The Robotics Inflection: Why This Time Is Different (ft. Joe Harris, Alloy)

These robotics companies that have a big problem, robotics is incredibly difficult, it's hard to make a robot that works, so we want to just take one of those hard things away. But I think that robotics will not necessarily have to be a consumer hardware business. There is so much opportunity in the enterprise and business space. Help us to imagine how soon and how big the robotics wave will actually be. I'm not sort of here to say, "Oh, you're going to have a humanoid in your house in one year." I'm not one of those people. I think we're actually massively overstating what that's going to look like in one to two years. We probably underestimate what it looks like in 10-15 years. Every founder begins with a question that won't leave them alone. For Joe Harris, that question was simple but profound. Why do machines still fail in predictable ways? Joe's story is full circle, from an electrical engineer to growth leader at Eucalyptus. And now founder of Avoi. A company building the missing feedback loop for robotics. A lowe helps robots see themselves more clearly. It captures the 1% of data that truly matters turning noise into understanding and understanding into reliability. The world of robotics is shifting. Senses are cheaper, computers faster, models are smarter. But the real constraint isn't hardware or capital. It's the ability to learn fast enough. That's the moment a lowe exists for. It's the system that helps every robotics company move up the reliability curve. From it works, sometimes to it works every time. Joe's thinking is rare. He doesn't chase markets. He notices the quiet and flexion points before they compound. He left one of Australia's most admired startups, not out of restlessness, but out of curiosity. A curiosity that keeps folding back on itself until it becomes conviction. In this conversation, you'll hear what it sounds like when it found a sees the future coming into focus. Not in hype cycles or headlines, but in machines of how something actually works. We talk about finding the truth inside systems, why great teams are built around people who want to learn faster than the world changes, and how a company can be both technically audacious and philosophically grounded. Joe also has this rare ability to zoom out. He sees robotics not as automation, but as a deflationary force, or even a lever of abundance. The idea that by making machines more reliable, we make life itself more affordable, more creative, and more human. If you miss Joe's first appearance as the operator obsessed with growth at Eucalyptus, go back and listen. It's an absolute masterclass. It's also a wildtards first, an operator to founder. This episode is a story of someone stepping into the founder's seat and building a company that just might accelerate the next industrial revolution. Let's dive in. First guest to go from operator to founder and wildhards. Welcome back. Are you not just a growth guy? I get that a lot. I get that. So you? Frequently. I mean, that's been what I think a lot of people came to when they met me for the first time. Or someone has come to know me. It's knowing of me as the growth guy from Eucalyptus. Say the last time I was on this being an operator running the growth team, running the product team, that was my identity for some time. What I think a lot of people don't know is I'm actually an electrical engineer. And almost 10 years ago, I did my thesis in machine learning for telecommunications. Which actually, so happens, has come back around and that's very similar to what I'm doing now. And before Eucalyptus, I worked at Atlassian. I was a software engineer. Working on DevTools. Again, not too dissimilar to where I've ended up back. I say that Eucalyptus experience was more so applying that systems thinking and systems design to a bunch of novel domains where I've got to learn so much from Tim and Charlie and Benny and Alexey and just watching them and helping them build that company that I've now got to bring along with that technical expertise from before into this new company at Alloy. When did you know you wanted to start a company? I think it's always been an idea. I come from a family of being self-employed, not necessarily in big tech companies or something like that, but just generally being normal to start a business, employ people, make your own money, work incredibly hard to pull that off, small businesses mostly. And so I made my first money online when I was about 12. What did you make? I was making HTML websites for small businesses around the area, making a couple of hundred bucks each time. And it was this first moment of like wow this is quite a lot of money for a 12-year-old. Teaching myself HTML for dummies out of a textbook. I also made some YouTube videos back in 2009. No way. Is that channel still up? I've hidden a lot of those squeaky voice videos. But racked off a few million views. And it was coding tutorials. It was like how to write Python, how to write Java, making games, Photoshop tutorials, After Effects, Cinema 4D. I was just dabbling. I was like teaching myself, making tutorials, teaching myself, making tutorials, watching other tutorials, making tutorials. Great art has steel, as they say. And that started to make some money. And it was again that that was the first time I made money where I wasn't actively doing the work anymore. I made the videos. And then there was trail every month. I just get a check from YouTube. I was like, oh this is quite powerful actually. There's some leverage here. I just think that was like a seminal unlock where I was like, I'm going to be my own boss. That's kind of what I'm going to do. Also it doesn't help I'm a terrible employee. I just don't want to be a terrible employee. I did not believe that for a second. Obviously, it went well at Euclifters, but I think I got a lot of autonomy there and a lot of trust. I think generally probably my managers at previous roles would probably attest that I've quite a difficult employee. Just that I want to move quickly. I want to test and learn. I don't want the answer of why we can't do something to be something bureaucratic. And as a certain scale, it becomes hard for that not to be a frequent reason why you can't do something. And so, yeah, kind of taking no for an answer, not being my strong suit. I think when I get told I can't do something, I often want to do it more. What businesses have you started after the age of 18? Yeah. A wide variety. There's a yoga studio. It's one one. I had an agency that did a bunch of smart contract development for Ethereum, EVM based blockchains around the time, 2021, 2022. Lots of NFTs going on. There's lots of projects that had sold NFTs with a promise of building some core technology and then had no way to deliver that core technology. And so, we were the people that they brought in to say, "Hey, please help us. Build this thing, this roadmap that we have promised." And I think I just started to get really interested in the technology and the opportunity to pull me into it and build the team out around that, train them on how to do it. And that was quite a moonshot opportunity in that. We did over a million dollars of revenue in that first year because there was so much opportunity and so a few people that knew how to do it. Given the wide variety of things you have started, when do you decide to say, "Actually, I'm going to build something and dedicate a huge amount of time and energy and resources into making this world class for what it is?" Well, I also had a creatine gummy brand. I've never been to that. Who hasn't? What's some creatine gummies between friends? Obviously, I don't know. A lot of people was a little side story. They'll have seen that that whole industry has been massively side-swiped recently. But that was actually the original insight of why we got into it. We saw that some of these big-name brands, like we got them tested, they had no creatine in. You eat them and you're like, "Oh, this is delicious. This is such a great experience. I can't believe I can have my creatine this way. Shokhara, you can't. It's a scam." So we formulated one that we felt was good, tasted good. It was very challenging, balancing the consistency with the dosage. Creatine is a hard thing to suspend in a gelatinous substance because it's quite a low concentration active. So I'm getting like, really nerdy on creatine gummies. But I think that's always been my mentality. Whatever it is, I'm going to go down to brass tax. I'm going to understand how it works and I'm going to try to make it the best it could be. But I think that I was always stretching myself across so many things that I don't think I ever felt that I did it justice. I didn't do it to a degree that I felt proud of as a legacy or something that I felt would have a lasting impact. So I've always had this question in my mind, what could I do if I dedicated myself to one thing? Well, right there, you just shared one curiosity, two, three years and three, get to the bottom of the root cause of something, and then three or four say it through. And let's take a yoga, for example. Why was holding you back for or even a creatine gummi example? What held you back from doing that full time? I was doing, I was already committed to a lot of other things. I mean, I spun down the creativity gummi and sold it onto a great home and it's still functioning today. It's a lasting business that hopefully people are very happy with. But I spun that down to do this. I think it was, I think it was the magnitude of the opportunity. It wasn't something that, like I was passionate about it because I really am passionate about the benefits of creatine and that more people should putake and that they will putake and it's going to be a big growth industry and is already. But it wasn't something that I felt I could work on for 10 or 20 years, at least. And that's what I was seeking. That's what I was really looking for. That was why I left you to know other reason, right? Was that I wanted my thing that I can go and do, my company that I can start, that's going to be a multi-decade journey. - How did you discover what that multi-decade journey would be? That's a big commitment. (laughs) - Yeah, and as I've established, I, you know, commit myself to so many different things, I seem to not be able to commit to one thing. I didn't know when I left Eucalyptus exactly what it would be. I just knew I needed to burn the ships because it was such a great opportunity to Eucalyptus that if I stayed there, I'd just keep staying there. And I status quo inertia. And so I actually, I gave very, very advanced notice, did all of the succession planning, the strategy setting, the execution, make sure that it was in a good spot so that I felt like I could walk away with my head held high, having sort of end to end had a great run. I think people will really only remember how you start something and how you finish it. And so many people finish up a great run, just kind of drifting and maybe become a bit dejected and a bit disgruntled. And they leave a bad taste. And actually that's what people remember. That was very important to me. But then I allowed myself, you know, I had a great opportunity, I worked with Immutable and helped them build out sort of a new layer of their strategy, built up a team, made some hires, get that ball rolling. On a contract while I was able to sort of spend some time reflecting 'cause it's been four plus very dedicated, hardcore years at Eucalyptus that I needed some oxygen for my creativity. And so I just followed my curiosity. And that led me to creating gummies. It led me to a variety of things. And one of those things was that I was going to be able to finally go back to my kind of original hopes and aspirations graduating as an electrical engineer, which was at the timing for robotics was incredibly ripe. And now I was going to be able to, you know, spend the next couple of decades and it would be the right time for that finally. - Share more about the creative process that you went through to surface what became a lawy. I'm curious, like, take us into the bedroom where you're getting creative, you're at your best, you're on your creating gummies. - That's happening in the bedroom. - Well, that's where I get a lot of my work done. - See, I see. All right, putting in work in the bedroom. It was actually born out of, like I believed I was going to be starting a fully-entern verticalized robotics company. We're going to build a robot. And as I left you, ClipTis, it was around the time that reusable rocketry, like, we're being able to relander a rocket that we launched was becoming incredibly reliable, almost perfunctory. It kind of wasn't even newsworthy anymore that we were able to do that, which was a soundly to me because every time I watched that fly up and land, it brings a tear to my eye. It's like, it's absolutely crazy that we get to be alive in this moment. This is, we've actually got from sci-fi to reality in our lifetimes, but they were continually seeing this cost curve come down. So I was infatuated by this idea that every time we have an age of exploration, I think about the East India Trading Company when we figured out joint-stop companies, and we were able to send boats across continents. Immediately after that you have industry, we begin farming, we begin mining, we do all of these things, and then we settle because we need people to co-locate with the industry. And so I was seeing things like Vada Space, Fleet Space, all of these companies that were spinning up, that were going to be this next wave of industrial revolution in space. And logically that followed, okay great, well that will mean we'll need settlement technologies in space, so someone's gonna have to be the space habitat business. That led me to, we're gonna need to be at a growth food in space, which led me to how do you grow food in a vacuum and a harsh environment, which led me to indoor agriculture or controlled environment ag, C-E-A, and indoor farming, vertical farming. That was around the time that plenty and these other companies that had raised like of a ciferous amount of money were then going out of business. - Which are vertical farming companies? (laughs) - Which are the other ones? - Which are vertical farming companies, they're going out of business. - Right, they're all vertical farming companies. And so it wasn't a great time necessarily to be like, I'm gonna go into vertical farming, but that's exactly what I decided I was coming to. (laughs) And so I sort of just tap the net work, I started asking people, who knows anyone that's worked in the space, who's been one of these engineers at one of these companies or founded one of these companies. And to all of the credit, like the VCs around Australia, sort of open doors and made introductions. And I rapidly was able to get in front of some of these experts and learn so much around the cost structure of those businesses and where it had gone wrong. And a lot around the very large footprint, meaning you needed a large amount of autonomy, a lot of robotics and automation, because the human labor component at that scale just made the economics not work, because they were growing lettuces and leafy greens, which had a very low dynamic range. People are not gonna pay $10 for a bag of spinach, because it's not noticeably that much better than a $2 bag. But there was one notable company in America called Oishi that was growing Japanese strawberries in New Jersey and selling them for a very premium price, because there was a story behind it, a brand, but also it dramatically tasted differently. So there's a lot around going for fruits that don't continue to ripen, so they benefit from being picked by the source of distribution. Growing crops that have that very high price point that they can achieve, they specifically grew leafy greens, because it's very easy to predict where it will fruit from, like it's gonna grow from, whereas with strawberries you can't really predict that. So trying to automate that is very difficult. So I think the main takeaway here is they went this path of technical least resistance, just get to market, but that meant that their unit economic hurdle was so, so difficult because they had to get the price point down, so low they had to make a huge footprint, and then because they built out that huge footprint, building automation at that scale was not ready, time and time. And so they just had a big leaky balance sheet, bleeding money, and the market turned on them. If the market hadn't turned, and they had one more big round, would they figure it out? Maybe it just felt like the culture was, we'll fix it in the next iteration, the next bigger factory, and they just kept not fixing it, 'cause it is quite technically challenging, especially at that large scale. So all of that aside, I think that led me to, well, the automation wasn't there, so I'm going to do this, I'm going to need to solve that automation problem. Well, I'm going to solve that automation problem, what does it look like? So I then went and spoke to 20 or 30 different robotics founders, and their lead robotics engineers, their lead machine learning engineers, and asked them questions about what it took to make their companies work. And there was a very common refrain between them, which was it all came down to the economics of buying a robot first as having a person to the job, and a lot of the time that came down to reliability, how out of 100 times it does the job, how many times does it do it right? Seems obvious. But that is the constant march of these companies is getting from 99%, but 99.99, 0.9999. If you're picking strawberries and it takes eight months to get to that level of maturity, and you crush one, you'll have a very unhappy farmer, right? The bar of expectation is incredibly high. And so I just kept asking why, and it got to this place where there are so many different use cases of robotics, and if I'm able to solve this feedback loop problem around gathering data from that edge of how that robot is doing, finding the places where it's not doing well to allow you to fix them, so you can make that continual march to 99.99, to 99%. 4, 4, 9, 5, 9, 6, 9, 6, 9, this is anyone who's a system reliability engineer is very common language. It's the same thing that people at Google and on Facebook do about making sure that the app is up, 99.99% of the time, and you're constantly trying to find these weird edge cases where it comes down that you're going to fix and protect against. And so this was all sort of very, I guess, analogous to me coming from web, having seen some of these patents before, even growth is that, right? You're looking for weird experiences on the website, on the app that cause churn, and you're trying to catch those in the analytics, and then you're doing an investigation, you're coming up with a theory, you're deploying the fix, and seeing if it improves it. Ultimately, I think I just came to realize that so much of businesses just feedback loops. And there is so much tooling and mature tooling in web and business to drive that and collect it and automate it. There isn't that level of tooling in robotics today. And that's what we set out to solve, realizing that given this problem exists, if I solve it, it's probably far better to be the horizontal provider giving them access to this platform to every other robotics company to accelerate versus become a fruit picking business off the back. So what, share an example of what that means in real life, what is maybe share an example of a robotics company, where this product can come to life? Imagine you have a robot that is a very self-contained, beautiful robot, it goes underwater, and it scrapes fouling and barnacles off of boats, cleans the bottoms of boats so they're more efficient. You may have a case where your autonomy makes an unexpected decision, maybe the light refracts through the water in a weird way that happens only every so often at a certain time of day. And it causes your models to make a different determination than they normally would. You might not catch that sometimes, because robots will produce upwards of one gigabyte per minute a lot of the time of data. And that data will be in so many different languages and shapes and sizes. It'll be images, it'll be a series of numbers, it'll be text logs saying, era, era warning, warning. You're trying to pay attention to all of that. And like when you think about something like a web app, it might produce a few kilobytes of data a minute, just in these logs or conversion metrics. So you're talking about many thousands of times, even sometimes millions of times more data coming out of these robotic devices than a web application. So none of the tooling that's built for web really translates over. So for that company who's dealing with that underwater cleaning, trying to isolate that issue is a lot of custom infrastructure they need to build, because it's just nothing I can take off the shelf really. The best solution in the market today is for them to there are great tools for replaying that data So they can scrub through it sort of play like a video and see all the images They can see the graphs being drawn they can see the logs coming through and so they can sort of put themselves in the robot shoes in a way And they're gathering context and they're trying to reason about what's driving this issue and For any of the AI engineers at home going that sounds like rag it is Right, it is retrieval augmented generation being done by an engineer looking across different data signals to come to a conclusion themselves And when I started to talk to all of these different companies and this was not like one isolated one company This was every robotics company I talked to the operator in the field that runs the robot will observe the fault They will tell the head office those engineers will hop on and they'll replay the mission replay the job And that is the way that it's done across the industry And that they may build some custom tooling and that will speed it up And as companies get mature like the most mature companies in this space have built a lot of tooling So it's not as always an unsolvable problem. It's just incredibly expensive and it doesn't make sense It's like a web company building its own GitHub building its own data bricks like why would you do that? It isn't what makes your beard taste better. It isn't your core IP and that is kind of that core inside of realizing this is going to exist And we should be the ones to build it How much time is spent In the status quo how much time will you save and Help us see like what does that mean for the business? Once the product is live. I think if you really boil it down Robotics is this fire hose of data Where you only really need one percent of it But knowing which one percent is incredibly challenging and time consuming And so a lot of the time when we do talk to certain companies and we're doing kind of discovery with them about their existing processes It will be upwards of sort of 90% of the data can't get looked at because it's just Think about The multiplicative relationship between the operator and the robots their goal is to have one operator Oversea as many robots as possible So if your main debugging solution is replaying in real time one second per second The data That's not going to scale when you have 10 robots with one person so it becomes the core bottleneck of the business It's actually the biggest unlock for all of these companies to be able to get the people of like one human Two 10 robots to 20 robots to 100 robots. That's what defines their unit economics over time And so we are in the core stream of improvement there. We're trying to accelerate their development And if you think about the time window that goes into Gathering processing curating training the models or finding and fixing hardware issues And then QAing that testing it and then deploying the new version About 80% of that is split between that curation analysis piece that we've talked about And then this QA and quality assurance happens at the end You make an improvement and you need to now go and test that robot to make sure you haven't regressed Anything else you haven't introduced any other bugs You have to feel incredibly confident once you're in production and you have 50 of these robots running around the world You're going to deploy over the air and you update to them You got to feel pretty confident that it's going to do what you expect And again, it's like there's limited tooling today for that kind of thing Part of it is that there just hasn't been an insanely mature Set of companies that have been commercialized to create a huge term for people to go oh, I should go and build this Right, and I think that's because most people spot opportunities the wrong way They're looking for where there's this big opportunity today To go and capture 1% of it 2% of it 3% of it But if you look at every single generational company that we have around us They all came up with the market that they have come to dominate It didn't exist if they if you know Amazon had gone Like what's the time of selling all of like to for AWS in the moment when they were originating They would never have been able to pursue it But it's that belief about the future that turns out to be contrarian and right That grows the market that you come up with you can reflexively accelerate And play a really key role in in helping to cold cold vet in nurture How do you earn the right to use the famous words that some listen as we'll understand To get into the door of these robotics companies where like the stakes are extremely high The security is extremely high I would imagine that the data is unique from one company to the next If companies are building custom tooling I would also imagine that people are using different tool sets How do you begin to sort of like a as I said get in the door of one and get a few design partners And then think about the product that is being created in a way that unifies all of those customer data sets Well we've observed actually from the cross section of companies that we've been talking to In different stages of development and commercialization and sort of becoming You know from research through two companies The ones who have been most receptive to this are the ones who are the most forward thinking The most accelerationist actually the furthest along Because I think they do appreciate what they do and don't do as a company Whereas when you're in the early sort of amorphous phases you're like okay well maybe we'll do that Maybe we'll pivot and go from making like slack like making a game to making a chat thing right Like people are concerned I think about the flexibility and losing flexibility by partnering with someone like us right But the ones who have had the best yield but best outcome working with us Who have been the most excited about it the ones who actually have robots out in the field They're actually seeing this problem in real time the sort of parallel data coming in For a single person or two people to deal with And so part of it is going to them and just just asking questions Right like we don't go in and go we know better than you We're there to learn from them realistically as you say like there's so much nuance and difference between the companies We want to understand how best to serve them And so that's been a lot of these early conversations And we have learned you know an almost amount about the best way for us to serve them It also has an interesting side effect as members of our team become experts Because of the horizontal cross section on nature Like we meet people in different sectors like we don't really mind which sector your robot is deployed into It could be we work with people in maritime in agriculture in logistics and defense These things all are at the data layer pretty similar We use a combination of the fact that a lot of robotics is organized around a few like Pareto distribution They're organized around a few open source frameworks Which output a similar data format So Those three kind of buckets of perception data numerical time series from sensors and text logs Actually covers a very large amount of the industry And then we do custom integrations for the edge cases that don't get covered I think you see that more and more with these AI First companies that are having these forward deployed models To do more of that custom integration at the beginning To help that company get the best out of the product It's just the new iteration of customer success And so we engage in that the same as them So one half of the question is why is robotics interesting The second half of the question is you mentioned earlier that a lot of Founders or folks have a hard time sort of underwriting markets that don't exist or are small Right now. How do you think about the two intersecting and feel free to handle that two Angled question anyway you want sure. I mean, I think the conventional wisdom is you can follow the cost curves Right, it's like historically throughout time as humans have brought down a cost curve We have disproportionately leveraged that thing exponentially more and more Right it comes down by half we use it 10x And that diffusion has held pretty true And if anything has accelerated as different technologies have stacked on top of each other Like the rate of which people came on to the internet or got their PC Compared to how the rate of which people got onto mobile To the rate at which now maybe people doubled in blockchain Maybe then we're talking about how quickly they got onto chat GPT with their 700 million weekly active Right like Fastest time to 100 million users These laws of diffusion like you look at any technological chart of time to 100 to a million doctors And it's just now their straight lines up right they launch in first year 100 million Things change quickly And so I think that's why it's important that you don't want to be sort of chasing the past Instead you want to be looking to the future Because things will move ahead of you and and you'll be left behind You obviously don't want to be far too early there are certain incumbents in the market already who have been around for five plus years They've made a bunch of Technology decisions five years ago based on a way the people approached robotics that maybe is become a bit less Relevant today maybe a bit more outdated And that's hard to pivot like I've been in tech companies that have been around for a while in the past And it's hard to just a fully rebuild your whole stack to be ready for the next technological shift so You know one of the pieces of insight that I got from a founder who had tried to make a robotics platform seven years ago It was like look at gen ai Like seems obvious everyone's looking at gen ai but specifically that was the one thing they didn't have when they tried to do it before And it's actually quite hard to use robotic data with llms because it's so multimodal because it's so heavy Like it's such big file sizes You could be easily looking at billions of tokens and the context windows of the most state of the art models Are measured in millions of tokens So for there to be a thousand x increase in the context window You're looking at it doubling every year for the next 10 years All right, so at least got 10 years Until the context window of a llm could technically be so large that you can just give it a Data file from a robot from a mission and ask it about it. Okay. Um, why does that matter? Because that's one of the ways in which you can speed up that curation process It's leveraging llms and generative ai to analyze mission files to summarize them quickly And we don't just say hey, he's a hey llm. He is a mission file summarize it We do a variety of traditional statistics on it. We have have a bunch of encoding models that we have built that allow us to turn each of these types of data into a vector representation, to then do similarity searches, query across them in natural language, because we do a bunch of contrastive learning. So there is a lot of fundamental machine learning to what we do that makes it possible for there to be sophisticated enough rag for you to build a context for an agent to even reason about the robotic data. So there is a lot of heavy lifting that goes in that it just doesn't make sense for every robotics company to figure out. That's why it makes so much more sense for their to be a platform that can partner with them to bring them that capability. And so that's why I think about that time horizon of, you know, we might, we're probably still 10 years out of the context window being big enough for them to just fling that bag in and talk to the LLM. These robotics companies that have a big problem, robotics is incredibly difficult, it's hard to make a robot that works. So we want to just take one of those hard things away. So specifically what we do is we allow you to search your data in natural language across your images, across the numerical time series, across the logs. And then you find those weird edge cases that you'd know are there but are hard to find. Then we allow you to find similar examples over time, all historical examples where this edge case has occurred before. So is this a severe thing that happens frequently or one time ever? Now you can get the answer. Then you're able to save and store that as is what we call a scenario. And if it ever happens again, you will be alerted. So now you have observability over this growing list of known issues. You're able to find those issues, save them. And beyond that, we also offer the summarisation which we talked about with the LLM component, where what we realize is that for a lot of people, they will go and do these tests or they have a growing list of tests and missions as they deploy these commercial robots. And it's pretty difficult to stay on top of how are things progressing over time, performance wise? What are the key events and issues and anomalies that you tend to see from those robots? Now we give you that summary right after the data is ingested pretty instantaneously. So we remove what was days of SQL queries and data transforms and just manual data parsing. And we give you that instant summary right after that you can just share with your colleagues or use this as source of truth and jump into the rest of that workflow with the search and the similarity from there. If we're talking about spotting the market on the move, it's cost curves coming down, it's a confluence of additional technologies coming on those cost curves but also just improving. So Moore's law has been the governing force historically. We see context windows expanding, we see the performance of these LLM's getting better, we see the side effect of this LLM revolution is that Nvidia has made better and better and cheaper and cheaper edge hardware. So the Jetson program is incredible. You can get phenomenal hardware for less than $1,000 that can run a lot of state of the art. It's still models but still really good. Something you wouldn't dream of five years ago. And that is doubling every year and getting more performant. And so the hardware is becoming more accessible. The cost curves of the components, LIDAR has come down, enormous amounts. I am using getting cheaper. It's like a lot of the core building blocks of robotics is getting more accessible and cheaper. Side effects of the LLM revolution as well has been there's novel approaches to robotics models that are powered by versions of these language models, which they call vision language models or vision language action models in the latest incarnation. And that is a special trained version of the LLM that takes in what the robot can see. Let's say some cups here. You give it instructions say poor water in the cup. Those are the two inputs. It then outputs cool. Well, if I'm going to do that in text outputs, I'm going to do that. I should pick up the cup. I should pick up the bottle. I should put the bottle over the cup. And then that goes into a diffusion model, similar to what stable diffusion or something like that. But that instead of generating an image, it generates an actuation state. It generates the joint positions that are more something should move to. And that thing gets output and fed back into the robot. And so it then says, okay, the next step is to move my hand towards the cup. The actuation state for that looks like this. I'm going to action it. Sorry. The diffusion model work with the action of what a joint should do. I think we're going to get quite. Maybe we won't. Yeah. But sorry, proceed. Essentially, it takes, hey, my arm is currently here. And the diffusion model is trained on a set of pairs to go, okay, well, it should probably move forward a little bit. And it maps that over to the numerical things that you tell the joints to tell them to move forward. And then it can reassess the situation and say, I'm closer. I need to keep going. Keep going. Keep going. And look at a lot of these VLA's in progress. They're very jolty because they're stopping to think every sort of movement. But recently, as of yesterday, every last week or something, Gemini Robotics 1.5 came out and it's able to have a much more smooth running VLA. There's a company I just met with the team in the US last week, who is called Generalist, who's similarly got a smooth running, is doing some more sophisticated planning before it acts. This thing is moving really fast. It's essentially the takeaway. I've never seen a technology move like this. Like I was sort of there when the sort of blockchain thing was happening. And people were getting sucked in. And I didn't see people adopting and applying the technology in such a profound way. I didn't see it feeding on itself like I do now. That's the thing that I think is fundamentally different. And so it's really just, can you track the gradient? Because maybe there's one company getting funded a week, two companies getting funded a week, ten companies getting funded a week, fifty companies getting it. That's been happening over the past six months. When I started talking to investors, maybe late last year, this year, this was not in the zeitgeist. This was not sort of a cool thing to be thinking about. There was a lot of focus obviously on Agentekella-Lem, and that was where a lot of the thesis was. But since then, there's been a massive sea change in the last six months. And now it's the place to be. One of the limiting factors, or perhaps differences between robotics and the 5, not 5, but the 12 months to a hundred million dollars of error are graphs in the software world that might be familiar with is that there is an audience to catch those products. And otherwise, another way to sort of frame it is like the distribution is there to be received by people who are willing to pay for it. How should I think about the distribution of robotics companies as they scale into our lives? And we actually start moving beyond what is actually pretty cool. But even just like a vacuum robot. Help us to imagine what we can, it's now specifically in my mind what the cost curves are looking like and how the technology is evolving rapidly. But then how do you take that and bring it into the real world? And yeah, I'd love to hear your thoughts. Part of it is that there needs to be a high enough rate of company formation. We just kind of touched on that the VC dollars are flowing in. That's a great sign. There's going to be enough swings at that to take these different lower cost inputs, some of the open source work that's happening and apply them and try and commercialize them. It will also bring a bigger mix of operators, which I think is also really important. A lot of core research has been done. That's really, really important, but it's not the same as commercializing to a company. And so being able to blend those two worlds together, I think is a key unlock. This year versus maybe five or ten years ago. More than ever, I don't know about you, but I feel like I see more brilliant people saying they want to work or are working on hard tech or hardware or things that are atoms not bits. Far more frequently than I did even a year ago, two years ago. That's a very positive sign for me that this is about the commercialization wave. Then there's also consumer adoption and consumer acceptance. As you alluded to, Rumble is pretty much the only consumer robotics company that's reached scale and survived. I found it by an awesome Australian, my robot. And Rodney Brook is obviously a legend and has founded a new robotics company, Robust AI, which is doing great work in logistics. I think they understood something very fundamental about the consumer robotics having a very tangible value proposition, having it be reliable. It does what it says is going to do on the box. It doesn't try to over bake the promise and flounder. They obviously did a lot of work in R&D to get it to where it needed to be. It wasn't as much of this broad base support that you can just leverage that we have today. But they were able to still pull that off. It has never been easier to now. Give it a crack. It was much, much harder than. There is so much opportunity in the enterprise and business space for the robotic side of things, dangerous jobs, things like going underwater cleaning rudders. That is something that is incredibly dangerous. There is a lot of heavy machinery stamping. That kind of thing where people lose limbs, they die. It is quite a hard environment. Those things will make a lot of sense quite quickly given the downside risk. But when it comes to consumer adoption, my Tesla just got a software update. They're now like drive itself. And when you get on a plane, you don't meet the pilot most of the time. You just get an audio thing saying, "Hey, I'm the pilot. I'm going to be flying this plane. I'm going to eat this 100 ton thing into space, essentially, and you're going to be on it." You're like, "Okay, sounds good." If we said that to our four-step grandparents 200 years ago, that was going to be a thing, 300 years ago, it would be like, "Absolutely. Fucking not." I do think it's going to be a generational shift where people getting born today will never learn to drive. It will seem archaic that we let people shoot around in these death machines. The time, huge amounts of momentum and impulse, you can just do that. You can just swerve the car. The amount of human loss that has happened and damage and hurt and pain has been caused by that. Obviously, it's been a huge unlock as a technology locomotion, but it came at a huge cost. And now we have, you know, some solutions to it. And I think that that when you have your first way in my ride and you go, "Oh, I'm never going back." Or like, "This is a huge unlock and I trust it and I feel comfortable." I think for a lot of people that's going to be this aha moment, that's that iPhone moment. And so it's already happened. And I think in retrospect, you know, the iPhone moment happened to everyone laughed at it. The literal iPhone moment, right? Like 2007, Steve Jobs announced it. Knock you, you know, everyone gets on the bell and goes, "This is stupid. No one's ever going to pay $600 for a phone with no buttons." That doesn't make any sense. So I think in retrospect, it will seem obvious. But it's happening right now. Like I say, last week, got the software update. So this is like real time. And I talked about like the Geminiar deep mind robotics lab dropped their update last week. Like new LLMs this week. This thing is evolving in real time. Like one of the biggest advantages we have at alloy is that we move incredibly quickly. And that we are able to keep up with that cutting edge. We're reading the research. We're utilizing the new technologies. And that helps us remain relevant and helpful to these other companies who are trying to focus on what they're doing. They don't want to be reading these research papers. I have them for all they are. But it's like we should be able to help them in that. If there was an improvement in the underlying LLM, and ultimately you're accelerating the insights that the team can work on to execute and make the product better. Like maybe you can share a bit about what some of the companies in the general foundational model business are doing like physical intelligence like general. Like how does your product fit into the next wave of general large language models in the physical world? Well, I think if we say that language models and large models are, they need to be trained. They need feedstock of data that's clean and labeled. For LLMs, there was the open web. It was this huge, huge corpus of exabytes of data that they could just pull on and train off of. Copyright aside, all of that doesn't exist for robotics. All of human society has not been feeding that data machine for the past couple of decades. And so there is an additional challenge that faces us now, which is how will those companies get to this critical massive data? And when they get that data, how will they manage it? We want to help with that second question. Is there anything else before we move on on the market that we should be talking about? I think just that it's bigger than people are thinking. Even if you're quite bullish on it, I think even then it is undistated. I'm not sort of here to say, to bang the drum and say, "Oh, you're going to have a humanoid in your house in one year. I'm not one of those people." I think we're actually massively overstating what that's going to look like in one to two years. We're probably underestimating what it looks like in 10 to 15 years. But I think similar to LLMs, if these are powered by VLAs, which looks like this main contenting technical architecture, but there will probably be additions to that and innovations to that same as we have seen with LLMs, we will see specialized use cases first. That came out GPT-2, GPT-3. There were lots of specialized fine-tuned versions of the model or something that was better at writing, something that was better at this or that. Over time, the models have gotten bigger and bigger and bigger to the point where the general model outperforms the specialist model on the specialist task. That's what we've just seen with the recent GROCH releases. Even today, there's a new Sonic model. Those are starting to outperform those. I think we'll see that similar trend of specialized use cases. We'll be the bread and butter. Those will get material traction. The autonomy is there and we're able to commercialise it in these industrial use cases and B2B use cases for the most part. Then I think we will see these general style models, but still they're kind of small, especially because they've to fit in that edge hardware, that smaller parallel computer that runs on the robot. That will be fine-tuned and prepared based on their specific data that they have from their specific use case. We'd love if they managed that data in our own. Then sometime after that, hard to say exactly when, these models will have enough critical mass of data that they will become so large, so many parameters, so much representation of reality that it actually can cross-embodied, it can be put into different robots with no additional fine-tuning and still be able to execute tasks with a high degree of reliability. I still think that that is a very long journey today. I think people expect that outcome and think of that as the robotics revolution, but I think it's more of a spectrum than that and we're already on it and that we will see an uptake like we're moving to a world where there's 10 million Tesla's on the road in the world already. If those progressively start to drive themselves, you will just be in a place where if you're out in the street, you probably are within 500 meters or a kilometer if you're in a CBD area, all of a robot. That's already true, or at least will be very true within the next one year of two years. That's before Waymos, that's before Robotax, you explicitly, that kind of thing. Then I deal every day with these robotics companies that are producing these more specialized, more verticalized use cases and they are seeing material success, commercial success, selling those use cases to other companies. In the backdrop of that context, how have you been searching for good customers? As I say, I think it's been much more about their level of commercialization that they have real customers and they have something where they're getting this feedback loop, like they are scaling their data throughput so materially, their current process is going to fall over. Their machine learning leaders like guys, we need to do something about this because if we continue this ramp, which is the curse of their own success, but you need to make your data an asset, not liability. Literally for a lot of companies, it can become a material liability because it's your cloud storage cost. When you move to global deployments, like if you've got robots in South America and robots in Europe, you're probably going to have to be using cloud to be collecting that telemetry. It's not going to live on some local infrastructure in your office. That's more of an R&D phase. We see this transition where people will need to move towards cloud over time, assuming it's not a sensitive use case. Those companies are the ones who are seeing that very high throughput. High throughput of data, they have their own customers. That customer base is growing and happy. Those are the people who have been the most excited to work with us and the people who we've had the greatest success with because pragmatic, real impact, we're able to actually show them commercial value of working with us, which makes them happy. Zoom above to the process or framework that you use to discover those good customers. We're going to have a lot of discovery conversations. A lot of those conversations is about asking those sorts of questions about where are they and their R&D. Do they have customers? What is their kind of general data throughput? What is the shape of that data? Who manages the data today? When it comes back, is there someone in the team? Do you have a whole team of people? Do you have an outsource partner? Understanding the shape of how they think about that today is the most important part of deciding whether it's the right fit for us to work together. We do turn away partners because I think for an early stage company it's really important that you have a good understanding of who you work with and who you don't. I know that you thought a lot about team design at Yug. This is maybe the first endeavor where you have, I guess, the time and resources and belief to design that team from scratch around the mission of a lawy. How are you thinking about the team and a lawy and in particular how people work together? What's the DNA? Who works though? Who doesn't? We raised this excellent round, led by Blackbird. The point of that is to build a fantastic team and build a fantastic product that makes customers ecstatic. That is the phase that we are in. What we've really focused on, and it's been a very rigorous process like we have interviewed a lot of people and selected very few. The reason for that is because when we're building this kernel of the culture, it's the first team, right? The most important thing that I can get right right now is that the team is missionary. They care deeply about the mission that we are on. They are hungry. They have something to prove and they're going to come and join alloy to prove it and they are coachable. They want to learn. They're not coming in thinking they're arrogant and how will the answers because this space evolves so quickly that if you think you have answers from one year ago, it's already outdated. You have to be agile. You have to be humble. That is the only way that you're going to survive. We've just got this team of curious, high potential, just super high horse power people. We have a lot of fun, but we work pretty hard. It's because we just believe that this is the only deflationary lever that we feel we have as a society to get somewhat out of this cost of living crisis that we have. Tinkering with interest rates will help to stem the bleeding, but it is not going to reverse the issue. Technology has always been that lever in terms of bringing the cost of things down. In a trite example, the cost of plasma TVs, and that didn't come down on interest rates. I came down on technology and food is similar. It's all supply and demand. It's interesting you just made the late from like tame design to like. Claw societal issues. Yeah. But that's the thing that keeps me up at night. When I think about what world are we leaving behind, that's going to affect the next generation, the generation after. If we do not make these changes, what does it continue to look like? We keep measuring the price of which things go up. The salaries don't go up. Something has to give. This is a way in which we're able to pass on those savings down through competition because we're able to deliver services and goods at a lower price. If people buy them at a lower price more affordably. The other end of that equation, I think this is a key part of robotics, is that it's only zero sum if GDP does not expand. If the number of companies does not exist, expand. But every time through human history when we've been able to make something cheaper or easier to do, we have just done so much more of it. And so there will be 10 or 100 times more companies being started because it's so much easier. I think that how many more companies got started after Wix and Versel and all of these enabling platforms and technologies came out that made it Shopify. Only more companies use this because of Shopify. We haven't even begun to scratch the surface. I think it was Ray Kurzweil said in this century, we're not going to have 100 years of progress. We're going to have 20,000 years of progress. And that's just relative to history, right? Like industrial revolution, all of that pre- that is kind of this flat line. And then it's just like boom. And it's because the technology is feeding on itself. And so I think that all of this together can be that deflationary force to bring more abundance back to people. So we're not leaving a world that's worse than it was when we started. In a world where everything gets automated, who's still working. I mean, I think that the short answer of what happens to Jobs, if everything is automated, is that everything that we think of today, the set of that finite things might get automated, but there are a bunch of things outside of that set that we probably can't imagine right now. I might not get automated, it'll probably won't be automated. Or it'll be this constant march of chasing, you know, as we are in society today, automating things, inventing new things, automating things. But I think Jobs will probably look like more like playing Starcraft, or some kind of real time strategy game of coordinating things, more strategic level, than they will be processing information. You think about how much time of human history in the last three, four hundred years, what percentage of human time has been spent encoding and decoding information? It is almost all of it. I think that will go away. And it will be replaced with a richer level of creative thinking, force of even more complex problems with much higher leverage. And that's what will be enabled by that lower leverage work being able to be taken off and executed. You should an example in history before. Yeah, I mean, well, there's many. When we made steam engines more efficient with coal, and they use less coal per trip, we ended up building a ridiculous amount more steam engines because it was so much more cost effective and we can get so much further with the same amount of coal that we ended up using more coal and aggregate. And it's kind of a weird example because I guess in this case, humans are the coal. But if we're using less humans to achieve each job, we're going to use more humans in aggregate because we're just going to do far more of those jobs. And if you just think it's similar to how everyone keeps saying, oh, we're with like LLMs, we're just going to go to a two day work week. I don't really think that's probably going to happen. Every time we've got productive productivity gains, we've still had people working the same 40 hour weeks. This is just going to be the next iteration of that. And I think there'll still be plenty of not more jobs created by this revolution than taken away. They'll look a bit different. But hopefully a lot more of that economics will be able to benefit from given how big the opportunity is for robotic applications. What's the analogy of a lawyer in that historical example? Where should we give us fast forward a year and fast forward 20 years? We are the sweaty coal boy in the front of the steam engine, shoveling the coal into the furnace. And I think for a lot of these companies, we really just want to help them on their journey to getting to that level of reliability that they need. And they are doing fantastic work and we just want to be a part of it. Love you, man. Thank you for coming back on. It's been an honor. Always. Thank you so much for joining us for another episode of Wild Hearts. If you want to learn more from other ambitious people building, designing and creating the world that we all want to live in, then please hit the subscribe and follow button. It would mean the world to us, the founders, the operators and the investors who join us in Wild Hearts. This podcast is a labor of love from the Blackbird team and day one. The show is produced by Kimela Herring and Melia Reiner, a Blackbird. Our marketing genius is Eva Telemacus and our editors are from day one. Annie Jones and Sanjay Chabarilla. Thank you all so much for listening and we'll see you next week.