Robots & AI in 2024: What Does the Future Hold? - Joseph Krause & Brannon Jones @AlleyCorp (E62)

Speaker A: Foreign.

Speaker B: So I'm back again with Joseph and Brandon of alicorp Robotics. How are we doing today guys?

Speaker C: Doing great. Thanks for having us.

Speaker A: Fantastic. Thanks for having us. Appreciate it.

Speaker B: Yeah, great to talk today. Let's just start with a quick two minute rundown, then for each of you, what you're about, who you are. Let's go. Let's start with Brandon.

Speaker C: Awesome. Yeah, thanks for having me on the show. Uh, Brandon Jones. I'm an investor at alicorp and I cover robotics and smart hardware. So that'll be pure play robotics. That'll be autonomous vehicles, underlying hardware, software, IoT, anything in the ecosystem, even ancillary services that support, uh, robotics companies. Those would all be things that we look at, uh, from an investment standpoint. Um, so yeah, my background, mechanical and aerospace engineering. Did that at SpaceX and built a bunch of factories there and uh, decided to be doing this on this side now.

Speaker A: Fantastic. And I'm Joseph. Um, I actually sit on a different team here at alicorp, the diversified team, but with the areas that I look into, Brandon and I have the pleasure of collaborating on a lot of different projects and think about a lot of the same areas. So I'm Joseph. I lead our computational and hard science practice here at alicorp. A lot of focus on semiconductors and advanced compute. And that's where Brandon and I spend a lot of time, uh, innovating and thinking about. I also spend some time in energy and uh, material science as well and not relevant to the conversation today, although always happen to digest. And so that's where I'm looking at here at alicorp, come by way of academia. So I was working on a PhD in material science from Rice University and there built a lot of neuromorphic chips, uh, using advanced materials and so trying to get rid of the bottleneck and operate chips more like the brain. And then I got selected for a fellowship and went to the US Army Research Lab which is right outside of Washington D.C. in uh, Adelphi, Maryland. Very fundamental in nature. We are mostly PhDs working on science, but the underlying objective of, you know, the army and the government is paying for work. And we want to look at how do we take things from early stage foundational research into applicable commercial products. And so I spent a lot of time on two core things, scalability and integration of that research that I, that I focus on. Uh, and then, and then been at alicorp for about two years. So excited to be here. And thanks again James for the invite.

Speaker B: Awesome. Many times. Thanks a lot for the rundown. I want to Start then with something I heard Elon Museum, I think a few weeks ago in regards to his Tesla robot, he said get the technology working and then make it efficient. I want to hear your guys opinion on that because from an outsider I feel like a lot of these hardware technologies actually when you're building you have to be very conscious of efficiency and efficiency is baked in. But he is in fact saying that we should race at the finish line, so to speak, and then get efficient afterwards. Do you guys agree?

Speaker C: You know, I think broadly I would say that I agree um, with that statement. I think that you need to push to understand what are the capabilities. And I think part of the problem with a humanoid in particular is that you want generalized functionality across a number of domains. And if you want to get there, you need to be able to build those data sets that you can pair with a foundational model or something like that to allow the robot to actually understand how to operate and a generalized environment and pull in context from the outside. So from one standpoint you do need to get it operational because you need to start generating the data. You need to also start showing the market that something is happening. You do have some traction now on the efficiency point. I don't think that anyone will adopt to the robot until you hit um, a high amount of reliability or efficiency from a cost perspective or a power perspective. Those other kinds of things are necessary for adoption. But I think in order to get there, um, I agree with that approach.

Speaker A: Yeah, well said. I would say the same applies when I think about semiconductors, especially when you look at a market like robotics, if we want to apply there, and also other places as well. While the early technology development is incredibly important, over time the efficiency must also come down. And I think you have to balance between R and D development, novel, uh, tech development, and then also building the organization. Of course, what we, what we invest in are companies that do make money and do grow to a large scale. That piece has to happen. And usually in hard sciences and, or things like robotics at scale, that efficiency does become better. And so, but one thing I think to mention is that the metric must come there at later stage. Especially when I think about computing in robotics, right. We talk about something like latency, um, where you have an advanced, maybe it's an AI acceleration chip or something like that. There really needs to be a stepwise change in performance for that robotic company or that robotic platform to want to adopt your processor over another one. Right. And that's just semiconductors. But it scales out across most heart Technologies as well, where if you don't have that stepwise change, you're not actually driving enough, uh, fundamental value to that end customer or client, in this case a robotics company, or in a robotics company's case, their own customer that they're serving. And so I think you have to thread the needle there, which is almost an awful answer to that question, but very true in that you need to push R and D, you need to do technological development, you need to build your team, but in that long run, where you're really driving value to your customer, you do need to have that efficiency metric, or any other metric that could be relevant performance, um, come around full scale and meet the expectation that you set out to build in the first place.

Speaker B: No, that's really interesting. And then kind of building on that point, then I feel like with the kind of all of these progress in generative AI recently, it's almost like the spotlight is taken off. Efficiency, right? It's like, how can we make this model as big as possible? How many parameters do they have? And almost fetishizing over, like, is it a trillion? Is it how many? Where does that fit in then? Because it seems like the future is going to be driven forward by software. How does hardware keep up? And when I say hardware, I don't necessarily just mean the chip stacks. It could be, you know, the actuators that need to keep up, or the physical robots operating in warehouses, et cetera. How does hardware keep up with this huge explosion in progress in software?

Speaker C: Yeah, so a lot of interesting points there, but I do think when you talk about these massive foundational models with all these parameters and the need to just say, hey, let's scale, let's scale, let's scale. As a way to get to generalized functionality. There are arguments for and against. Although, um, I guess I do think that they are a critical component. I mean, on the one hand, those models have actually worked quite well with computer vision and natural language processing. Now I think that's because there's so much of that data just available on the Internet. Um, just when you look at all the number of photos, all the text, all of those kinds of things, you can build a massive model. I focus on robotics. So talking about, from that perspective, there aren't those kinds of data sets already. Um, it would be kind of weird for someone to kind of have uploaded a whole bunch of signal inputs and robotic outputs too, for, you know, whatever their actuator is or whatever their, the form factor of their robot is. And it's a bunch of different form factors, and you're controlling the individual torques across a robot to various levels. At every second you're sending a different signal ultimately to get a complete action. So that data set just doesn't exist. So, yeah, we can talk about scale up, scale up, scale up. And you're going to make a robot functional, but not, uh, if it doesn't have anything to base itself off of it. Now we're seeing some people combine the two in an interesting way. Um, RT2, for example, transformer model, vision language, action model, uh, came out fairly recently, or Google announced it fairly recently, and they basically took RT1, which was a whole bunch of training set data that they took them like a year and a half to develop of a robot that only spent time, and I think it was a kitchen moving things around. Then they combined that with, um, some of the more foundational models and then you're able to get, okay, a little bit generalized understanding of what does it mean if I say clean that up, that kind of idiomatic expression, um, how do you transfer that to the tasks that it already knows how to do in a kitchen? And so they are starting to see some progress there, but still it's coming year and a half to even get the data set that they need to be able to combine it with that foundational model. So, yeah, people are talking about scale up, scale up, but it does work in some contexts. But I still think that there's a long way to go in robotics. Um, that just makes it kind of impractical. Just the different embodiments, the variety of environments, the variety of form factors, um, and then also simply the energy intensity and capital intensity, time intensity it takes to build those models up, I think are going to put these further out. Not to mention you can't ask GPT to do everything for you, even if, I mean today, for example, I put in a data set and I said, hey, return these answers from this data set. And it kind of made up some answers, it got some of it right, it kind of made up the rest. So if you don't have that reliability, you're not going to be able to put a robot into an environment where there's people, where it's a dynamic situation, where it's kind of guessing the outcome. And so until you can get to that 99.99% reliability, or put a human in the loop with a 70% accurate, um, algorithm and have the human kind of adjust, it's going to be harder to get those things deployed. So that's a Little bit of my soapbox. But I do think there's a slightly too much focus on scaling and probably too many narratives saying the future is here right now, um, for robotics, I do think to your point, the hardware has to catch up, and not so much the performance of the hardware, but the data we have on how to control that hardware. And there are people working on it.

Speaker B: That's fantastic. And then do you think that the future of robotics is just going to be some kind of walled garden where we can really control the inputs, um, and have kind of very specialist, very niche applications as opposed to these kind of, you know, what people are dreaming of right now, which is, you know, humanoid robots can do all manner of things, um, which, you know, the more you research and you read into and you see what people are doing in the space does feel like decades away. So do you think the near term we should focus much more on these kind of narrow use cases where you can really control the inputs into these, uh, models that we're training?

Speaker C: I think in some cases, yeah, I think if you want to specialize. So there's kind of this argument where if you want to do what RT2 did, stay with a very specialized use case, build a complete but small data set for that, combine it with a large data set, you get some functionality out. That is one way to do it. And I think that can work. There are people doing teleoperation of robots, um, as a method to get more of that data across various industries. And that's sort of a human in the loop version for robotics, because otherwise it's kind of hard to pair a, uh, robot and a person in the same place without the person potentially being in danger. So that's another way where you can start to get out there and get some generalization. But I do broadly think in the near term you're going to want to pair classical methods of motion planning and robotic control with kind of learned control from robotics. So that's another sort of human in the loop, for example. But stitch together the pieces that do work, um, and that probably will start in unlimited context now. I don't think it's 10 years away. And I do think those models will get good enough to have fully learned behavior in a generalized context. For a robot, for most contexts. Um, but yeah, near to medium term, I think you're going to have to combine, uh, some summer or specialized quite a bit.

Speaker B: Yeah, far away, Joseph.

Speaker A: Well said, Brand. I think, I think the argument for efficiency depends on, back to your original question, how you define as, well I think robotics is a great example when brand just laid it out very well around you can't afford to have, uh, lower levels of performance or kind of accuracy, uh, right. In your end model. And so if you define efficiency as the cost to build those models and then bring that data in and compute that, well, then, yeah, that cost is going to be incredibly high and not efficient. But on the flip side, the end use, end application use of that in robotics will actually not be efficient at all and that it's not usable, as Brandon correctly said, in my opinion. And so I think there's this unique dichotomy between the need to continue to build these models, um, and then once they're well understood and really the core training parameters can be used around the different markets of what they're trained for. That's when you're actually able to focus on that scalability and efficiency. And I think we see a lot of that with models like lambda or 7 billion or other type of, uh, size parameter models. You're seeing that become more of a reality now. We're seeing a lot of edge deployments of models where things can run on your, uh, cell phone or smartphone device. Very different model, right. Than what's running in ChatGPT4 or something else like that. But again, depending on how you define efficiency is either a positive or a negative. And so I think it's very early in this cycle and that continues to need to proliferate. I think the second piece around that feeds into that is very much the data that we're using to train them as well. If you're building a email or large language model for text or photos as well, there is a large amount of data, of course, via the Internet that's able to train that and in a precise way. Right. I think the way you should actually think about language models personally, um, is that we don't teach rules to models, right? And we don't teach them in the English language, that there's a subject and a predicate and an adjective and where they fit in the sentence. It rather knows that in context from the information or data that it's pulling in, right. From scraping the web and Wikipedia and whatever other sources that they may use to train those models. So when you scale it out to a problem like robotics, well, how does it learn those rules, right? How does it know what to do with the robot or the actuator, the sensor, whatever piece of the robotic system that you want to dive into, and there are many to dive into, how does it learn those things without seeing those examples before. And I think if you were to correlate that back up, is there a Wikipedia size version of robotic examples on sensors or actuators or the way robotic arms should move and operate with humans in the room? There probably is not a large data set there. And so I think these models, as Brandon, I again agree with Seth, they'll continue to develop, we'll need to continue to put more data into them and they'll really continue to evolve to hopefully move that accuracy number up and in the long tail, that efficiency number down, uh, or I guess higher, more efficient. Um, but again I think that takes time. I don't think that happens overnight and certainly doesn't happen in complex industries like robotics or other places where that can exist. The only other thing I would add to that thought would be I do think there will be advancements in semiconductors that enable the change here from a strictly power perspective. AI acceleration is a very large market. Alicorp is invested in that market to be fully transparent. Um, but as chips get better and are able to process specifically large language model, um, data sets, right, rather machine learning, et cetera, that will actually enable power to come down and efficiency will rise from that. Now implementation will be really important. Where does that implement it and who feels the end effects of that? Is it the OpenAI or is it the end customer that's using OpenAI? I think that's yet to be determined. But semiconductors will play an important part there, I think in reducing power and therefore increasing efficiency.

Speaker C: You know, I actually love you said that, um, especially the edge compute piece. In a lot of ways I feel like robotics relies on the work Joseph's doing in semiconductors and on that industry to come down. I mean, when you think about edge compute, all of these robots are going to most likely need to operate on the edge or in these localized contained network environments, these 5G environments. But you can kind of, if you thought the brain as a biological computer, kind of a poor analogy. But if you did think about it that way, the brain uses 12 watts or something and your computer uses 60, 100 watts or something like that. You almost can think of yourself as right, Edge compute. When you go into a new area, that's of course GPS denied, right? You can go anywhere and you can still make these intelligent decisions. So Joseph's question, I think that's so fascinating is what efficiency level is really going to be required to get generalization across industries? I don't know that we're going to get the power down to, you know, millions of years of evolution. But, um, it does, it does beg the question, you know, where's the threshold where you're going to start to get runaway adoption of these robots in various contexts? It's a great point.

Speaker B: Yeah, it's almost philosophical. I love the point you brought up about, you know, are we machines? You know, is our brain modeled around neurons and this and that? It's a very interesting space to explore. Um, but before we do, Joseph, please do touch on, um, kind of the progress that's been made over the last couple of years in terms of the hardware side of things. Because I think most people would have been exposed to the progress that Nvidia's made with the GPUs given their presentations, I think at the end of last year and earlier this year. Is there more to it than that? Kind of like, what were the big takeaways from the progress?

Speaker A: Great question. I think there's been a unique development in. If we're talking about hardware, we'll focus on chips in the Nvidia piece, but probably scales out to other pieces of the stack as well, um, around the right application, or rather could say the right architecture for that application. Right. And we've heard a huge push in application specific integrated circuits, or ASICs as the market likes to call them at Ali Corp. And between us, uh, as a firm, we very much believe that will continue to be a trend. Um, Automotive doesn't need the same chip as Defense, which may not need the same chip as aerospace or Space, for example. There will be differentiations there and I think chips are moving very much towards that development. That's been true in AI acceleration as well. Right. There's a handful of companies that are building AI specific acceleration chips that kind of plug into the existing hardware or combine in the existing hardware, whether that's M, M2 or PCIe connection, et cetera. That allowed the acceleration piece to become solved in a more efficient energy kind uh, of conscious way. Um, with more power, you know, it's able to compute more operations, or tops per watt is the common metric that's used in that space. So that's been fascinating from an actual architecture perspective. There's a very unique situation unfolding in that Nvidia has a large piece of market share for two reasons. One, their GPU is very good. Absolutely. It has a very high tops per watt and is able to do great calculations. But they also built Cuda, um, which is the development framework they use and have given to the developer community, frankly, over the last 15 or 10 to 15 years, um, to build that out. And Jensen, I was listening to a podcast with the CEO of Nvidia the other day and he very much was intentional about that. When they first came out with their first GPU around going to need developers that are using these algorithms and other tools. And of course Nvidia started in gaming and it evolved to AI. It wasn't always there. But the core principles actually do remain around the development of the infrastructure needed to use that architecture. And so people ask me all the time, you know, why is Nvidia such a market leader? Well, one, they are very good at computation, but AMD has a very good, uh, gpu, but they don't have cuda. And so that's been a big piece in the industry. The reason I bring that up to your question is there's been a lot of attempts right now from both startups and from some of the bigger incumbents in the industry around trying to chip away at cuda, essentially trying to chip away at the moat that CUDA has built around Nvidia. And so whether that's using different frameworks that allow, you know, multiple language, language agnostic as they like to call it, or on the flip side, hardware agnostic, I can come in with multiple languages and that can get fanned out to multiple pieces of hardware and it doesn't impact what we're writing at the front end. Um, that's been a huge trend. I think we'll continue to see this decoupling or decentralization for lack of a better word, often video into developers that want to just do machine learning, uh, and deep learning and AI and frankly don't have too much of a preference on what runs down the back end. Right. There are even opportunities to segment different type of machine learning and deep learning operations from normal CPU operations. Right. And so when we're running your browser in the web or something that doesn't need a, uh, gpu, that's of course your CPU and other processes like that. Same with memory, you know, how does memory get constrained by the CPU versus the gpu? And where can you innovate there? That I think is where the chip wave is really moving to. And so architectures are always evolving. Um, I spent too long working on architectures and I think that will always be intellectually interesting, but rather these frameworks that are being built right now to access the different type of architectures. Right. Almost in a way what docker did with x86 versus other type of architects, when intel and AMD and ARM kind of added competing Designs around that. A similar framework to that I think will unfold in the AI ML space.

Speaker B: Interesting. And then maybe one quick question on the back of that. Is silicon here to stay?

Speaker A: That's a great question. Uh, and I think Brandon smiling because he knows something that we're cooking up internally looking. Silicon's an incredible material. Um, semiconductors as a whole are going nowhere. I often joke in a very jokingly manner that Silicon Valley is named Silicon Valley. Uh, and. But investors tend to shy away from semiconductors. There are unique opportunities. We think about expanding Moore's Law. Right. We're getting close to the fundamental limits of Moore's law now. 3D architectures have been a large component around, trying to move around that intel, amd, um, Micron, Applied Materials, other advanced materials and research firms have looked into 3D architecture. We actually have a current incubation ongoing inside Ali Corp, which is still in stealth and hasn't been released yet. But that's looking at new materials to replace silicon in the semiconductor supply chain. Um, and the point there is to drive actually performance and applications. As I mentioned, that application specific integrated circuit, very specific markets that need highly advanced and specialized chipsets. If we need to move past 3 nanometer and lower node sizes, what comes after that? We're building the materials at scale to do that with a new startup. So more to come on what comes after silicon in the semiconductor world?

Speaker B: Exciting times. Hopefully in a year's time we'll have some more developments and that we can come, hopefully come along and get the founder on and talk about it some more. M. That'd be fantastic. Then I want to switch tack again, kind of. We actually talked about this off camera before. Um, it's a kind of something that's been playing in my head a little bit as well, which is human optimization versus human augmentation when it comes to robotics. So I think when you say robots, industrial robots, a lot of people have in their mind, you know, this kind of arm in the factory building cars or this, uh, you know, robot across the floor of the Amazon warehouse. These feel like jobs that a human could do. But obviously a robot has come in and done it, you know, 100, a thousand times more efficiently. But I think it's still, you know, something that a human within our capabilities could achieve. Is there anything on the horizon, kind of thinking, you know, the next decade or next five years maybe, where we could really look to use cases that are going above and beyond what any one human could achieve? Um, Brannon. Yeah. Let's start with you.

Speaker C: Yeah. So I think absolutely is the short answer. Um, now we see so many of these robots where they replace directly what a person does, because that's the most understandable context for a robot to be used. So if you're a warehouse manager or you run a, uh, manufacturing facility, you've got all kinds of internal KPIs and heuristics to understand how well your operations are performing with human labor. If I can tell you that I'll just replace a person and replace it with a robot one to one, and they'll do that same bin moving or packing operation. It's very easy to build a business case around that and to say we're going to bet on this as a business. And so that's why I think you see warehousing being, uh, uh, honestly has always been, and still the test bed and the launchpad for robotics and the number of the most valuable, um, robotics companies are there. Symbolic locus and brightpick and dexterity, and the list goes on and on of the top robotics companies, all kind of in that warehousing space. Now, the newest crop of robotics companies, I think that they're still doing, um, tasks that people would do, but doing it much faster. For example, we invested in Renovate Robotics, or roofing company, is functionally what they do. But they've built a robot that can lay down 12 times as many shingles as a person can. And so ultimately they double or triple the productivity of the overall team. And so that's kind of moving into the hybrid zone where, okay, yeah, this is the task that a person used to do, but instead of doing it one for one, we're now doing it 12 for one. And so internally, this roofing company has going to have much higher internal margins than any kind of comparison. Um, that's a legacy roofing business. Now what you're getting at is, are there things that are currently just not possible to do at any level by a person that a robot's going to be able to do. And I think a lot of that is going to come from our ability to capture and analyze data intelligently and respond to that. So some of that will look like customization, uh, in agricultural context, being able to adjust the nutrients that a particular plant needs, maybe even down to the leaf level, adjust the lighting, both the spectrum of the lighting and the intensity and the location of the lighting based on observed characteristics and behavior of the plant, knowing when that plant is active versus when it's dormant, and being able to tailor, um, all of your, um, agricultural efforts to each individual plant to ensure not only more harvest, but harvest at the right time. So then you only need to harvest it in one shot. Those kinds of things we can't really do with a person right now. We don't exactly have the data to do right now. Um, so I think that's one area. I think you're seeing robots being able to, um, do a lot of innovative things in the medical space that a surgeon can't do, simply because the physical form of a surgeon's hand doesn't allow you to go in there and do a 360 degree ablation around, you know, a vein. And if you can have a robot come in there and do that, or travel up, um, you know, intravenously or even up your spinal column to do some procedure in your brain without having a very intrusive, um, operation. Those are kinds of areas where robotics will definitely outperform a person now. Areas that need to have high reliability. But there's such a clear use case because it's tough to even do things like a tracheal intubation, uh, in the medical space where you just put a tube down someone's throat without damaging them. And if you can use AI to better understand how to position that, do that more repeatably, you're not relying on a person with a joystick so much. Those kinds of areas I think robotics, uh, will do quite well.

Speaker A: That was really well said. Uh, I'm not sure I could add too much more there. I would think one thing I would echo would be the challenge. It is to really move from 1 to 1 to greater scale. Renovate's a fantastic example of the 12 to 1. But that problem really requires both expertise and also data aggregation and differentiation to really make that become a reality. I think the Renovate team is very focused on that and successfully doing that to date. One of the reasons I think our robotics, um, folks back then, but that I think the underlying idea that you really have to have this some type of scaling mechanism, right? Maybe it's not, you know, 12x, but some type of scaling mechanism allows more of an easier adoption curve. Uh, I'd love for Brandon's thoughts on this, but farming, I find another good example of this, right, where one to one is a tough comparison, although could be valid in some cases, especially where labor is challenging or impossible. But 1 to 12, if we use just the same metric as Renovate, that feels like it scales much more appropriately in that use case. And I think that that challenge is where companies today need to focus and I think futuristically or I should say rather just five to seven to ten years out, we could bat around what that looks like, uh, and the different metrics and or definitions of robotic integration looks like. But today I think it's very much one of the pieces is what is that ratio and how are you actually impacting market that you work in. In Renovate's case, they're, they're deeply impacting the market they work in. There's a huge problem there. It's a serious pain that their customers feel. And so the more that ratio increases, you know, 1 to 12, say it started 1 to 5 and 1 to 7 and 1 to 12, the more beneficial that system really becomes and in turn the more data and performance and experience that system gains, which in time could actually increase that ratio more. So, and so I think the, the bigger you can build this ratio just to expand on that point, uh, I find very fascinating in today's technology. Ten years down the line, I'm not sure maybe we will have humanoids walking around and just doing all our normal jobs. Brandon's the expert, I know in the world, in this space, and I defer to him on that one. But I think in today's world that that's where one of the focus points should be.

Speaker B: No. So it's an interesting one then. You use the word replacement a number of times. But then if we think about augmentation, I think is something we touched on earlier in terms of always having a human in the loop. I think a lot of kind of AI and robotics, um, kind of roboticist specialists who have been around for many decades and have kind of seen all of these hype cycles and boom bust cycles, kind of say the same thing, that you always ultimately need a human in there somewhere, whether that's supervision, whether that's kind of looking and checking the input data, um, what are the use cases for that. Then if we pull out some specific examples where it's not necessarily replacement, but you know, in the same way I might go on ChatGPT and it helps me write a short email or you know, helps me fill out blog posts, for example, in the world of robotics. What are some of those examples?

Speaker C: Yeah, absolutely. So I think you see it across a range of spectrums. Um, I think one that uh, is quite popular is the idea of an exoskeleton, so the augmenting your literal physical performance. And there are a number of startups working on this. But can you use this in a context, especially in DoD or for um, outdoor recreation and hiking? I would Love this. It's very easy to hike, you know, as high as Everest, um, kind of thing. Um, or you see it in the case where, um, aging patients need help just staying mobile around their house and being independent for longer, taking a lot of burden off of the healthcare system if they get hospitalized, those kinds of things with an exoskeleton. So that's one kind of example. I mean maybe a more tangible example would be even um, driver assist, um, even the simple technologies lane assist and rear view assist. And then you can imagine as you move up in levels of autonomy all the way to self driving, um, I still think of that as human augmentation. Even though maybe you're replacing the human from driving. Now what you're doing is reading, working, thinking, having a conversation while you're not driving. And so I see that as really freeing up the person who's still doing the task that they need to do, which is go from point A to point B, but you're now enabling them to turn that into very productive time when before all they could do was talk on the phone or listen to a podcast or something like that while they were driving. And even doing things like that increases mortality risk. So that to me is a much closer, um, area of human augmentation that I think would be real. And I mean I also, yeah, I think teleoperated systems too. I know there are a lot of teleoperated surgical devices. There are a lot of folks working on teleoperated ultrasound devices. Um, so especially in an area that doesn't have the same kind of resources to have a physician present, that could make a lot of sense. Um, when the alternative would be patients have to travel 50, 100 miles to get care. Um, now if you can have a surgeon or a physician from um, a city that's even far away perform the procedure, um, ultimately that's better for everybody.

Speaker B: Yeah, I think driverless cars you just mentioned, um, it's actually a really interesting space to kind of dig into because it's one of those that you always felt was on the cusp of happening for, you know, the last five, 10, 15 years. If you keep getting told the same thing over and over, can you guys, obviously, as insiders who, people who gonna understand the space a lot better than most, shed some insight onto why that's been the case? Is it kind of, you know, hubris on the part of automotive manufacturers? Is it kind of like us overestimating where our AI ah, models are? Um, kind of like give some insight into why this is the case. Uh, Joseph would love to hear your thoughts.

Speaker A: Yeah, so I think one of the things I love to talk about space is that it's the classic predicament of a chicken and an egg problem, right? In that you have two axes, um, that are crossing. And in my opinion, self driving cars get better when you remove human error, right? In theory, not necessarily in practice because we've never had the experiment to run yet. But in theory, if you reduce human error and reduce it all to math, there should be very minimal or at least less accidents that occur from that problem. And so the more self driving cars there are, the better the performance that those cars actually exhibit from reducing the human error that's involved. The problem is that we, the first variable in that equation is the more self driving cars there are. And humans don't want to do that because they don't trust that those systems are strong enough yet. There's a variety of reasons. One of the reasons is they don't trust that those systems are strong enough yet. And so it builds this natural chicken and egg problem. And so I think this is a challenge where maybe back to your previous question with human in the loop and, or other implementation or artificial augmentation could be interesting, um, potentially, you know, virtual designs or things like that to actually improve data. We're seeing a lot of work also in theoretical work, right? How do I simulate driving miles? And so can I simulate driving 100 million miles and actually train that AI or system in the robotic uh, end application, whether that's automotive or other, to accept that, inherit that, and then actually use that to try to circumvent this chicken and egg problem until we can reach a critical mass and actually, you know, topple over. And so I think that's the first piece, um, or one of the pieces that I always look into. Um, I think the second thing that's fascinating is that there's a really hard balance between what should exist, what people want to exist, and then what will inevitably exist. We'll never be able to answer that question because there's so many components and things involved there. But as investors we're always viewing that. Right? And to put it more simply than those three topics, what's the macro view, what's the technological development and where do those meet in the middle? And I think in self driving cars that's been an interesting thing where people have a difference of opinion on where that should meet in the middle. Right. And should all callers be self driving and we don't own personal cars anymore? Should Only certain transportation or different aspects of city moving be, be autonomous or should really uh, only certain actions be autonomous. Right. Lane changes, speed, parking and high incident areas. Right. But when you're driving on, you know, the freeway or highway, maybe you don't need as much latency power and raw training data to drive straight in between two marked lines. Right. Humans are fairly okay at that, I think. Maybe not so. So I think that, that I don't have the answer to that question at all and think it's one that can be debated for a long time. But that's an interesting meshing of intersection. I think we've seen in autonomous cars where you have technologists and people really pushing this forward and absolutely some reservation from some of the publics. Right. And, and some of that's well founded in of course horrific accidents that have happened and things like that. And some of that is also a push around that it does exist today. How do we remove that? By bringing in autonomy. And they'll always compete. And so I think that's where problem one I mentioned up front, around the chicken and egg, coupled with this macro and technological smashing, that's where they link together, is you really need to thread the needle between what do people want, what does the macro environment dictate, where is the technology today and what problem can you really solve? That's a unique ecosystem and it's hard to solve for. I think the ones that do will be very successful. And so I'm optimistic. I think self driving cars would be really cool. Um, I've been with quite a few bad drivers in my life. I'm a fantastic driver. Uh, I think that's uh, an interesting

Speaker B: piece to consider, I think, to your point. I think what people want is perfection, to be honest, because I think we're so used to technology just seeping into all aspects of our lives. And there is also the almost kind of moral question about if a machine car, completely autonomous, ends up killing one person, is that actually kind of worse than human error killing a person? I think a lot of people would argue yes, it is worse because it's almost like we have a responsibility to only put things out into the world and engineer things that can kind of resolve our inefficiencies, if you see what I'm coming up. Um, so I think yeah, people are just going to look for uh, perfection. And even if we manage to cut down this cooler number of road deaths per year, which I, you know, personally, I think that is the goal of autonomous vehicles is not so we can just sit in the car and kind of have some wine while, while traveling from A to Z. I think the point is, right, that we kind of completely reduce this silent killer, which is road accidents every year across the world. Um, and people will only accept it when we get that kind of number down to 99.999% of accidents, uh, depleted. But yeah, Brandon would love to hear your take as well.

Speaker C: Yeah, I honestly couldn't agree more. I think you guys hit two amazing points. Joseph, with the data set is not there and it won't be there until the car is out there. But they won't perform well enough until we have the data set. So no one is willing to adopt this. No one's willing to accept some level of accidents. To your point James, because of the perception that how can you hold a robot culpable for anything? Um, how can we put out something there that we know is going to harm people with some amount of occurrence with a person. I guess people probably feel more like there's always some chance that a person does not injure somebody else or they have the control to not if a robot's just going to have errors at random. Even if on aggregate there are fewer accidents than what people cause. Um, how can we consciously and good conscious put that out there? I think maybe the third piece is more on the technological side that slowed down adoption of autonomous vehicles and that's like, I mean a person makes decisions on the road purely using visual and that's been the Tesla approach is just to use RGB cameras and stereo all camera technology.

Speaker B: Right.

Speaker C: And try to get enough training data. Now they are, I mean maybe they've got a hundred thousand cars out there, driven millions of miles. They do lots of simulation as well and their autonomy is still not there yet. So you see Waymo and cruise stacking on LIDAR and radar and ground penetrating radar and infrared, all these different sensors. Um, one that's expensive, um, um. So Tesla had the right approach in terms of let's keep low cost hardware. The problem is that it's, it's simply not working. They don't have enough data. So then you look at ah, these others who are using expensive hardware stacks makes it harder to scale for 1 and then for 2. Sensor fusion at that point incredibly challenging from a data integration perspective. How do you take the LIDAR data, how do you take the radar data, how do you take those in all these different circumstances and combine them? And then in driving in particular there are so many edge cases. So I mean I think what scientists are most Worried about is not, is the car on the highway just going to go haywire, um, under normal conditions, but what happens when a person has, um, you know, another car painted on the back of their car, or the sky painted on the back of their car or something else, and the robot doesn't actually understand where it is? Um, those edge cases can have really, really catastrophic outcomes if they're not accounted for. And so the bar actually is higher. And then the technological burden is proving itself to be much more challenging than we thought. So your question, is it hubris? Isn't it always? I guess, is my response. And maybe some amount of that founder vision is required in order to make something transformational happen. But I think that's what's slowing us down. So we're not there yet. We obviously are moving in that right direction. I do think as these systems are out there, um, you'll start to get a little bit of resolution in the chicken and egg problem, and that can start to shift public perception. But to your original point, James, it feels like something that has to exist, right? It feels like we can't just continue to accept the, uh, number of people that die in car accidents each year. So, um, that's definitely the tension in my mind and why I think ultimately it will happen.

Speaker B: Do you think maybe it's not just a case of inserting technology into our lives in the same way that a smartphone was, or a laptop or a PC, which is relatively straightforward and kind of just put it on your desk and it's there? It's actually more of kind of a habit change and an infrastructure change around it, and the world will almost change around. It's another question that I've seen, um, from an interview on Lex Friedman the other day, um, basically saying that kind of robots aren't just going to appear in our world without the world changing around them and human activity changing around them? Um, I guess my first question on that point then would be, is that already happening? I mean, the human behavior is always changing, um, year by year, month by month. With the advent of, as I said, smartphones, with the advent of AI, now we can all have kind of basically a personal assistant at our fingertips, which would be unheard of even 10 years ago. Are there ways that you can see right now our lives are changing because of robotics, um, and may even continue to change in the next 10 years to the point where we're like, our lives today are unrecognizable.

Speaker A: I think the answer up front is absolutely, uh, and what I mean by that Is think about modern manufacturing, right? And what exists today versus call it, let's just call it 75 years ago or even less than that, maybe 50 years ago. Uh, it's incredible which robotics have really enabled and allowed us to do. And manufacturing is one small discipline of that. I think it absolutely scales out. I think what's exciting is what also can come next in the near term future. I'm biased, but think semiconductors play a big piece of that. I think software and better iteration on models play a big piece of that. Uh, Brandon had an incredible point, in my opinion on centrifugion, that's an incredibly challenging problem. And then by the way, it needs to be done in femtoseconds, milliseconds, right, to really have it. If you're talking about autonomous car approach, and this is where I think the distinction between technology and assistance, uh, versus replacement is very important. It's important to always realize that the brain is incredible, right? The computational power and energy efficiency of our brain is really remarkable. It's a remarkable, remarkable, remarkable thing. And that's very hard to do. And I think that in some environments people think they're going to replace that. Uh, and I'm not convinced yet we ever truly can. I'm not necessarily saying we never could. There are a lot of approaches on biologics and different type of computation. Like I mentioned, I worked on normal for computing, which is an attempt to mimic the brain in computation. Not replace, but mimic. And so maybe over time that does come. But there are a lot of, you know, secondary, uh, and tertiary functions that come with that incredible performance that I think require the dictation around what is true help. Right? And that's where I think robotics and manufacturing is a great example. The sheer volume of cars or planes or any type of item that we can manufacture today is directly correlated to robotics and we're able to do that. Um, and I think that's an interesting use case that really demonstrates the level of proficiency we've developed to date when compared to 50 years ago. Yet the amount of efficiency and our performance technologically we have yet to go compared to where we are today. And that's a really interesting dynamic. One of the reasons I think robotics is so interesting, um, is that it really continues to straddle this, this realm of unbelievable world class technological advancement and still massive technological advancement to come over the next 25 years or whatever that may be. And so that's a unique market. I don't, I don't think many others. There's a Few, but not many have that unique straddling of both sides of improvement.

Speaker C: I like that and I think those are some great answers. Um, maybe just a few other examples of how robotics is already influencing our lives, um, and will probably continue to. Obviously we beat self driving to death, but all the driver assist, we talked about that, that's one way that's already shaping how people think about driving. I mean, even when you think about cooking, all the smart appliances that we use, um, already are changing at the consumer level. How people are interacting with smart, uh, technology in the home and having like Google Nest or those kinds of things, um, in the home and having these hubs that control technologically all of their appliances. There's been a pretty dramatic uptick in consumer acceptance of technology in the home. Now there are some subtle ways that it's also changing behavior, or rather behavior changes pushing adoption of robotics that aren't as in your face as a Roomba driving around your house. And that's simply the fact that, uh, I enjoy two day delivery. I expect two day delivery now.

Speaker A: Great point.

Speaker C: Right. And many times, um, it's even faster than that here in New York City. And so much of this is driven off the back of Amazon's acquisition of, uh, Kibo, which used to be Distroviral, which is an AMR platform for warehouse management. And their ability to deliver exceptional service is because they're augmented in these ways. And I'm not so much aware of that. I'm just aware that my package shows up same way I go to Chipotle. I expect them to have guacamole. And, uh, earlier this year they announced development on a robot called autocado, which takes in hundreds of avocados and peels and cores them and mashes them and kind of gives you a whole bunch of guacamole, um, right away instead of doing that all by hand. And that's because I go in there and I say I want guacamole. And I don't care that it's the middle of winter or wherever I am. And so there's a lot of ways where we're getting more and more spoiled, uh, with our choices. And that's pushing people to find innovative solutions. I want my coffee faster. So there's a lot of technology coming out to grind beans, to mix drinks, to do all that with less intervention. Um, so those are, I think, some ways in which consumer behavior is, um, enabling and driving more companies to look at robotic solutions in a cool way.

Speaker B: If we get driverless cars right, what we want is quack on demand. I love it. That's right. It's been a minute since I had a Chipotle as well. Well, you got me craving. Um, I want to pull an offer

Speaker A: there, by the way. Uh, that was just throwing something. There's also a big market, if I may, and thinking about things we probably less think about, but are incredibly important in something like Defense. Right. And there's a fantastic book on robotics and defense called army of none, which was recommended to me and I've read and a lot of my thinking comes from that. But back to a lot of the points we've talked about today. One, around data, two, around the brain, three, around the human augmentation or in the loop piece of this, you know, puzzle. That is, I think, an area of robotics that we see a huge push in it. Maybe not just robotics, also just autonomy autom, generally speaking. And will be there. There's a lot of nation states developing that. And there's a real question around, what does that look like? Should a human ever be in the loop? Right. Yes or no. Is mathematics better at determining when they should strike a target than a human or not? It doesn't use emotion, but doesn't have emotion. I do not have the answer to that question, and I don't pretend to, but I think there are situations like that where we'll not realize developments and advancements and the implementation of robotic systems until we see them, but that are actively ongoing. And I think Defense is one market that I'm top of mind on because I think about it a lot. But there are plenty of others. I think, uh, the guacamole and other things like packaging are really well said by Brandon, where I think that'll be a really big iteration in robotics around things we will not see or will not touch firsthand. Maybe we'll know. Right? We know that Chipotle ran this commercial with guacamole, but we don't touch that firsthand. That'll really change the way we perceive, operate and do things. Just like if we use a micro example in defense will change the way we operate and do things in both global conflict and kind of future escalation and hopefully de. Escalation of conflict as well. So that think that's an interesting piece to tie in all the points there that. That occurred to me when Brandon was talking about the things we won't see. I think that that is a great example as well.

Speaker B: That's. I really like the way you said. You said a phrase in there, kind of like is basically can mathematics replace humans? Is what I heard, which I think is a really interesting way of framing the whole robotics problem. Um, and then I kind of throw that question back to you guys. Do you think it can? Do you see a pathway to where it can? Because I mean then we start getting into the whole conversation about AGI. Um, and another point that I wanted to make, as always ties in is kind of that famous quote from um, Elon said about, was it Larry at Google, uh, accusing him of being speciesist when he wanted to basically put humans before AI and almost in certain people's minds, kind of robots slowly approaching, you know, the rights and, and everything that makes us human. Um, do you see a path to that? Obviously we're not there now. And I'm, I'm not kind of like stoking the fire of AI doomerism or all of this stuff. But where do you, where do you see that headed? Because I think it's a really interesting question. Can mathematics ever fully replace basically the human experience?

Speaker C: You know, I don't spend a lot of time uh, worrying about that. Honestly I don't lose too much sleep over that. I don't think that that will happen now. I guess potentially what I'm more worried about is this more Wally esque future where we've automated away all the inconvenience out of our lives and those have some deleterious side effects that we didn't really expect and does things to our mortality. And so potentially, but it's hard to think that there's going to be such a strong use case that people would um, continue to iterate if it started to worsen the quality of their life, take away their livelihoods completely, all those kinds of things. Uh, that's kind of the first piece. And I guess that would only happen if you have a runaway Skynet AI situation, which I honestly I don't think that's even going to be possible, um, personally. And the reason is kind of back to that data collection piece, um, but also um, the fact that it's trained on models and that's full of data from the past. And so when you think about what a person does and the creativity of creating something new, even what's happening in a PhD lab, you couldn't really have GPT, write a, uh, paper, discover a novel science because it doesn't quite have the capability to do that. Now maybe you could say there's some generalization that you can train it to say I should combine these different things and run these tests. But still uh, that seems really far away. To be able to pull in enough outside unrelated context and fuse those things, um, to have a model with enough data to have enough generalization to do that seems pretty far away. So I don't imagine people are going to be inventing, um, use cases right now that are going to worsen or dilute the human experience right now. Um, that's, that's my personal take. I think we're far away from it.

Speaker B: Sorry to interject then, but kind of on your point there, what is, what is kind of creativity and what is inspiration from a human point of view? Right. Because at the end of the day, we're all sentient beings that just take in lots of different inputs throughout our lives, and then something happens, something magical happens up here, um, and creates new outputs. I mean, you know, all of the famous inventors and discoverers of the past, they had just, I would imagine, above average number of inputs from a kind of wide variety of sources. Um, obviously we will, maybe, you can argue, have different natural intelligence and intellect levels and being able to put the pieces together. But I feel that, you know, that is maybe not an insurmountable problem for a robot to solve. If we just model after, uh, as you say, the way humans ingest and then kind of spit out and regurgitate or reform data and information into new forms, um, if we kind of break it down into mathematical models, at least in my head, it seems like something that humans can achieve. Um, not saying that's obviously going to happen, and we're probably many, many, many years off that happening, but it's not in my mind, at least outside the realms of possibility.

Speaker C: Yeah, I mean, my immediate response to that is probably not outside the realm of possibility. It does seem like, I guess technologically you could imagine what would need to happen. But when you look at innovation, and you're right, so much of innovation and creativity in people is taking something old and making it new in a way that just fits the current context. But they are taking in data from many different diversified sources. So that one action is actually a very computationally intensive exercise. That's how you have all your tortured geniuses. This is my personal take on this. But then also what happens is you have one person who puts out the first piece of jazz music, and it doesn't actually look like what jazz became.

Speaker A: Right.

Speaker C: Because you have somebody else who's got their own AI model trained on their own experiences, who take that in and then spit out something else. And so when you think about, like the entire network of artists who are creating to kind of create a cultural movement to something that truly becomes innovative. It's not just one individual who's pushing it forward. There are so many people, so many famous jazz artists who we say, oh, he was the father of jazz. No. Miles Davis was the father of jazz. No. Louis Armstrong. Right. On and on and on. And if you were to try to create basically that computational complexity across an entire network of people that all have these data sets, and because the brain can do it, people can do it at such low power.

Speaker A: Right.

Speaker C: In so many different contexts. The problem starts to look a little bit more intractable to me when you look at how innovation has kind of built on itself over time now still could happen. I think that might mean m. Maybe it happens slower, um, if you had robots trying to create jazz music. But I guess to your point, I suppose it could happen. It just, um, seems like it'd be tough to get the full confluence of things that needs to happen at once. My opinion.

Speaker B: I see your point. Yeah. I mean, I think where I'm coming from personally is, you know, the personal shock. I go from seeing things like mid journey in 2021, 2022, and, you know, not even realizing these kind of things are possible. And now we basically take it for granted. And I'm sure you guys would agree. Anytime I see anything online, I basically the first question I ask myself if it's an image or a video is like, is this AI real? So this is already an example of how it's seeped into our lives completely with us registering. But yeah. Joseph, same question. Would love to get your take.

Speaker A: I think it's hard to answer mathematics replace, uh, the human condition. Uh, I don't know if I'm qualified or have an answer for that. I think the real thing I would apply today would be probably each win in certain regimes. Right. And what I mean by that is, if you think about supply chain or factory automation, a human cannot physically calculate 57,000 transactions or moving parts in milliseconds and then place the right boxes on the right, uh, you know, lines on to the right station, to the right truck, to the right, uh, dispatchment. Right. That of course is a mathematics problem that wins there. To use a similar example for the audience, just to keep it, you know, relevant to the conversation in the defense or, you know, war space. Maybe that's one where I don't know the answer there. Right. What about empathy? What about human condition? What about the actual decision that gets involved there? And thankfully I'm not making Those decisions. So that's absolutely not up to me. Would never actually pretend it would be up to me. But what I mean is that is it's a great example where we probably don't know, um, that answer. I think, to bring back a point I wrote earlier. The brain is incredible, and it's hard to imagine. You know, I don't think mathematics ever fully encompass the things that we can feel, uh, and do and comprehend and articulate, um, that do exist in the brain. Right. Uh, and not, not to sound, you know, too cheesy, but very true. Compassion, love, understanding even. Right. In mathematics, things are most of the time binary. Um, they're a, you know, a 0 or 1. And, and even in the new models, they're weighted across whether it's more towards a zero or more towards a one. Right. And that can change and fluctuate, but very much still needs to reach an end state. And that doesn't happen in the brain. Right. We can be in the middle, we can be torn on things. And so I think, I'm not a philosopher, I don't have answers on if we can replicate empathy with that mathematics, but I certainly do believe that those things are very, very hard to do if we ever have a chance of doing that. And so there are certain things where I think human and the brain and consciousness does win out, and then certain cases, like I mentioned, where mathematics went out. I agree with Brandon that I don't worry much about doomsday of AI and thinking through that. I actually don't think that will come. But I also don't, um, ignore or not recognize that there are incredible researchers who do think that's a possibility. Right. With some of the OpenAI drama and other things that have existed, whatever may happen there, there are certainly incredibly intellectual people who have worked on this technology, who understand this technology much more deeply than I even do from a software fundamental mathematics level and are concerned about it. And then there are again, on the flip side, some that are not. And so I don't think we can answer that question. I agree with Brandon strongly that I'm not worried about that world. I think we should continue to build technology and we will continue to do so in a way that's beneficial to humanity. But certainly there are incredibly intellectual people who think the opposite. And so that's probably a problem that is answered over, over time or maybe not, um, and something we'll see. But I'd say as a whole, I think mathematics are good for some, the human condition is good for other, and it will be almost near impossible in my opinion, to replicate fully the human brain, potentially 95% of it, maybe 99% of it. But I think that one shred of consciousness that brings love, empathy, compassion, understanding, those things are unquantifiable, um, to put it simply.

Speaker C: So the last thing I will say on that, though that did make me think is, you know, on one hand there's so little we actually understand about how your brain works, how it processes information, how it adapts over time, responds to external stimuli. Um, um, that's why we still have a whole bunch of neuroscience PhDs who are still studying, because we don't know a lot about the brain. At the same time, you look at some of these GPT4 models and uh, there's actually a lot of neural nets that we don't understand how they're making decisions and how they work. And you see different emergent properties like, oh, we didn't expect this to happen. There's not a ton of theory around it. So it is actually interesting to say we don't know actually what these neural nets can do. So when I'm sitting here saying, oh, it's only based on the past, we don't fully understand that, just like we still don't fully understand the brain. So it makes me think, okay, it'd be hard for us to build a brain, but it might be possible for us to build something that's more interesting than what we expect.

Speaker A: Right? Absolutely. That's a great point. I mean, Grok is a great example of that. Right? It has the ability to have humor to put simply. Right. And you can't teach humor per se in terms of, that's a personal experience. But again, in the same relation I made back to learning rules in the English language, you can observe situations that humor was present and identified or marked as humor. And so maybe that is a way that compassion or empathy is marked around. This was a moment of compassion. Can you exhibit that at least in a reflect, reflectory or reflection based method versus ones that are very, uh, proactive or spontaneous and so. Great point. It's uh, a good thing to think about. Um, we have, it appears Neuron has solved how to be funny with an AI. I don't know. We'll see.

Speaker B: So I think probably by not putting any limits on it. Right, because we've all seen that. I think it's one of the first thing you said today, Brandon, is some of the kind of cock ups that ChatGPT has done is just absolutely embarrassing and it's very easy to Expose kind of like the pre programmed barriers have been put in place, right, by just simply tweaking the questions. Um, and again goes back to the whole argument of having a human M input. I think just to put like a full stop in that whole discussion is there does still seem to be that disconnect between reality and, you know, what happens in the GPUs and what happens in the models and the weights and all of this stuff. Um, one could argue that even language itself, I mean this is a kind of, you know, the model that has caught the most headlines in terms of the transformers etc. Bears no resemblance to, you know, humans lived reality. I mean it's just something that we have come up with and evolve to use as a means of communication. Um, actually bears no, um, kind of comparison to what we experience. Um, it's something that we've created and then the extension of that would be how do you then teach a robot to interact with the real world, which is full of unpredictability, where something that is a lot more predictable, which is language and the inputs that we give it. So, um, yeah, whole fullness of co argument. We don't have three hours, but we should definitely, definitely get into that sometimes. But guys, it's been a pleasure. Before we wrap up, I want to just finish on one question that um, you guys can answer in turn. Is there a robotics company out there today that you think in the future will be worth 1 trillion in value? So a company doesn't have to be something that alicorp have invested in, can be a robotics company that you've got your eye on, that you know someone there, you really like the research they're doing that you see in the next 10, 20 years, this can be a real game changer.

Speaker C: That's a great question. Um, and frankly my initial thought is it's a shame if we're not invested in whatever company. Um, and if I saw that company, I certainly would do everything I could within my power to make sure I was in it. Um, all jokes aside, um, we see some people innovating in some interesting spaces and unfortunately not a good answer to your question because again, if I saw that company, we'd definitely be in them now. I think our portfolio will all be very valuable. But to externalize it outside of that, because everyone knows we love our own portfolio. I think actually just in the manufacturing space that might be where we see a big winner. Someone who can go into the existing ecosystem where system integrators are kind of operating and bring automation to These manufacturers, I think the people that can do that in a way that allow the 85% of the market that's not automated because they can't afford it, or don't uh, have the space for it, or the conventional solutions don't work because they're too dynamic. If someone can solve that automation issue, um, maybe disintermediating system integrators, maybe augmenting them, I think that will be enormously valuable because so many of the tailwinds today are pushing for that. Um, just from the labor challenges to the onshore challenges and all the enabling factors from connectivity and compute hardware costs going down, all the things that we talked about already, um, to me the company that gets that right, bringing automation to the mass market of manufacturers will be incredibly valuable. So a startup, uh, listening to this, if you're building that also please, please reach out because uh, I would love to chat.

Speaker B: Fantastic.

Speaker A: I'll be quick. I'll give two very quickly. One very practical I think and a little bit of a stretch and one a little bit more theoretical or futuristic on that timescale. I think the one. I mean trillion's a big number. Uh, right. There's only a few companies that really reach there. If we envision a world in 15 years, call it, or 10, whatever that metric may be, where it's like Star wars, for lack of a better example, things are flying around. We're not driving places like you walk out of the office, you get in this thing, it hovers to your next location. Everyone does that. Let's take the theory of car ownership as almost vanishing just for purpose of this example. Tesla to me really is the example or maybe another automotive company, but they certainly have the lead in compute and training that actually becomes valuable actually at a trillion so real economic level there and has the tools and systems in place and the physical hardware and cars to do that. And so not exactly the fair answer. I understand your question, but really in a realistic sense I do see them continuing to push this autonomous, you know, non ownership type. You know, you have four people and two benches sitting in the car, having a conversation, moving around. In Teslas, I could see that um, existing in the more futuristic case a company that I don't know if exists but something like an ex machina. If you've seen that movie, um, where you really have this human like incredibly intellectual at your fingertips personal assistant, you, uh, know, a mixing of Iron Man's Jarvis with a real humanoid version of that, um, and there's a few companies building there, maybe one of them is it. But that to me feels like something a majority of industries, people and really the world would want to ingest and have. And so there's probably a potential there, if I think about how many people have an iPhone or smartphone. And it's still of course not the full world yet and the valuations that come with that, that it could be a good form factor that a company that builds that there, and I mean really sophisticated, almost indistinguishable between a human and a robot. Um, that could be a, uh, definitely a trillion dollar company as well.

Speaker C: It reminds me of my feature answer matches that one. Yeah, it does.

Speaker B: I love it.

Speaker C: I'm not totally sure that the form factor will look like a person. It may not when it becomes that Jarvis assistant, but I 100% agree. Joseph. In the future that would be incredibly, incredibly nice.

Speaker B: I was going to say, to be honest, like, I know Brandon, you're a big fan of Humanoids. I just, at the moment I just find them a bit messy and I just think that something like a Jarvis, almost like a God that just kind of hangs around and can hear you at any times, plugged into all of your devices, your electronics, your wearable, understands your health state, etc. Etc. Understands your relationships with other people. I can see that being, you know, not even a trillion dollar company. I can see that being as commonplace as anything, you know, by the, by the end of the century, definitely, um, in the first world, in the developed country that we will just have this device that I think it's, I mean we already, you know, on the first couple of percentage points there, um, so I would, it doesn't take, it doesn't take uh, a genius or ah, someone to actually extrapolate out into the coming decades. That's the future I'm excited for. But thanks so much for your time guys. Absolutely been an absolute pleasure having you on. Um, all the best. Have a great rest of your day.

Speaker A: Thank you.

Speaker C: Thanks, James.

Speaker B: Take care of yourself. Bye Bye.