embodied technologies. And I've worked on either side of that but I think both of those it's core to think about the user and the person and I think a lot of times what gets lost in robotics and with immersive technologies is. Sorry I actually don't know which is which. Which one is in the brain? So what I would say is with AR we are projecting ourselves into a digital space and with robotics we are pulling out. Oh okay, I misunderstood. And robotics are embodying the AI itself. So it's two sides of the same coin.
SPEAKER_07
But in both senses you are developing technology to interface with people and there are interface problems all over the place in terms of making things that people actually want to interact with. They can be optimistic about the future of the role we play alongside AI in various forms and I think there's a lot of innovation to be had there. So just as core like how is the human interacting with all this capacity? Yeah and I think you see a lot in the devices space right now.
SPEAKER_09
[SPEAKER_07] Personal, what a personal AI is going to look and feel like. [SPEAKER_07] Is it going to be a phone or is it going to be glasses or air pods or whatever it is. [SPEAKER_07] And I think there's a lot of things to happen there. [SPEAKER_08] Do you have one you're cheering for? [SPEAKER_07] Not yet. [SPEAKER_07] Pim, space you think people should be running at? [SPEAKER_07] Yeah. [SPEAKER_07] So a few things. [SPEAKER_07] One, I very strongly disagree with the claim that we are at AGI. [SPEAKER_07] I think we have better than random idea generators that get increasingly better than random at an increasingly rapid rate.
SPEAKER_09
[SPEAKER_07] And we just so happen to have so much data on the internet that we can construct them to now do useful things so long as those things are represented largely in text space. [SPEAKER_07] Right. [SPEAKER_07] And that is such a black hole of where everything is being sucked in that companies are increasingly exposing everything in text space like tool calls and things like that. [SPEAKER_07] Because otherwise you miss out on being part of that interface.
SPEAKER_08
[SPEAKER_07] Right.
SPEAKER_09
[SPEAKER_07] So I think my definition of AGI is far from being met. [SPEAKER_07] With that said, it's incredibly difficult, even being in the space, I just find it incredibly difficult to predict what somebody should do. [SPEAKER_07] I think that people should be spending a lot of time, we're spending a lot of time thinking about how do you enable jobs or education in this new world. [SPEAKER_07] I love like half my childhood. [SPEAKER_07] It's not AGI, but everyone might lose their jobs. [SPEAKER_07] Yeah, I well, but how many jobs do people spend on their computers?
SPEAKER_08
[SPEAKER_07] Right, well, your center's very high. [SPEAKER_07] Aren't humans somewhat random and inconsistent in our reasoning? [SPEAKER_07] Yeah, I think we can match it with execution and checking whether it's correct. [SPEAKER_07] I think there's a lot of things that humans are still very special for. [SPEAKER_07] But I also see the pace of the changes of technology and I can't say that I think that that's going to be a forever thing, right?
SPEAKER_09
[SPEAKER_07] So, yeah, I don't know. [SPEAKER_07] I spent a lot of time thinking about there's just really interesting. [SPEAKER_07] There's two things consecutively broken. [SPEAKER_07] One, hiring is broken because everything is LLM generated and bullshit. [SPEAKER_08] And then second, nobody's getting hired. [SPEAKER_08] So this is one thing that I just run into every single day. [SPEAKER_08] It's okay, it's just because we're not servicing the right people the right ways. I think science is, and maybe this is my framework of thinking through this. There's a lot of fields that were enabled by text. Science is one of them.
SPEAKER_09
They sort of came after the innovation of text itself, which is largely in my head, a way that we compress real world information, right? So we see a bunch of things, we describe it.
SPEAKER_08
[SPEAKER_09] It was a room full of people, a bunch of people wear different color shirts, right? [SPEAKER_09] Eric Newcomer was there. [SPEAKER_09] We compress all that information into something that is relatable and quick to communicate. [SPEAKER_09] A lot of fields came on top of text. [SPEAKER_09] We as humans invented text, right? [SPEAKER_09] Math, physics, a lot of these things. [SPEAKER_09] And that's where you see the completely rapid innovation, really without limits, because you can build verifiers in text domain.
SPEAKER_03
[SPEAKER_09] They came after the innovation of text itself, which is largely in my head, a way that we compress real world information, right? [SPEAKER_09] So we see a bunch of things, we would describe it. [SPEAKER_09] It was a room full of people, people wear different color shirts, right? [SPEAKER_09] Eric Newcomer was there.
SPEAKER_08
[SPEAKER_09] We compress all that information into something that is relatable and quick to communicate. [SPEAKER_09] A lot of fields came on top of text.
SPEAKER_03
[SPEAKER_09] We, as humans invented text, right? [SPEAKER_09] Math, physics, a lot of these things. [SPEAKER_09] And that's where you see the completely rapid innovation, really without limits, because you can build verifiers in text domain.
SPEAKER_08
[SPEAKER_09] You can do the whole thing.
SPEAKER_07
[SPEAKER_08] So I think building things on top of the rapidly accelerating field that can be described mostly in text, where most of the communication of that field is in text, something around there is probably the right way to bet on the LLM curve or financial model curve. [SPEAKER_09] And then to your point, I think we're going to see rapid innovation. [SPEAKER_09] If you really like physical systems, now's your time. [SPEAKER_09] Go figure out how to manufacture these things and be really good at it, because you're going to get a ton of optimizations in the next year or two.
SPEAKER_07
[SPEAKER_09] You're welcome to respond to AGI if you want, but we don't need to get lost in that. [SPEAKER_09] Otherwise, just an area you think people should be running out. [SPEAKER_09] You came up through the army, right? [SPEAKER_09] So you've seen some of these problems from a totally different lens. [SPEAKER_09] Are there other areas you think people should be building companies? [SPEAKER_08] To me right now, and I work in hardware, but there are three areas I just think are most valuable in the world or accrue. [SPEAKER_09] And I think there are subsets of each area that you can go very deep within. [SPEAKER_08] There's AI and data.
SPEAKER_07
[SPEAKER_08] I think interfaces is a perfect example. [SPEAKER_08] That is a sub area of that. [SPEAKER_08] If someone can build holographic technology for my agent, I'll pay you today for that product. [SPEAKER_03] But I think we're seeing interfaces be important. [SPEAKER_03] I use an agent every day. [SPEAKER_03] It's a part of my workflow. [SPEAKER_03] The entire company is required to use agentic frameworks. [SPEAKER_03] That's the world we're going to.
SPEAKER_08
And so I think interfaces is a cool example.
SPEAKER_07
[SPEAKER_08] I'd love to see it. [SPEAKER_03] Compute and energy to me are industries that have been historically very hard to disrupt. [SPEAKER_03] Compute has had a better wave of disruption. [SPEAKER_03] There's been in-memory compute, neuromorphic compute, different alternatives trying to get ASICs based accelerators for AI. [SPEAKER_03] I think we've seen a wave there, but we're still on a Moore's law trajectory of compute. [SPEAKER_03] So what comes after two nanometer? [SPEAKER_03] Well, it's 3D heterogeneous structures. [SPEAKER_08] There's a lot of unknown problems when it comes to that. [SPEAKER_03] From the compiler down to the transistor.
SPEAKER_07
[SPEAKER_03] There's a lot of interesting work to do there. [SPEAKER_03] Energy. [SPEAKER_03] Energy had a wave and crashed in a wave because it was early in the technology. [SPEAKER_08] Probably just wasn't fully there yet. [SPEAKER_08] But now we're coming back to this place where we're energy constrained.
SPEAKER_07
[SPEAKER_08] And so what is the future of solar? Energy storage. Nuclear is certainly interesting. Of course we spend time there. I just think those areas. And I'm sure there are a million things you can think about inside of them. They're going to power the future of the world. Great. I'm going to open it up soon. Start thinking. Rob, it's easy to get lost in the science of it all. But I mean, you've sold two businesses. I feel with frontier technology, it's build the biggest business ever, change the world. What advice would you give to founders about thinking about an exit and big companies and how much do you court them?
SPEAKER_08
[SPEAKER_07] Just give us some tactical advice from your experience doing this. [SPEAKER_07] Yeah. [SPEAKER_07] Well, so I think anyone building in big tech is taking on some massive, massive opportunity.
SPEAKER_09
[SPEAKER_08] And whether it's humanoid robots, world models, materials, these are all huge categories. [SPEAKER_07] And it can be tempting to try and boil the ocean. [SPEAKER_07] But you already have huge ambition. [SPEAKER_07] Huge ambition. [SPEAKER_07] So the real challenge is how do you define a set of milestones that make sense, that can align to funding goals you have? [SPEAKER_07] How do you focus constrained resources towards a problem that is going to differentiate you from what are very competitive spaces in the early days? [SPEAKER_07] And so, for us, that was building a safe robot to be around. [SPEAKER_07] And that was true mechanically speaking.
SPEAKER_09
[SPEAKER_07] But then we invested a lot from an AI perspective in training robot behaviors that were capable but compliant. [SPEAKER_07] And so, creating a clear through line, a set of milestones, and being very deliberate on how you spread your investments when- [SPEAKER_07] Are the fundraising milestones and the acquire milestones the same? [SPEAKER_07] Because for the acquirer, it needs to be something they want. [SPEAKER_07] Whereas, I guess the funder, especially if you have investors who are willing to come back around, you can sort of pre-negotiate if we do this, you're still on the journey?
SPEAKER_08
[SPEAKER_07] I think they're similar, but you always have to play it a couple rounds ahead.
SPEAKER_09
[SPEAKER_07] And I think, in both spaces we're talking about, but in robotics specifically, it is one of the most capital intensive endeavors that one could undertake. [SPEAKER_07] And so thinking very hard about if you take this next round, what does that mean for the subsequent round and the round after that? [SPEAKER_07] What is the trajectory that gets you to either an independent, sustaining company, or are there partners that are important to develop along the way, such that you have a clear path to building the business that you set out to build? [SPEAKER_07] You could become unacquirable if you get to a certain scale. [SPEAKER_07] Yeah.
SPEAKER_09
[SPEAKER_07] All right. Questions from the audience? [SPEAKER_07] Great. [SPEAKER_07] I have a question for Rob.
SPEAKER_08
[SPEAKER_07] Here.
SPEAKER_09
[SPEAKER_07] I'll bring it to you.
SPEAKER_08
[SPEAKER_07] I have a question for Rob. [SPEAKER_07] So I've been learning about Sprout in my courses, and I find it really interesting that it's a smaller robot. [SPEAKER_07] I just wanted to ask, as compared to the average humanoid, where do you see Sprout in a person's everyday life? Yeah. So for those, Sprout is our first robot. It's three and a half feet tall, 50 pounds. It has two arms, two legs, articulated eyebrows and LEDs around the face.
SPEAKER_03
[SPEAKER_08] So it's cute and engaging, but it's also capable of interacting with doors, refrigerators, elevators, navigating autonomously. [SPEAKER_07] I have a question for Rob. [SPEAKER_07] Here. [SPEAKER_07] I'll bring it to you. [SPEAKER_07] I have a question for Rob.
SPEAKER_08
[SPEAKER_07] So I've been learning about Sprout in my courses, and I find it really interesting that it's a smaller robot. [SPEAKER_07] I just wanted to ask, as compared to the average humanoid, where do you see Sprout in a person's everyday life?
SPEAKER_03
[SPEAKER_08] Yeah. So for those, Sprout is our first robot. It's three and a half feet tall, 50 pounds. It has two arms, two legs, articulated eyebrows and LEDs around the face. [SPEAKER_08] So it's cute and engaging, but it's also capable of interacting with doors, refrigerators, elevators, navigating autonomously. [SPEAKER_08] And so I think there's a lot a robot like that could do in the home. And our starting premise was it has to be safe so that people trust it in their home to start with.
SPEAKER_03
[SPEAKER_08] And that same robot could learn to climb a ladder, right? And that might actually be the first thing that people want and are excited about and is capable of doing many of the household chores that we one day care about. [SPEAKER_08] And so that was our approach to the problem of how do you build a general purpose machine that then can be useful near term. [SPEAKER_09] Pim, I did want, what is your agent stack? Or how do you supercharge yourself as the CEO? Clearly very technical.
SPEAKER_08
[SPEAKER_09] What is, to you, the cutting edge of having the models today assist you in your job?
SPEAKER_03
[SPEAKER_09] You can't give that away, dude. [SPEAKER_09] No, yeah, and I do want to make it very clear. We are on an incredible curve, right? The system, the capabilities are incredible. [SPEAKER_09] I'm a big cloud code power user. One of the things that we're rolling out internally are systems that connect more pieces of the whole organization so that anybody can hook into cloud code. [SPEAKER_09] We also have entire, so one of the things that's interesting that we did, we have a data analyst. So we were always understaffed on data analysts.
SPEAKER_08
[SPEAKER_09] And then there was a QA person from a team. We did a hackathon where we told everybody to go build things around cloud code and also the different models that we did at the beginning of the year. [SPEAKER_09] And one of our QA team members built this thing called ML Data Buddy, where it's just a Slack bot. [SPEAKER_09] And then you can just talk to it and it's your data, right? It's SQL queries, you get it back. It's Slack, you can just talk to it.
SPEAKER_07
[SPEAKER_09] And then now he has made it so that it scales alongside the number of questions it's getting. And it feels like you're just talking to a coworker. [SPEAKER_09] And this felt very natural and very good. And we all started using it so much to the point where the database was getting overloaded because so many employees at the same time were asking highly concurrent select queries. [SPEAKER_09] And so that worked really well. [SPEAKER_09] My— [SPEAKER_09] Is that you at the company level then rather than thinking of yourself?
SPEAKER_07
[SPEAKER_09] I do it in my personal life too. So I have some automated workflows that go and download my personal files from a lot of the servers that I use. [SPEAKER_09] And then I can ask finances. I can ask cloud code to put together a report of this or this or this. And then I just give it data sources and it just goes and does it. [SPEAKER_09] It's quite helpful. I have it connect to my calendar. I use it as an interface to all the complexity of clicking around. And that's very helpful. [SPEAKER_09] Another audience question? [SPEAKER_09] Yeah. [SPEAKER_09] Okay. [SPEAKER_09] It's good.
SPEAKER_07
[SPEAKER_09] Hey, I'm Anjit. I'm CEO of Pillar Space Systems. We're building some space hardware. I had a question for Joe. How did you approach the Raytheon angle? It's really cool that they're investing in you. How did that come about? [SPEAKER_09] Yeah, so I think, as all good relationships do, it came from a business need and the customer side. So we met with their research teams around what problems are you trying to solve and what are you. [SPEAKER_09] How do you do materials today, which morphed to you should meet the venture's arm. And I think we're in a very unique time, which I might get in trouble for saying, but who cares.
SPEAKER_08
[SPEAKER_09] Innovation. And so they had a different lens to looking at a company like us, where I don't know for sure, but five years ago might have been harder to understand this early.
SPEAKER_07
[SPEAKER_09] I think later on, most likely, when we have a product and they know what we're doing. But I think their team was deeply convinced on, okay, we know AI is going to change materials. We know self-driving labs are going to make a big impact. We don't have the expertise to do that. We build jet turbines. So we should work with someone who does. And I think that was where there was a very congruent relationship. [SPEAKER_09] Hey, dude, we're never building the jet turbine. We don't want to compete with you there. We want to build materials. And yeah, great. We're not a materials company. We don't want to compete with you there.
SPEAKER_08
[SPEAKER_09] There's a really good synergy that we formed together. And I think the rest is history from there.
SPEAKER_07
[SPEAKER_09] Super interesting. Yeah. Right here. Sorry, in the front. [SPEAKER_09] Hi. Thanks for coming today. I think we all acknowledge how constrained supply chain layer is foundationally. Given how fragmented the hardware supply chain is, how are you deciding what parts of the stack you need to control versus rely on external? [SPEAKER_09] For us, you actually find the common denominators. We found that common denominator to be the game controller interface, for example. But you got to find the inner, again, you find the interfaces that actually have a simple way of deploying into them. And I think a lot of things are downstream from there.
SPEAKER_07
[SPEAKER_09] And sometimes you have to create the interface, the robots. And that's really hard.
SPEAKER_07
[SPEAKER_08] It's cliche, but true, first principle analysis. Well, if there are tools that we buy that have 36 month lead times. So the first question we ask is, what is the benefit of acquiring the tool? Why are we acquiring the tool? What is the use case of the tool in the future? If it makes sense to custom build that, or make some type of alteration to use it quickly, because we see it has x, y, and z use case in our business moving forward, then that makes sense. If it's a one off, or we don't have any other people that want that tool, we just have to do that for this contract, you're less inclined to want to go take on all that. That's so much technical risk, capital expenditure, and even focus. And I think one of you mentioned focus earlier. You can boil the ocean very quickly. You start talking to customers in deep tech. And next thing you know, you have 17 different problems to try to solve. You need to be very good at that. That's a cool problem. But we don't work on that. That's a cool problem. We don't work on that. And I think you are constantly doing that as a founder. If you are not doing it as a founder, I think you're actually failing the organization, particularly early on, because the best skill is saying no. And I think you have to be really laser focused on we are solving this because of this if a.
SPEAKER_07
[SPEAKER_08] founder I think you're actually failing the organization particularly early on because the best skill is saying no right as the saying goes and I think you have to be really just laser focused on we are solving this because of this if a hardware requirement meets solving that I will build it if it doesn't I'm going to find another way to get it that's the way we think about it all good succinct audience questions I love this crowd so far who's next all right yeah I was wondering how you think about with materials for example the kind of throughput constraints that exist on building stuff out and so like for example I think Google DeepMind had 2.2 million materials they discovered they think 40,000 are stable it's like okay how do I make all of these and then how do I convince people that they work right and to your point around we're using a hundred year old alloys I think part of it is that people know the stress strain curve and they've used it for a hundred years and it's like even now that we have 40 different concrete alternatives you can't build a house with them because you can't get that in building code and it's like how do you think about that when you're talking about critical industries like defense like oh you should use this new alloy I've invented that trust me it works and we should put pilots and stuff in there yeah really good question I'll touch both parts so on the latter part is called qualification in the aerospace industry and for those in the audience qualifications essentially the process of confirming what this gentleman said which is hey we feel safe to fly that like that's going to fly in a fighter jet that's going to fly in a manned aircraft or it's going to fly on something else and ironically I think in some of those industries you're actually seeing the companies who make the end products vertically integrate and we were surprised by this when we went to market we thought we would have to vertically integrate immediately to get material to production and instead they're actually vertically integrating from the top down so these aerospace companies are saying we are tired of dealing with the qualification requirements of the US government or any other administration or organization we're going to do it ourselves and I can list off the top of my head multiple companies who you are all aware of that have material teams in-house bypassing qualification for their own systems now in air which means the government then certifies the whole so there's these lists that the government has I think the FAA runs there's a military spec list as well for manned aircraft as well as missiles that control like if something is qualified to be used in those systems and so obviously like for commercial aircraft you have they're going to follow that requirement they're the slowest least innovative people in the world but like for aerospace space they do not in some of their systems need to follow those they're launching their own rocket on their own dime and as long as they're doing that in a way that they can prove it's safe they have already done own custom alloy development I've seen it before so the qualification piece I think is you're seeing this again the stack come the other way and the throughput will be the bottleneck there and how much they can actually make because they're not material companies they're aerospace companies so that's the second one on the first one the throughput side is the hardest part in science and so one of the reasons that that genome or matter gen or any of these papers that have come out have been poor is they're trained on computational data only and computational data is cool it's fun to simulate electrons but it's not the real world and if you call any customer in the world and say look at this amazing material I made you on my computer screen they'll say yeah cool can you send it to me in the mail can I use it can I actually put it in my product and if the answer is no to that then I don't think you've actually discovered a new material and so when you see these models coming out with new materials they're not new materials they're new ideas of materials and until you can make them you don't have new ones that's why self-driving labs are super important that's how you solve the throughput problem sure hey thanks for the thoughts here the LLM revolution was powered primarily by publicly available data sets there were mostly accessible to a lot of the labs and a lot of the people when we switch to physical intelligence robots world models how do you see the data bottleneck how do you overcome it how does it how do the use cases map to data so we hear about game data and so on but in games primarily it is first-person shooter and so on and if I want to teach my robots how to lift boxes in a warehouse how do you actually map that use case to what data you need how do you get that yes so that's why world models are so important because you don't want to train a model to predict violent actions but you also want the model to be able to simulate environments that can do that so that you can evaluate your agent against it and you want to train on the positive actions and evaluate on the negative ones right so that's why you don't just run large scale imitation learning and then shoot that agent into the real world you use world models such that you have an environment that understands all the possible unfolding dynamics and then you use an agent and reward them specifically for the behaviors that you actually want to see so it's actually the counterpoint that argument is actually that you need games to simulate the extreme scenarios that you want to avoid against now I would argue that the entire field is still in the research phase there are no known scaling laws that are as clear as in text where you can just predict what type of data you need for what type of thing so that's the difference I think everyone's still figuring it out it's not like Tesla inhales all this data and has a huge advantage they do and it's for sure useful but I think the general solve which is what I think you're asking the general solve like LLMs those scale like there are scaling laws less for specific embodiments for specific situations like driving is better than others but the general solve is not as clear as LLMs I think there's ideas yeah I mean I think what is working today is in places like self-driving cars where it's a fairly straightforward controls problem and a more complicated environmental spatial understanding problem and you can build a model around that but the same is true in robotics like the things that people get working are actually quite narrow stair climbing door opening packing a box like these aren't one model that solves all of these problems and you know that can either be with real-world data we rely a lot on simulation is that true with physical intelligence I mean they present it as very broad but they will you don't think the folding laundry is generally applicable I mean I don't
SPEAKER_07
[SPEAKER_08] ideas yeah I think what is working today is in places like self-driving cars where it's a fairly straightforward controls problem and a more complicated environmental spatial understanding problem and model you can build around that but the same is true in robotics. The things that people get working are actually quite narrow: stair climbing, door opening, packing a box. These aren't models that solve all of these problems. And that can either be with real-world data—we rely a lot on simulation. Is that true with physical intelligence? I mean they present it as very broad but I don't think the folding laundry is generally applicable.
SPEAKER_07
[SPEAKER_09] I don't think it's solved and I don't think we yet understand the scaling laws associated with current approaches to say that with just a bit more data laundry is solved. I agree fully. If we have some time I can try to go into how to think about the scaling laws but it might take me a few minutes. You tell me—it's hard to preview if this is where I don't know. Give us just a couple bullets and people can follow up with you. So there's complexities related to the degrees of freedom of the robot, right? There's humanoid—it's like large degrees of freedom. Self-driving system is like very low degrees of freedom. And so your scaling laws are different depending on the degree of freedom. Then within the degree of freedom you can have different types of predictions, right? So if you're predicting at the level of a game controller you can just predict a bit or a digital value. But if you go for what I mean, you might need to predict motor torque or something. And so every degree of freedom has different values. So there's complexity there. Then there are scaling laws on the degrees of freedom and the input space. Then there are scaling laws on the environment right—where the robot operates. And then there are scaling laws on the sensors themselves. So anybody claiming they have a very general solution is full of shit.
SPEAKER_07
[SPEAKER_08] Yeah, I love this crowd. A very New York thing, yeah. [SPEAKER_09] All right, thanks. So you have a material science company and one thing that you said about cars 100 years ago—they used all these old alloys but they only needed five or six elements. And then a car today needs 150. Can you talk about where the raw materials are coming from and how those raw materials get into the pipeline of what you're doing in your company?
SPEAKER_07
[SPEAKER_08] Really good question. Probably the question I get the most when I go to Washington DC and I'm on the hill or at the Pentagon or at the Department of Energy—where the heck do you source those things? And I think you're seeing a big focus on critical minerals. It's the derived products category of critical minerals that is super interesting and there has been a focus from DC on what are these things actually going into. So I'll give you a perfect example. C103 is a super common alloy in aerospace and defense. It's got cool properties. It's made with hafnium and it has a high weight percentage of hafnium. Hafnium is almost entirely—I think it's like 90-something percent owned by China. So now DOE and DOD are very worried about hafnium content in their materials. So not only is there a question of can you make a new alloy that's got way better performance, we actually get questions like can you make the same alloy but with no hafnium at all? And we've been able to do that. We've done that successfully. But that is almost a different materials problem because now you're trying to hit the same property constraints but just remove the inputs. And that's a different problem than different inputs with maybe different property constraints that allow you to extrapolate the periodic table more. We can use a refractory metal to drive strength. We can't use a refractory metal in C103 or a new refractory metal in C103. So critical minerals are going to be really important.
SPEAKER_07
[SPEAKER_09] You're saying your philosophy is mostly to try and cut out minerals you don't want rather than find new ones?
SPEAKER_07
[SPEAKER_08] There are places where it's really hard to get rid of them. Nickel-based superalloys just have elements that we don't control. And I think that's why you see an effort in dollars and on-shoring because there has been an acknowledgement: no, we probably cannot get rid of them. We better onshore or near-shore that technology rapidly. There are some elements where I think there is a push to get rid of. I think people have shown that niobium-based alloys in the high entropy field had really cool properties and there might be materials in there. I can guarantee there are. I know what some of them are that don't have hafnium. And I think the field was convinced on that and it's kind of where it's pushing today. So it depends is the unfortunate real answer.
SPEAKER_07
[SPEAKER_09] One last question. I think I've been avoiding you. So in each of your sectors, how much do you worry or used to worry about competition from mega-capitalized labs and corporations?
SPEAKER_07
[SPEAKER_08] Zero. I have a very unique view on competition. I rarely think about competition. I just think it is such a waste of mental brain space. You have one job as a team: to execute on what your mission is and be the best in the world with that. I pay attention to what our competitors do. I'm not naive to what the market's doing. I'm just not drooling over the fact that they've come out with a new release. I mean, otherwise what the hell are you doing all day long? We're doing the same thing. So you have near competitors and startups. I think you should be aware. Like big tech or big labs in very unique spaces, I just think they have no chance of competing in more generalized spaces. Where they have raw compute or raw data—like ads—you know, ads is probably a tough space to compete with Meta and YouTube. I don't know. I don't work in ads. But in materials, I don't think Meta is going to compete with us in materials or Microsoft or Google or really any of them. Personally, that's just a personal view.
SPEAKER_07
[SPEAKER_09] How much do you think about competition? I think why Anthropic is winning is because they're close to their customers and they put their evals and model training runs—how much do you really care about the 1% benchmark versus what your customers are actually using? They just move faster and close to customers in a way that OpenAI and Google can't because they are larger, bloated organizations that are more distracted. Your answer specifically worried me about Anthropic. It's between the lines I'm giving you. It's an example of how Anthropic was not the incumbent in the space that is now looking like they're actually winning, right? They were in that position and they won by focusing on their customers and ignoring everything else and by focusing on code. And I would argue—and I don't know them, I didn't work on Anthropic—that they did that by not worrying about what OpenAI was doing and like...
SPEAKER_07
[SPEAKER_09] The 1% benchmark versus what your customers are actually using there—they just move faster.
SPEAKER_07
[SPEAKER_09] Customers in a way that Open AI and Google can't because they are larger or bloated organizations that are more distracted. Your answers—I specifically worry about Anthropic is between the lines I'm giving you. It as an example, Anthropic was not the incumbent in the space that is now looking like they're actually winning. They were in that position and they won by focusing on their customers and ignoring everything else and by focusing on code. And I would argue, and I don't know them, I didn't work on Anthropic, that they did that by not worrying about what Open AI was doing. Customers, customers, customers—what are they? What are people? Well, you're the flip side of big companies can be good. You gotta be super careful with this answer. I think there are things that small companies can do that big companies really struggle to do. I said it before—focus and attention and a very opinionated product is very hard to come by in the earliest stages at big companies. So I think you can build up momentum on a trajectory and you got to be right about where you're choosing to focus, that that is a problem that matters and it'll be differentiated. But pushing hard on that thing is something that only startups I think can do very effectively. So we got to a point where I think we had enough of a kindling fire that it made sense to have a big partner and to help that grow. But in the earliest days it's about how do you get that spark into something meaningful, and that's harder for big companies to. Couldn't imagine a better last word than that—an ode to startups. All right, this is awesome. Great crowd, thank you.
SPEAKER_07
Yeah. All right. Questions from the audience? Great. I have a question for Rob. Here. I'll bring it to you. I have a question for Rob. So I've been learning about Sprout in my courses, and I find it really interesting that it's a smaller robot. I just wanted to ask, as compared to like the average humanoid, where do you see Sprout in like a person's everyday life?
SPEAKER_08
Yeah. So for those, Sprout is our first robot. It's three and a half feet tall, 50 pounds. It has two arms, two legs, articulated eyebrows and LEDs around the face. So it's cute and engaging, but it's also capable of interacting with doors, refrigerators, elevators, navigating autonomously. And so I think there's a lot a robot like that could do in the home. And our starting premise was, you know, it has to be safe so that people trust it in their home to start with.
SPEAKER_08
And that same robot could learn to climb a ladder, right? And that might actually be the first thing that people want and are excited about and is capable of doing many of the household chores that we one day care about. And so that was our approach to the problem of how do you build a general purpose machine that then can be useful near term.
SPEAKER_09
Pim, I did want, like, what is your, like, agent stack? Or like, what is your, how do you like, supercharge yourself as the CEO? Like, clearly very technical. Like, what is, to you, the cutting edge of having, like, the models today, like, assist you in your job? You can't give that away, dude.
SPEAKER_09
No, yeah, and I do want to make it very, very clear. Like, we are on an incredible curve, right? Like, the system, the capabilities are incredible. I'm a big, like, cloud code power user. I, like, one of the things that we're rolling out internally are systems that, like, connect more pieces of the whole organization so that anybody can hook into cloud code. We also have entire, so one of the things that's interesting that we did, like, we have, like, a data analyst. So we were always understaffed on data analysts.
SPEAKER_09
And then there was a QA person from a team. We did, so one thing that we did is we did a hackathon where we told everybody to go build things around cloud code and also the different models that we did at the beginning of the year. And one of our QA team members built this thing called ML Data Buddy, where basically it's just a Slack account, like a bot object. And then you can just, you can just talk to it and it just is your data, right? It's SQL queries, you get it back. It's Slack, you can just talk to it. And then now he has made it so that it scales alongside, like, the number of questions it's getting. And it feels like you're just talking to a coworker.
SPEAKER_09
And this is just felt very natural and very good. And we all started using it like crazy to the point where, like, database was getting overloaded because so many employees at the same time were asking, like, highly concurrent, like, select queries. And so that worked really, really well.
SPEAKER_09
My, I... Is that you like things that are at the company level then rather than thinking of yourself? I do it in my personal life too. So I have a lot of, I have some automated workflows that, like, go and download my personal files from a lot of the servers that I use. And then I can ask, like, finances. I can ask cloud code, like, put together, like, a report of, like, this or this or this. And then I just give it data sources and it just goes and does it. It's quite helpful. I have it connect to my calendar. Like, I use it as an interface to, like, all the complexity of clicking around. And that's very, very helpful. Another audience question? Yeah. Okay. It's good.
SPEAKER_09
Hey, I'm Anjit. I'm CEO of Pillar Space Systems. We're building some space hardware. I had a question for Joe. How did you approach the Raytheon angle, I guess? It's really cool that they're investing in you. How did that come about? Yeah, so I think, as all good relationships do, it came from a business need and, like, the customer side. So we met with their research teams around what problems are you trying to solve and what are you... How do you do materials today, which morphed to you should meet the venture's arm. And I think we're in a very unique time, which I might get in trouble for saying, but who cares.
SPEAKER_09
Innovation. And so they, I think, had a different lens to looking at a company like us, where I don't know for sure, but five years ago might have been harder to understand this early. I think later on, most likely, like when we have a product and they know what we're doing. But I think their team was deeply convinced on, okay, we know AI is going to change materials. We know self-driving labs are going to make a big impact. We don't want to have the expertise to do that. We build jet turbines. So we should work with someone who does. And I think that was where there was a very congruent relationship.
SPEAKER_09
Hey, dude, we're never building the jet turbine. Like, we don't want to compete with you there. We want to build materials. And yeah, great. We're not a materials company. We don't want to compete with you there. There's a really kind of good synergy that we formed together. And I think the rest is history from there. So. Super interesting. Yeah. Right here. Sorry, in the front. Hi. Thanks for coming today. I think we all acknowledge how constrained supply chain layer is foundationally. Given how fragmented the hardware supply chain is, how are you deciding what parts of the stack you need to control versus rely on external?
SPEAKER_09
For us, you actually find the common denominators. We found that common denominator to be like the game controller interface, for example. But you got to find like the inner, again, you find the interfaces that actually have a simple way of deploying into them. And I think a lot of things are sort of downstream from there. And sometimes you have to create the interface, like the robots. And that's really, really, like, you know, hard. And then, yeah.
SPEAKER_08
It's cliche, but true, like first principle analysis. Well, if there are tools that we buy that have 36 month lead times. So the first question we ask is like, what is the benefit of acquiring the tool? Why are we acquiring the tool? Like, what is the use case of the tool in the future? If it makes sense to custom build that, or like, or make some type of alteration to use it quickly, because we see it has x, y, and z use case in our business moving forward, then that makes sense. If it's like a one off, or like, we don't have any other
SPEAKER_08
people that want that tool, we just have to do that for this contract, you're like less inclined, to want to go take on all that. That's so much technical risk, capital expenditure, and even focus. And I think one of you mentioned focus earlier, like, you can boil the ocean very quickly, like, you start talking to customers in deep tech. And next thing you know, you have 17 different problems to try to solve, you need to be very good at, that's a cool problem. But we don't work on that. That's a cool problem. We don't work on that. And I think, you are constantly doing that as a founder, like, if you are not doing it as
SPEAKER_08
founder I think you're actually failing the organization particularly early on because the best skill is saying no right as the saying goes and I think you have to be really just laser focused on we are solving this because of this if a hardware requirement meets solving that I will build it if it doesn't I'm gonna find another way to get it that's the way we think about it all good succinct audience questions I love this crowd so far who's next all right yeah I was wondering how you think about with like materials for example the kind of just like throughput constraints that exists on building stuff out and so like for
SPEAKER_03
example I think Google deep mind had like 2.2 million materials they discovered they think 40,000 are stable it's like okay how do I make all of these and then how do I convince people that they work right and to your point around like we're using a hundred year old alloys I think part of it is that people know the stress
SPEAKER_08
strain curve and they've used it for a hundred years and it's like even now
SPEAKER_03
that we have like 40 different concrete alternatives like you can't build a house with them because like you can't get that in building code and it's like how do you think about that when you're talking about critical industries like defense like oh you should use this new alloy I've invented that like trust me it works and like we should put you know pilots and stuff in there yeah really good question I'll touch both parts so on the latter part is called qualification in the aerospace industry and for those in the audience qualifications essentially the process
SPEAKER_08
of confirming what this gentleman said which is hey we we feel safe to fly that like that's gonna fly in a fighter jet that's gonna fly in a manned aircraft or it's gonna fly on something else and and ironically I think in some of those
SPEAKER_03
industries you're actually seeing the companies who make the end products vertically integrate and we were surprised by this when we went to market we thought we would have to vertically integrate immediately to like get material to production and instead they're actually vertically integrating from the top down so these aerospace companies are saying we are tired of
dealing with the qualification requirements of the US gov or or any other administration or kind of organization we're gonna do ourself and I can you know list off the top of my head multiple companies who you are all aware of that have material teams in-house bypassing qualification for their own systems now in air which means the government then certifies the whole yeah so there's like these like lists that the government has I think FAA runs run there's a military spec list as well for manned aircraft as well as like missiles that control like if something is qualified to be used in those
SPEAKER_08
systems and so obviously like for commercial aircraft like you have they're gonna follow that requirement they're the slowest least innovative people in the world but like for aerospace space they do not in some of their systems need to
SPEAKER_03
follow those they're launching their own rocket on their own dime and as long as they're doing that in a way that they can prove it's safe they have already done own custom alloy development I've seen it before so the qualification piece I think is you're seeing this again the stack come the other way and the throughput will be the bottleneck there and how much they can actually make because they're not material companies they're aerospace companies so that's kind of the second one on the first one the throughput side is the hardest part in science and so one of the reasons that that genome or matter gen or any of these
SPEAKER_03
papers that have come out have been poor is they're trained on computational data only and computational data is cool like it's fun to simulate electrons but it's
SPEAKER_08
not the real world and if you call any customer in the world and say look at this amazing material I made you on my computer screen they'll say yeah cool can you send it to me in the mail like I can I use it can I actually put it in my product and if the answer is no to that then I don't think you've actually
SPEAKER_09
discovered a new material and so when you see these models kind of coming out
SPEAKER_07
with new materials they're not new materials they're new ideas of materials and until you can make them you don't have new ones that's why self-driving labs are super important that's how you solve the throughput problem sure hey thanks thanks for the thoughts here LLM revolution was powered primarily by publicly available data sets there were mostly accessible to a lot of the labs and a lot of the people when we switch to physical intelligence robots world models how do you see the data bottleneck how do you overcome it how does it how do the use cases map to data so we hear about game data and so on but in
SPEAKER_07
games primarily it is first-person shooter and so on and if I want to teach my robots how to lift boxes in a warehouse how do you actually map that use case to what data you need how do you get that yes so that's why world models are so important because you don't want to train a model to predict violent actions but you also want to model to be able to simulate environments that can do that so that you can evaluate your agent against it and you want to train on the positive actions and evaluate on the negative ones right so that's why you don't just run like large kill imitation learning and then shoot that agent into
SPEAKER_07
the real world like the use world models such that you have an environment that understands all the possible unfolding dynamics and then you use an agent and
SPEAKER_08
and reward them specifically for the behaviors that you actually want to see so it's actually the counterpoint that argument is actually that you need games
SPEAKER_07
to simulate the extreme that the scenarios that you want to avoid against now I would argue that the entire field is still in the research phase like there are no there are no like and again I'm kind of somewhere artistic like like no known scaling laws that are like there's some but not like in text world where you can just use can predict like what type of data you need for what type of thing so so so so that's the difference I think everyone's still figuring it out you know it's not like Tesla inhales all this data and a huge advantage you don't think there's they do they do and it's for sure useful but I think the the general solve which is what I think
SPEAKER_08
you're asking the general solve like LMS those skilling like there are skilling less for specific embodiments for specific situations like driving is better than others but the general solve is not as clear as LMS I think there's ideas yeah I mean I think what is working today is in places like self-driving cars where it's a fairly straightforward controls problem and a more
SPEAKER_07
complicated environmental you know spatial understanding problem and model you can build around that but the same is true in robotics like the things that people get working are actually quite narrow stair climbing door opening packing a box like these aren't one models that solve all of these problems and you know that can either be with real-world data we rely a lot on simulation is that true with like physical intelligence I mean they present it as very broad but they'll you don't think the folding laundry is like generally applicable I mean I don't
SPEAKER_09
think it's solved and I don't think we yet understand the scaling laws associated with current approaches to say that with just a bit more data laundry is solved I agree fully agree I can if we have some time I can try to go into how to think about the scaling laws but it might take me like a few minutes you tell me it's it's it's hard to preview if this is where I don't know give us just like a couple bullets and people can follow up with you so there's complexities related to the degrees of freedom of the robot right there's come and then so humanoid it's like large degrees of freedom self-driving system is like very low degrees
SPEAKER_09
freedom and so your scaling laws are different depending on the degree of freedom then within the degree of freedom you can have different types of predictions right so if you're predicting at the level of a game controller you can just predict like a bit or like a digital value like you but if you go for what I mean might need to predict like motor torque or something and so every degree of freedom has like different values that so there's complexity there then so there's scaling laws on the degrees of freedom and the on the input space then there's killing laws on the environment right so the where the robot operates and then
SPEAKER_09
there are scaling laws on the sensors themselves and so like anybody kind of claiming they have a very general service like full of shit yeah I I love this crowd a very yeah New York is doing yeah I should okay great enthusiastic go for it all right thanks oh so you have a material science company and one thing that I you said about like cars 100 years ago they use all these old allies but they also only needed like five or six elements and then a car today needs like 150 and so can you talk about where the raw materials are coming from and how those raw materials
SPEAKER_08
get into the pipeline of what you're doing in your company really good question probably the question I get the most when I go to Washington DC and I'm on the hill or at the Pentagon or at the Department of Energy like where the heck do you source those things and I think you're seeing a big you know critical minerals has
SPEAKER_01
been talked about area a lot it's the derived products category of critical minerals that is super interesting and there has been a focus from DC on like what are these things actually going into and so I'll give you a perfect example you know C103 super common alloy in aerospace and defense it's got cool
SPEAKER_07
properties you can read about it it's made with hafnium and ha it has a high weight percentage of hafnium hafnium is almost entirely I think it's like 90 something percent owned by China so now DOE and DOW are very worried about hafnium content in their materials so not only is there a question of can you make a new alloy that's got way better performance we actually get questions like can you make the same alloy but with no hafnium at all and and we've been able to do that like we've done that successfully but that is almost a
SPEAKER_09
different materials problem because now you're trying to hit on the same property constraints but just remove the inputs and that's a different problem than different inputs with maybe different property constraints that allow you to kind of extrapolate the periodic table more we can we can go use a refractory metal to drive strength we can't use a refractory metal in C103 or like a new refractory metal in C103 so critical minerals are contained to be really important you're saying your philosophy is mostly to try and like cut out minerals you don't want rather than find so you can you can there there are
SPEAKER_09
places where it's really hard to like nickel based super alloys they just have elements that we don't control and I think that's why you see like an effort in dollars and on-shoring because there has been an acknowledgement no we we probably cannot get rid of them we better onshore or near shore that technology rapidly there are some elements where I think there is a push to get rid of I think people had shown that niobium based alloys the high entropy field had really cool properties and there might be materials in there I can guarantee there are I know what some of them are that don't have half new and I
SPEAKER_09
think the field was convinced on that and it's kind of where it's pushing today so it depends is the like unfortunate real answer one last question I think I've been avoiding you so in each of your sectors how much do you worry or used to worry about competition from mega capitalized labs and corporations
SPEAKER_08
zero I have a very unique view on competition I rarely think about
SPEAKER_06
competition I just think it is such a waste of mental brain space and that you have one job as a team to execute on what your mission is and be the best in the world with that I pay attention to what our competitors do I'm not naive to what the market's doing I'm just not like drooling over the fact that they've come out with a new release like I hope you did I mean otherwise like what the hell are you doing all day long I mean we're doing the same thing and so you have near competitors and startups I think you should be aware like big tech or like big labs in very unique spaces I just think they just have no chance of competing in more generalized
SPEAKER_06
spaces where they have raw compute or raw data like ads you know ads is probably a tough space to compete with with meta and YouTube I don't know like I don't work in ads but in materials I don't think meta is gonna compete with us in materials or Microsoft or Google or really any of them personally that's just a personal how much do you think about the competition I
SPEAKER_09
think sort of why is anthropic winning it's because they're close to their customers and they like put their evals and like model training runs and they like how much do you really care about like the 1% benchmark versus what your customers are actually using there they just move faster customers in way that open AI and Google can't because they are larger or bloated organizations that are more distracted your answers I specifically worry about anthropic is between the lines I'm giving you it as an example of like anthropic was not the incumbent in the space that is now looking like
SPEAKER_09
they're actually winning right they were in that position and they won by focusing on their customers and ignoring everything else and by focusing on code and I would argue and I don't know them I didn't work on an anthropic that they did that by like not worrying about what opening I was doing and like customers customers customers what are they what are people like what are people like well I'm you're sort of the flip side of big companies can be good you gotta be super careful with this answer I think there are there are things that small companies can do that big companies really struggle to do I said it before but
SPEAKER_09
focus and attention and like a very opinionated product is is very hard to come by in the earliest stages at big companies and so I think you can build up momentum on a trajectory and you got to be right
SPEAKER_08
about where you're choosing to focus that that is a problem that matters and it'll be differentiated but
SPEAKER_09
pushing hard on that thing is is something that only startups I think can do very effectively so we got to a point where I think we had enough of kind of a kindling fire that you know it made sense to have a big partner and to help that grow but in the earliest days it's about how do you get that spark into something something meaningful and and that's harder for big companies to couldn't imagine a better last word than that ode to startups all right this is awesome great crowd thank you you you you you