Michael Cho - Data, Robots & The Future of Humanoids - ep 180

Speaker A: I think I would have basically quit or given up if not for the fact that at those dark moments, I'll compare to that, uh, plane crash and say, okay, this is very bad, but then it's not as bad as dying.

Speaker B: This podcast is for informational and entertainment purposes only and is not investment, legal or financial advice. Opinions are those of the host and guests and do not reflect any affiliated entities. Investing involves risk and past performance is not indicative of future results. The host and guests may hold positions in discussed securities. Now, please enjoy the show. Michael, thank you so much for joining me today. Super excited to chat with you. All right, Michael, you are the CEO and co founder of frotobots Labs. Frotobots is a crypto powered robotics platform that uses low cost remote controlled robots to collect real world data for training AI models. You're also a third time founder and a, uh, not so fun fact. You're actually a plane crash survivor. All right, so before diving into all the fun and you know, wonderful world of robots, I'd love to, I'd love to hear about, you know, that experience. You know, obviously it's not a happy thing to talk about, but I mean it's, it's so rare and I know you've told me the story before, it's impacted you in a major way. Could you tell me like, how being a plane crash survivor has, has impacted your life?

Speaker A: Yeah, I mean, uh, well, first of all, thanks for having me on the pod. Um, um, yeah, I would say honestly on hindsight, the fact that obviously I survived, uh, it is actually probably one of the best things that have ever happened to me. Obviously it's not a fun experience, but it is a very, very good thing that have happened to me because, uh, it's a very good, I would say bullshit barometer, uh, for myself, right. I think before the plane crash, um, um, I think I lie to myself a lot, you know, like in what I think I'm doing, like what motivates me as a person, things like that. Like, I don't want to get in too philosophical immediately, but it is very much like that, right? Because I was literally seconds away from death. Like, you know, I was, um, so the plane crashing in the sea, um, and I, you know, it's a true miracle that I survived. I was drowning basically, right? And towards the end, you know, I was basically blacking out. Um, so I mean, I have a very long scar on my shoulder because of that, but obviously I survived, so. And um, yeah, I mean, I would say at that time I was actually, um, uh, head of investment for a family office in Abu Dhabi. You know, I was earning a, you know, drawing a decent pay, managing, um, you know, I would say billions of dollars, you know, on the balance sheet. So I have a lot of, you know, I, I, I think I was a very cocky guy, let's say, looking back. Um, but then, you know, when, when I was being so close to death, I realized that all these like, it's, it's all bs, right? You know, it doesn't really matter. Um, and uh, yeah, so the fact that I survived, I realized that life is, can be really, really short and you never quite know, um, uh, when you need to leave. Um, so yeah, and then I realized I just want to build things. I'm a very geeky person, nerdy person, so I just want to build. So I started doing startups. And honestly, like I said, uh, mentioned just now, this is my third time co founding a startup. Um, there are many times I think I would have basically quit or given up if not for the fact that at those dark moments, I'll compare to that plane, uh, crash and say, okay, this is very bad, but then it's not as bad as dying. So let's see. And I think in this case for photobots as well, um, when we first started like three plus years ago, like, robotics is a very dirty word in, uh, VC world. You know, I think the mantra back then for sure is you don't touch hardware, number one. Right? And robotics is the hardest of hardware, so all the more you shouldn't. Right? So, uh, I got to give kudos to those like, um, very early, like um, you know, investors that wrote us a bit of checks, right? Not um, big checks, but sufficient for us to get going. Um, uh, but of course, like the world has completely changed, right? Ah, since ChatGPT, a lot of the advancements are starting to creep into robotics and happy to go in depth into some of those. Um, but, um, yeah, I think that experience, in summary, set me up well in terms of the mindset of being a founder, because founder is a very lonely journey and it needs to be because you only have, um, I would say, um, asymmetrical upside if you are contrarian and. Right. So for some period of time you have to accept that the whole world disagree with you, which I think for the first two years of our project, certainly that's the case. I think now the world has completely flipped. I would say maybe too many people agree with my original thesis, so we need to run a lot faster. Um, I mean, it does help with fundraising obviously, but I mean, you know, so it's, it's hard to tell. Um, but anyway, I do think it helps me a lot in just ah, uh, being an entrepreneur, being a founder. And when it gets lonely, when it gets down, I just thought okay, it can go further down so I'm not at the bottom. And so yeah, maybe uh, unrealistically optimistic, but I think you really need that as a founder, um, especially in the early stages.

Speaker B: Yeah. I'm curious as to, can people experience, you know, without some traumatic events like what you experience, is it possible to have this uh, such a major impactful event on your life where you kind of reframe your whole life and decide to chase after very large goals, hence like you know, building, you know, in the robotics field before um, chatgpt and before kind of humanoids hit the scene in a major way. Um, I'm wondering if it's possible to experience these, these life altering events where you make these big changes and decisions without experiencing trauma. Or maybe trauma is required in order to, to, to, to, to, to. For that to happen.

Speaker A: Yeah, I mean you certainly don't need a plane crash for you to pick up, become a great founder. Like you know, look at all the great founders, you know. Um, but if you really look at a lot of their earlier lives, you know, they all have trauma of a different kind. Like mine is just dramatic, right? You know, I nearly died. That's it. Um, but trauma can be, you know, could be growing up, you know, like a lot of them tend to be immigrant because like uh, by default if you're immigrant, you're, you know, you're a fish out of water in a new place, new language. And that could be a different kind of trauma. I'm an immigrant myself. Like in Singapore. I grew up in Hong Kong. So um, so coming to Singapore was also quite tough. Like I didn't speak English like nearly at all I would say when I was a teenager, when I came. So it was a tough adjustment period. Um, um, but I uh, think you, I think a person can be smart, but smart is not useful in my view. To be a good founder, um, you just need grit. And sometimes you can be too smart for. Yeah, as a founder, right, you need to be quite naively optimistic. Like I say, I'm realistically optimistic and that is good in my view, a good trait. Um, but you cannot be too unrealistic at the same. So it's, it's, it's um, it's, it's, I think for every Founder is a different. And also depending on the, on the cycle, like, uh, and also the particular industry, the dynamic in the market, I think it's quite different. Um, you know, obviously there are many, many great founders who have very, very different personal traits and they all kind of make it. Uh, I certainly have not made it at all. Like, you know, I have failed, basically. My view. In my view I failed for many years. Right. I'm still trying. Um, and I'm still trying because like, you know, I had that experience. But definitely you don't need a brain crash, I don't think. Uh, but having some, um, you know, being able to just get back up your feet consistently is a must. Is a must.

Speaker B: Uh, so they see the difference between the master and the apprentices. The masters failed like 10,000 times while the apprentices failed once, you know, so. So I think you're right about failure.

Speaker A: Yeah, it's the same case as I was just chatting with my teammate just now, uh, before this call, maybe the way he used ChatGPT. I say you just ask one time, it's not going to be sufficient. You're going to keep on going back and forth, back and forth, uh, until chatgpt might fail you on the first prompt. But m, if you really give a chance, you eventually will surprise you. So in life, um, I don't know whether that's a good analogy, but, um, yeah, you cannot expect to get something great on your first try, uh, unless you are just lucky.

Speaker B: Um, so you mentioned ChatGPT and robotics, and when I think of ChatGPT, I think of AI as kind of a software type of product. But you mentioned, hey, we started pursuing photobots before ChatGPT and has the robots, how has the robots industry or field changed since ChatGPT? Because you would imagine from my standpoint, oh, it probably hasn't changed at all because we're talking about a purely software product, not like robotics.

Speaker A: Yeah, so I was, uh, fortunate enough to be, um. So I mean, I think like most people, I was still like completely wowed by ChatGPT and the advancements is there and the pace of advancement so fast. Um, but I was a firm believer in the trajectory. So, um, during my first startup I was actually in the Bay Area and I remember crashing, um, this course in Stanford, um, which was taught by this PhD student. His name is Andrej Karpathy. So he was still a PhD student. So I started crashing the course, um, and then that course really set me into this. So this is circular 2016 17. And I really caught the Bug, basically. So actually since then I've been reading, uh, AI, you know, like back then they call it machine learning. Okay, um, papers. So during the first, my second startup I actually do a lot of machine learning. And throughout the years I've been reading papers and so, um, even though I'm not researchers, I'm not super technical myself, but I do spend a lot of time reading papers, um, and trying to. So I would say I'm quite up to date. And what I observed, maybe three, four, like when we first started this project was that um, my hunch was that the bottleneck towards a lot of the, well, let's take the cyborg robot, uh, because that's what we started and that's what we're still doing. And the main thesis, actually why me and my brother started this from a hobby project to a company, is because my observation back then was that there are a bunch of cyborg robot startups, um, some of them in fact projects, uh, done by large companies like Amazon. Uh, Amazon used to have a program called Cyber Robot called Scout. Um, I think they put a lot of investment, maybe 400 people working at the team. Um, there are still great startups working on this problem. Uh, but my observation back then was that most of them tried to do the AI thing from day one, which means that they will have a robot that is pretty expensive, um, around $10,000 actually because, because they put it an expensive GPU, fancy sensors and all that. Right. And this is three, four years ago. Um, I guess back then maybe zero rate was still, I mean interest rate was still zero. So maybe it's somewhat doable, but clearly it's even harder to do it now. But uh, even back then I thought that if you take the view that these things, um, maybe the bottleneck is not so much hardware or lack of algorithm because back then I think some of these algorithm that has, since, you know, has, has proven a lot very effective, have started actually three, four years ago, started to show some signs of life offline, uh, reinforcement learning being one of them. So, and if you take that view of the world, which is very much by the way, a Tesla worldview, okay, like you try to train things end to end, then then actually you, this is a data collection problem, right? So having $10,000 robot, basically, if you look at those robots as data collection devices in the real world, $10,000 robot is a no go, right? How are you going to get the data? Um, and so we thought very naively at the time, um, because we started out honestly, this whole Project started as a hobby project. Um, me and my brother were just hacking, trying to build a, um, cyborg robot by ourselves. Just a maker, typical weekend project that we do together. Um, but we naively thought that, okay, someone should build a toy robot that is a dumb robot. Um, dumb meaning there's no GPU on it. So the idea though is that instead of trying to have the AI in day one, you yank out the AI, you just have it be 100% human. Remote control now three, four years ago, is still very much unproven, right? And first of all, no one has tried to do a $500 cyber robot right at the time. That's why we decided to do it. The other bigger thing is that can you even get a latency good enough, right? Because if it needs to be human control, then you need your latency to be low. And we took that bet because this happened during the pandemic where, as you recall, that's when the whole world jumped to zoom. And my observation was that there was just a lot of, uh, infrastructure level investment that's brought forward by literally a decade. Right. Obviously we have video calls before then, but that one year or so just fastly accelerate all this. So I felt a bit more confident that you can just ride that wave, um, such that you can be based in Chicago and then remotely control a robot halfway across the world, let's say in Africa, and somehow be able to do it with a level of latency that for some robotic tasks is still feasible. Um, and Sauro robot, by the way, I've got one here. So this is our current version. It's literally like a toy, right? Um, um, so like I mentioned, our initial target was to get this to $500. This bucket. Right now we retail for $200 and uh, it's pretty good spec. But anyway, um, so clearly it can be done, but at the time, no one has done it before and therefore we wanted to try. Um, but, um, I would say maybe at that time, uh, this was my third startup, right? Um, but it wasn't a crypto thing. Honestly, I, uh, was very much a crypto skeptic at the time. Um, I thought it was all scam until I would say only like the second month into it, when at that time it was only me and my brother doing full time. Um, uh, so we discovered helium. Um, and this is like, I think, uh, uh, I want to say April 2022, right? So this is like, um, yeah, when the market is still pretty hot, um, before the collapse. Um, um, but yeah, but helium really blew my mind on honestly. And I thought that it's fundamentally very different. Obviously the cool kids call it Deepin these days. Right. But um, obviously filecoin, things like that, you can argue those are all Deepin, but I didn't know any better. I don't know what's Filecoin. Um, I don't even know what's a crypto wallet and all that. But I was just blown away by helium and I thought, okay, I need to learn from them. Um, because we were trying to have $500 devices out there in the world, they also have $500 hotspots out there in the world. I look at their world map and say, hey, I want that world map. Um, and that's when we started really going to this, uh, I think took the red pill on crypto. Yeah.

Speaker B: So going back to the initial kind of thesis there, because you were saying you were taking this class and you realized that the issue was with, I guess robotics broadly was the data. So, so how, how could you. I guess what was telling you that? Because that to me is non intuitive. It's like, you know, the issue with robots is, I don't know, the, the actuators in the arms or you know, you know, who knows, right. I'm not like a expert here, but I would think it'd be something more on the hardware side. But you're saying. No, it's, it's the actual kind of the, the um, the data that's used to train the brains essentially. So how did you come to that conclusion?

Speaker A: Yeah, I mean I should say that definitely you need good hardware. Uh, this is the physical world after all. And physical world has no mercy. It's very brutal. Uh, having good hardware definitely is required if you want to replace a human being, let's say ultimately, um, it's just that the path to go there, um, I mean it strikes me at the time, um, so I don't know whether you're familiar with, especially in self driving, they have so called level one to level five. Um, I think back then, but even now, although some of them will claim otherwise, I would say all the cyber robot projects, effectively they are what I would consider level two, which means probably the AI can drive maybe out 100 meter, maybe 99 meters off the way. Maybe there's one meter somewhere that human need to intervene, let's say. Um, but because you don't know when that one meter is going to happen. So you ended up, you have humans watching over the bots kind of all the time. Maybe one person can watch over two or three robots, but there's a limit to how much you can scale because you do have to watch it quite carefully. And so if you work out the unit economics, I realized that level two is a very awkward place to be in terms of unit economics. You have a very expensive hardware but you're not offering a service that can fully replace a human which is so hard to do because human. A lot of people think that, let's say for example against Dawak robot which is mostly used for last mile delivery let's say. But if you really think about the doordash guy, yes, maybe 99.9% of what he does is doing the trip, but that last couple of meters he does on both ends is so general, uh, it's so hard, uh, all the way down to let's say whether he can press a lift, the button of a lift, right. Or maybe doing basic customer support, right? Ring the bell, have a chit, chat with the person or maybe you know, at the restaurant, pick up, oh, this is the wrong order. Like you know, that's it's um, um. And uh, I think if you work at Doordash and Uber is you see all this, um, um. But as I thought about all this I realized that like um, maybe the way to do, to get a level five, um, doing it, trying to do the AI prematurely, it's not going to get you there. Um, and since no one is trying to do the so called dumb um, robot version, I thought maybe you can do that. And um, you asked what gave me the conviction. I guess you can say that I buy into the wealth view of Tesla. So um, like I said, I took a class on the Andrew Karpathy, right. Obviously he was at Tesla for a while. Um, and I've been tracking what they do and it strikes me that what Tesla was doing is very much collect real world data in the real world, learn from it. It's more like they are learning from humans, right? So if you look at each Tesla, to me they are just like uh, robots on wheels basically. And effectively the data that they are gathering, they obviously gather the video, but actually more importantly they gather what the human driver is doing. And that is actually a very critical part. Not just the video, I think having video is required, but having what the human does is the real gem. But once you have these two things then I think you can set up a pipeline to literally learn from humans at scale. And I thought cyborg robot needs a similar kind of so called data flywheel um, so, and uh, I can totally understand though, like why these cyber robot startups do what they do because how do you, I mean you unlike a car, let's say, you know, people will buy cars because they drive in them. Right? A cyber robot, you know, no one can see the cyber robot, right. So how do you deliver value kind of in day one so that people will pay for something and obviously last month delivery. So I do understand the go to market. A lot of them end up still trying to tackle a very difficult problem, um, from day one, a lot of them are still trying and I really give them kudos of doing so. Um, in a way, um, what we ended up doing is this, I guess the gaming component, but honestly I also kind of stumble onto it. I don't want to say I have it from day one. Uh, we were also trying to actually do last mile delivery. To be honest, um, I have driven our robot into a Starbucks at least 100 times or more. So the way I'll do it is because these robots, uh, especially, I mean our current version, but our earliest version already have a speaker microphone. So what I'll do is let's say order something on doordash, um, uh, and do self pickup. I'll drive it into Starbucks and then I'll shout through the microphone and then usually, um, uh, the Starbucks girl will probably freak out for a bit. But eventually I get my coffee and so I did it maybe 100 times. We did a bunch of TikTok videos. Some of them actually went viral and all that. Um, but I realized out of that 100 attempts, maybe 10 of them failed. Right? So it sounds like, oh, we're nearly there, right? 90% we're nearly there. And I would say, yeah, I mean during those 90 time where it worked, I think it was magical, right? Because I can totally see how you can do, for example, um, some kind of labor cost arbitrage. Right? Because I said this thing can be done across the world, right? So, um, hiring a person in New York to deliver versus getting someone from, let's say the Philippines or Indonesia to do it for you is a drastic difference. Um, but then I look at those 10 times where it failed, I realized, um, if I broadly look at the way it failed, you literally need to solve level five, literally have an AI that's as good as human. And also the hardware needs to be way better than even though $10,000. Once I said pressing the lift, it's a different thing going up the escalator, ah, going up curbs, maybe going upstairs. Going downstairs, um, avoiding the dogs along the way, just all these hardware, so very, very good hardware and also very, very good AI And I realized those are almost nearly insurmountable. But then, so I was honestly a bit disappointed. I was like, oh gosh, maybe this is just too hard to solve. Maybe this is still not the time to jump into it. Um, fully. Uh, but then along the way I discovered that me and my brother and a couple of the guys on the team were all fighting with each other to drive the robot because it's kind of fun to drive it. Um, I don't know whether you know, gaming title like Euro Truck Simulator, um, I mean it's a pretty niche kind of gaming experience. But I'm a sucker for those gaming titles. So I totally get why some gamers, some genre of gamers would like this kind of like, uh, it's not a recent game but it's just a way to, I don't know, chill experience. Right. But except that this happening in the real world. Right. Um, and I would say since then we have really lean in very, very hard on the robotic gaming aspect. Um, to be honest, I think we have not fully proven it. Like we uh, have a lot of doctor evidence I would say of where this can go, uh, as an entertainment and fun gaming experience. Um, um, um. But uh, I do think if you can figure out gaming literally is a very nice wedge to allow these robots to generate some revenue even though they don't have real world utility, but they're just funny to look at and play with.

Speaker B: Can you take me back to the initial thought process? When you're first starting frotobots, you're thinking, okay, sidewalk robots are going to be huge for delivery. Therefore we want to create uh, you know, the, the, the top or you know the, the leading company in that field. And we're going to start with, with this model and we know by making these cheapish or not, not cheap, but less expensive, uh, you know, sidewalk robots which will scale through crypto incentives and et cetera, et cetera. But was that initial vision like we want to take on delivery or we're purely going to collect the data and then we're going to somehow monetize that?

Speaker A: Yeah, I mean initially definitely I wanted to do delivery. Um, because if you can really do delivery, then there's cash flow. Right. And then you can support the growing of a fleet and then once you have to grow the fleet then you collect data. Right. Um, it's just that I think are definitely very strong teams still going after this thesis, right? And I give them a lot of credit for still trying. But you can imagine that is a three party like marketplace. Uh, running that three party marketplace is non trivial. And who's your competition? Your competition are human beings who are extremely versatile in all kinds of scenarios. Right. So I think uh, it's really um, Actually the question is it's not so much like the robots cost $10,000. Even if it's $20,000 if you can amortize over a number of years it's still way cheaper than hiring the dude running around in San Francisco or New York. Right. I think the bottom line is that these robots are still nowhere as good as doing the full thing end to end as a human retail door dasher, um, and being 99% dev versus 99.9999% like Six Sigma. Right. I think it's uh, and oftentimes what I realize for a lot of robotic tasks it doesn't have to just have to be last mile delivery. It could be folding clothes, it could be if you ask a robot to cook a meal for you, things like that. Right. Um, some people think that because of safety reason for self driving you need Six Sigma. M. Right. But I would argue actually for the more mundane tasks that you maybe eventually want a humanoid to do, let's say for you at home, let's say cook you a meal. Um, Imagine if the way I think about it is um. Imagine if you think about typically the way they talk um, about success rate in robotic research paper these days is they call it episode. Right? One episode could be or one task could be let's say cracking eight. That could be one task. Now if you think about the entire um know series of frying is a series of many many subtasks. If each of them give you, I don't know like you are only 99%, you're 99% to you know, if it's a series of 100 tasks, 99% to the power of 100, you realize you, you're less than 50% right. So you really need it to be nearly 100% for every, every single sequence. Um and so, and I, so I do think um, I'm a slightly more pessimistic maybe. I, I, I, I know too much now these days because I'm so deep in, in the woods and talk to researchers all the time. I, I, I know a lot of the bottlenecks that still exist. Um, um. But gaming it's, it's a good wedge like I said, you know, because a 99% robot is still not good enough to replace uh, the DoorDasher. But 99% robot, um, or 99% like kind. It makes for a great game because it takes a lot of effort to succeed. Right. If you fail, it's just you fail the mission. That's it. Right. Um, and I think that could be another very interesting wedge um, to get the data flywheel going as well.

Speaker B: So could you take me through the evolution of kind of your journey with Frotobots, where it went from? We want to do delivery. Okay. It's actually extremely difficult. We need to uh, get more uh, data. So let's do uh, more data, more experience. We'll do teleoperation, um, and then you know with, with humans driving it's some portion or, or the majority. And then uh, actually we're going to gamify this because you know it'll increase engagement and let you know it'll be easier to get these teleoperators and, and we can reward them with different incentives. Is it was that kind of the evolution where like went from. We're going to do the delivery to okay, we actually need to do teleoperation with people and we're going to first collect the data then to okay, we're going to do all those things but in order to get there, or in order to get there maybe the most efficiently we're going to gamify. It was that kind of the three part evolution we can say.

Speaker A: Yeah, pretty much. It uh, I would say also actually um, um, this gamification part I think also leans in very well with uh, just trying to get a robotic, I mean a robot body that is like less than 500 bucks. Obviously we have to make a ton of hardware trade off. Um, and because we pick it as a gaming right the requirement actually for the robot is like it can be a toy, right. Can be very plasticky. Right. So I think it just fits in very well to all the things that we wanted to do anyway. Um, so we just really double down on the gaming side. Um, um, like I said, I don't think this is a fully validated thesis by us. Um, but we have a number of anecdotal evidence also. For example, we do have gamers who um, are very much normie, I guess you could call them Web2Gamers, um, who don't know this has anything to do with uh, crypto, but then would have paid us hundreds of US dollars to play the game. We have gamers like that, like a handful of them and occasionally we have some Pretty big uh, twitch streamers that would just all of a sudden show up and that just live stream the whole game. And so uh, to me there's just so much like um, is a very, esports, very in fact augmented reality if you think of those kind of gaming angle. I think there's just a lot of things that we haven't fully actually explored yet, uh, that can lend quite well um, to robotic gaming as a small niche genre. Um, and so I mean my sense on robotic gaming is that if you kind of look at again like a gaming title like Euro Truck Simulator, right, It's a ten year old game. I believe his daily active is like 20,000 to 30,000 um, gamers. It's not huge, it's definitely not all those huge million uh, concurrent games. Um, it will never be a Fortnite. Right. Um, but it has such a loyal folks group uh, that would just play in uh, year after year. Um, so as a game maybe it doesn't look like a crazy gaming title, but then as a robotic network, um, the largest sitewide robot network to my knowledge, even uh, though they've been working on it for I think nearly 10 years is like 3,4000 robots. So if you have a fleet that can support tens of thousands, you are by default the largest robotic fleet of its kind. Right. Cyber robot. So it's um, it's uh. Yeah, so I thought it actually if you look at the order of magnitude it is possible.

Speaker B: Can you talk about your fleet today and how people are interacting with the robots, the sidewalk robots now?

Speaker A: Yeah, so I would say we went through, I would say broadly maybe two phases of development. The first phase is um, I mean, I would say even till today we don't have a token. Right. So we don't have the actual Deepin thing going. Um, so, so what we had to do actually even to now is that we still have to put all these robots on our own balance sheet and we have very little capital, um, until more recently. So we have to be really careful. But because the robots are so cheap, right. So it actually allow us to deploy these bots in. So we have deployed robots in over 50 cities actually, which is something that's never, I don't think anyone has done it before, um, most of these other sidewalk robot projects there will be maybe in maybe 10 cities and that would be considered a lot. Right. Um, in our case it's just toy. Right. It's something that, especially this version is something that you can live up by your hand. Um, so that has allowed us to Actually very quickly deploy them in the real world in many different cities, um, in very, very different environments. So, uh, we have bots in Europe. Uh, I mean I would say one of the most memorable drive that I've done was we have a couple of bots in Sweden. And this happened during the winter time. I was in Singapore, it was super hot. I was driving, you know, probably to the, to the, you know, the Swedes who basically saw this little bugger, you know, driving around is like, you know, what is this? Right. It's probably a very mundane like regular, ah, neighborhood. But to me it was just magical, right? You know, driving through snow. Um, so yeah, these things have been through snow, have been through really heavy rain and surprisingly they hold up quite well, um, even through summertime in the Middle East. We have some bots in Abu Dhabi as well. I was quite surprised they do hold up. Um, um, but having this kind of really cheap hardware allow us to scale geographically even though they are all in a way sitting on our own balance sheet. So the way we've been doing it, like I said in the first phase is we hire local host. Um, and these are usually students actually. Oftentimes, uh, we pay them an hour a day, something like that. And to put it out they charge robot every day, put it out there and then they'll do their own thing. They literally leave the bots there. And turns out, um, it really depends on the neighborhood where you put it in. But there are many places, uh, in the world where you can actually do this rather safely. Um, so there are some locations we've been running for like two, three years, no problem. Um, and somehow, meaning the robots will still come back at the same place, although in the intervening hours actually maybe some gamers would have driven it around and completed some mission in the neighborhood. Right. So, um, I would say that set up where effectively you as a gamer you can dial in and assess any random robots on our network. Right. Um, it's definitely have its own like, appeal because it's, it's a little bit like virtual touring. Like I said, I was driving aboard in Sweden. I've never been in Sweden before myself. Right. So that was very interesting. Um, but I do think that this next phase we're going into which we're focusing a lot more on, um, is, is, is a different kind of game mode where we are basically getting the uh, gamers to drive the bot that they own. Right. Instead of driving someone else. Bottom, um, the reason why we're doubling down on this particular mode for now is because you can imagine if everything is on our balance sheet because I don't have the whole Deepin thing yet going. So I only have hundreds of bots. And you can imagine even if you are a, ah, gamer who really like this kind of gameplay, but after a while, um, you realize it's all the same bots, especially in your own time zone, because only the bots in your own time zone will start to show up. Uh, uh, and my own gut feel is that we need to be at least in the thousands before this game start to feel like an open world game that you can never finish playing. Because in a way every robot to me represents roughly, let's say one hour to two hours of unique gameplay. You can think of it almost like that. Um, I think at our scale we are not able to give people that experience yet. Um, but I think I would say once we have the token going and we truly have the DPN thing going, we will definitely switch back to bot one because that is maybe potentially more fun for some people. The fact that you can maybe, I don't know, stick some token with some dude and then as collateral you drive that bot around, pick up some nft, which is the basic gameplay. It's very Pokemon Go kind of gameplay. Um, and so, um, and then along the way you experience the world. Right, um, that way. Um, but I think for now, especially because the robots are really cheap now, um, so I do think it's quite affordable for most folks to just buy and maybe just drive around in their own neighborhood. They kind of like. Exactly. Pokemon Go. People play Pokemon Go in your own city, collecting items in your own city. And I think that alone, uh, will already get us somewhere, especially once we have the crypto incentives.

Speaker B: So, so today for example, I would be driving around and there are missions for me to accomplish and there's things for me to do or the gamification aspects with the, I guess the crypto incentives are not yet in, but the other still little things for me to collect in, little missions for me to accomplish. And then that's like from, from my standpoint as the gamer, but from your standpoint as a company, what, what is the purpose that these bots are doing for you guys? Are they collecting really valuable data that you can use for like, for monetization?

Speaker A: Yeah, great question. So uh, like I said, it was our initial thesis that the data is useful, but honestly it was a thesis, but I would say, uh, it has been proven, um, um, uh, from very, very early on because we care about the data so we have set up the whole data pipeline. So from very early on we started collecting data. And then I think once we have about a few hundred hours of data collected in the real world, um, which back then I thought it wasn't a lot, but anyway, I just tried reaching out to some researchers and uh, I found that the reception was very good actually from them. So it turns out that actually the kind of data that we gather is extremely rare. Um, and when I say extremely rare, I want to emphasize again, if you're only talking about video data, um, that is not rare at all. There are a ton of this kind of first person view dash cam video on YouTube. Right? Uh, um, the thing that is rare is that if the video is timestamped very precisely against human actions, because that is what you need to kind of set up the training pipeline to learn from human. And because ours is a game, we record basically whatever the gamer is doing on his game controller, basically. Uh, and because we have the video and we have the action labels, then this data set becomes very unique and extremely rare. Anyway, um, fast forward to today. We have actually a lot of ongoing research, uh, collaboration, um, with various um, researchers who do this kind of, they call it urban navigation research. So uh, for example, a couple of weeks ago there was this lab, um, from UC Berkeley. They publish a paper using our data set, um, and also they ah, built basically, uh, more or less kind of like state of the art kind of navigation model. And then more importantly they evaluated that model using our fleet of robots. So I want to say what I started to realize is while our uh, robots is out there collecting real world data, that exact fleet of robot also allows the researcher to evaluate their models in the real world. So that is actually this evaluation part, um, people doing AI development work in the digital world would never think about it because benchmarking is easy, evaluation is easy. You let the AI do an Olympic math exam in a second. So it's never a bottleneck towards iteration cycle, um, for digital AI, but in the real world you have to test it. In the real world you can't speed up time, right? So you can test it in simulation all year on. But ultimately the proof is you need to do it in the physical world. Um, I started to realize actually some of the researchers do care a lot about the fact that we give them data. But actually some of them don't care as much. Especially for those who rely a bit more on the more classical robotic techniques, like less learning based, less data driven kind of techniques. But I would say all of them care a lot that we have data in the real world for them to test. And so anyway, yeah, so that's how we started working with a lot of these researchers. Um, and along the way actually we got to know um, some of the folks from, especially from DeepMind, um, robotics. Um, I'm very good friends with a bunch of them. So uh, I think for the second time now we've actually co organized this competition called the Earthworfer Challenge where it's a very unique competition where we basically have human gamers. So we have the top human gamers in our game representing humanity if you will, and then competing against university teams who are ah, um, researchers who do urban navigation research. They build the AI models that basically uh, take on the same game exactly as the human gamers. And um, happy to report that at least for now the humans are far ahead of the AI. Uh, for now, for now. And so um, but yeah, but I would say um, so I think this kind of collaboration like the fact that um, I think we like I say it's more than 10 universities where we had this kind of collapse and these are some of the best labs in the world doing this kind of research. And so I think our initial thesis that the data is useful and then also kind of like uh, towards the end more recently we realized that the, the being able to offer them physical robot for evaluation is also extremely useful to them. Like these two things uh, just allow us to really do a lot of very close um, collaboration with them. Um, so I mean right now we don't monetize the data set, we uh, choose to um, freely give away. I think that's the right thing at least at this stage to do because that allows us to just basically spread away and get a lot more mindshare like from this top uh, researcher to work with us. Yeah.

Speaker B: Why is this data, why is urban navigation data valuable? Is it you know, because the future is going to be robots giving us all these deliveries and we need all, you know, need to know the nitty gritty details of every neighborhood and kind of the sidewalks and everything like that. Or uh, why do people find this valuable?

Speaker A: Yeah, I mean ultimately I would say that probably the biggest economic unlock would be if you can do last mile delivery. Right? Um, um, you know it could be food or grocery. Right. The doordash, the uber eats. I mean already huge category in and of themselves. Uh, if you look at logistics, typically the last mile cost of like even cross border delivery logistics, actually the last mile is easily 30 to 40%. So actually the last one is extremely costly. Um, and so, so, yeah, I mean, if you can really build an AI that it can basically do what a door test should do, or like, you know, what the Amazon delivery guy can do, then yes, you know, it's a huge, uh, economic, um, unlock, slash, capture, perhaps. M. So I think that's where, uh, how you would fully, fully, um, capture value and lock economically. Um, I also want to say, like, actually, obviously I'm a bit biased because I started off doing urban navigation, but I do think actually, uh, sidewalk navigation is probably one of the hardest robotic tasks to solve. Maybe, um, harder than what I call robotic manipulation, which is like, let's say robot arms folding clothes. And in my view, probably also harder than self driving on the road. Um, the way I think about it is, for example, Andrew, when you drive, do you think you pay more attention to driving on a highway or driving in a, uh, crowded car park?

Speaker B: Car park, for sure.

Speaker A: Car park, right. So that tells you that actually the complexity of the task is less related to the speed, is a lot more related to how dynamic is the scene. And so if you take that view, right, uh, when you drive on the road, at least there's traffic rules. Most of the time people do follow. All the cars move in the same direction in general. Uh, whereas in cyber environment the edge case is the norm because there's no rules. Right. So, uh, you have all kinds of weird shit that. I mean, one of the most weird shit I've seen one of our robot was, I think we had this bot in Berlin. And um, one of the guys was calling me, hey, you know, I don't know what's happening to our robot. And then we started looking at the camera. It turns out there was this, uh, I think a stray dog that just kept humping our robot for like 10 straight minutes. It was so odd to watch. But anyway, like, you know, shit like this, right? And so, um, and therefore the more the edge cases there are, the harder it is to gather a data set that represent enough of those edge cases. Right? Um, um, but yeah, anyway. But I guess you can also argue though, maybe unlike self driving on the road, if you fail, no one dies, presumably. So maybe the requirement is not like six, nine. Maybe you just need four, nine. Um, so it is possible that way as well. Um, yeah, but I would say whoever can really correct last, um, I mean, urban navigation. Um, yeah, basically you can be doordash, Uber eats, and the last mile for Amazon. Um, and also if you imagine not just delivery, I Mean it might be the other way actually. Like when you finally have humanoids, that is everyone will have a humanoid at their home. Your own robotic butler. You're going to get that robots to help you do stuff, run errands. Right. I mean that robot still had to navigate in the neighborhood. Right. If you wanted to do more things than just stuck at home. Right. And so that robot will require navigation in urban environment. Yeah.

Speaker B: This Six Sigma that you keep mentioning, is this some sort of like holy grail of uh, I guess safety measuring where it's like it can't be 99% safe or good or effective, it has to be 99.99999, you know, et cetera. Is that like the, considered the Holy Grail or. Why do you keep mentioning Six Sigma?

Speaker A: Yeah, I think it's uh, Six Sigma is like, it's, it's, it's like kind of like uh. I think the Japanese use it a lot. Uh, um, um, you know they, they use it for like to, to, to get really the qa especially the, the Honda, the Toyota. Right. They really treat it like, like, like that Bible. So uh, I mean I just use it as a hand wavy way to say. So in my view. Why I think, for example, I think we're still some ways away from. There's a lot of Talk recently about ChatGPT for robotics kind of moment. I think we're some ways away because like I said, I do read a lot of papers, uh, robotic papers, um, um, and most of these robotic paper will always have this chart, you know that they have a success rate chart. All of them are still measured in linear scale. So usually a lot of time, the state of the art, like you know, let's say success rate will be like oh, 70%, 80% even 90%. Right. But to me, so long they're still measuring things in linear scale. They are not talking seriously enough. It needs to be log scale. It needs to be log scale. Right? Once you plotting into log scale like every 9.0.9 and then every 0.09 like it's, it's, it's is a next uh, logarithmic improvement. Um, and I think you basically need this kind of success rate in order for you to consider replacing human um, um and I should highlight replacing human labor in very generalizable kind uh, of settings. Um, obviously for decades now we have robots that automate in manufacturing config line and those um, already done very well by classical robotics where you have one robot that does one thing, one task repeatedly thousands of times a day, super accurate, super precise. Um, and those are, I would say, fully solved by now. So that is not, in my view, what is exciting about robotics today, what is exciting about robotics is because we start to have some line of sight on, maybe these robots can be very generalizable because humans are so generalizable. And so, um, I think we are on a path, but it's still a lot of, I think, things that are holding us back. And not just the AI, by the way, it's also the hardware. For example, I'll just give you a quick example. Um, right now we don't have, uh, touch sensors anywhere near what humans are capable. Right. So, for example, I have this object in my hand, I can put it behind my hand and I can do this. I don't need to see it. Right. Obviously you need touch. Right. Um, but I would say nearly everything you see in robotics today, like all this, um, robotics research paper, 99% is computer vision. Um, but if you think about, there's a lot of tasks that actually, the way humans does it and the performance we do come from touch, we take it for granted because we don't even think about it. Right. Um, um, so I would say hardware like this, we, uh, also need hardware improvement. Um, yeah, I would say the other thing is also like all these actuators, um, that you see in humanoids today, they are nowhere near as agile and as good as, let's say, human joints or animal joints, basically. So, uh, yeah, I think we have few factors away, even on the hardware side, a few generations, I would say, before we start to approach what humans can do.

Speaker B: Yeah, I want to talk to you about the use of crypto in the sense of, hey, you're thinking, okay, we want to build out this entire network of hardware devices. Crypto's a great mechanism for that. Helium being a perfect example. And then alternatively, we want to create a system that encourages usage and engagement. So we're going to gamify that. So in your, in your opinion, the crypto aspect is, hey, we're going to scale hardware because people want to, you know, buy these devices because there's a crypto element to it. And we're going to. People will engage with the actual, uh, product on a deeper level because they're playing and they're getting rewarded potentially with some sort of token or NFTs. So could you kind of describe that aspect?

Speaker A: Yeah, I would say, um, deep in. And maybe I would say all the play to earn games in a previous cycle, uh, they all had pretty Bad name. But I think it's harder to imagine those working well because what is the utility of playing a game? Right. Uh, but why should you be paid to play a game? But in this case, because the data generator, let's say it could be useful, then there's a solid reason why you should be paid to play the game. Um, and I think honestly, my very naive way of thinking about, or simplistic way of thinking about why I'm using crypto and why I'm in crypto is that if you set this up right, it's a very fair way to distribute m, um, rewards per your efforts put in. It's just a crowdsourcing at scale in steroid. Yeah. So, so, um, you do have to set up the tokenomics or basically this incentivization properly so everyone contributing feel that they are getting their fair share. But, uh, once that is done, then it's kind of permissionless. People will just show and do stuff.

Speaker B: Right?

Speaker A: And this has happened many, many, many times across many different crypto projects. It has grown very, very big, very quickly. So I think crypto has that ability. Um, I mean people call it capital formation, but it's not just the, uh, financial capital. It's also just human labor. Right? So, um, human labor could be, um, let's say again, the gamer from Philippines who are putting in time to teleoperate a robot. It could also be down to maybe not the gamer who teleoperate, but maybe a guy who love or don't mind having a bunch of robots in his home and then he will maintain it, charge it every day and then put it out there. Um, um, so some humans need to do that, right? Because, uh, I think the robots are so good that they can charge themselves. You know, you need basically babysitter or bot sitter to learn after these sports and you need a lot of human. Sorry, a lot of humans involved. Um, my own take about this is that, um, I mean, I do work a lot with, um, or partner with many of these, like really brilliant researchers. But one maybe thing that I sometimes disagree with them is that I think it could be just that because they're researchers, they need to think this way. But I find that oftentimes they're trying to remove the humans in the loop far too quickly. Um, if you look at again what Tesla is doing, uh, their human to robot ratio is one to one. They are still one to one because every car has a human driving it, babysitting it essentially. And so I think you shouldn't. My view is that you shouldn't remove the humans too early until you truly have level five. And so, and so I take the view that actually a lot of things you do need a lot of humans involved because the humans are way more capable physically and also the physical intelligence as well. Um, um, and crypto, again, it's a great way to incentivize people to spend time, spend human labor. Um, of course, things like compute and all the hardware stuff I think is all proven, right? I don't need to shoot this anymore. I think it's is fairly proven but I would say like manual labor is very much needed. Um, and that's something that um, maybe for, for people who don't. I would say a lot of times for outsiders looking into robotics, they um, would think oh, the Capex, right? Obviously the Capex is, is non trivial. Um, uh, especially humanoids. Humanoids are still pretty expensive, right? Um, you know, if you want to get a decent one from unitree or booster, it's, it's still like 30, 30 grand, right? It's like a car, you know. Um, and so, but um, I mean I would say the operational cost of maintaining those robots, uh, it's, it's broadly speaking, you know, in my mind usually it's the rough math I have in my mind is if you put $1 to buy the Capex, the hardware, you probably need to budget around another dollar to run the opex for another year.

Speaker B: Do you think the next five years, uh, the future of robotics, whether it be, you know, humanoids, etc, will they all be teleoperated?

Speaker A: Uh, did you say teleoperated?

Speaker B: Yeah, but, well, in order for them to actually be useful in like real world scenarios, will they have to be tele operated, you think? And I, I know you know this is a departure from, you're talking about the cost which is I, I, I can't believe it's dollar, uh, one of Capex is, is a dollar of maintenance per year. That's crazy. But just, you know, hearing this, I'm like oh man, maybe we're, maybe we're going to need the human in loop for much longer than I expected. I thought it was going to be, I'm just reading things on Twitter and whatnot. But oh, we're going to have humanoids in two, three years and it's going to be a crazy future. But the way you're talking, I'm thinking, oh man, maybe over the next five years we're going to need teleoperated, uh, humanoids and robotics, uh, in order for them to be active in real world environments.

Speaker A: Yeah, I think depending on the task um, I think within two years probably you, you will have um, robotic arms that can autonomously fold your clothes pretty well and it could be around. It would still be an expensive thing but it would be like an expensive washing machine let's say. And so I think some household will definitely start to want to uh, maybe buy that. Um, yeah so but then again a machine that can fold clothes is still not doing the thing end to end. Right. Because what you may want is also for you to hang clothes and for you to maybe I don't know, uh, set aside clothes, maybe some clothes you need to keep for longer term you put in a box and then some clothes in your hand. Um and it's kind of like maybe a little bit like um, all this robot vacuum. Right. Um, so I used to have one a top of the line one of those Chinese uh, vendor that do even like um, um uh mopping the floor as well. So it's both water plus plus uh, it's really really good. I find that ultimately I still end up using my Dyson I'll do it myself because even it takes time away from me but I'm just, just that much more efficient and that much more thorough. So I think it's going to be like that there will be people who will be okay maybe with 99% or maybe 95% of the work. Um but again the last 1 to 5% will still take another order m maybe another 5 years after that. So you start to see I would say some of these more say bespoke system that will start to come into household. Uh certainly I think manufacturing setting as well. Um, personally I know that there are a lot of bets like for example figure is doing trying to deploy humanoid in manufacturing settings. Um, yeah I think it's very impressive what they're doing. So maybe famous last word but I sometimes wonder especially in a um, control environment where maybe it might make sense to just build a more purpose built ah robot um, that need not to have that humanoid form factor. Why do you need two legs? Um, the whole warehouse is flat. Presumably you can probably just have a wheel base which will make the whole thing a lot easier to solve. Um but in any case I think to answer your questions it's not going to be like the digital AI thing. It's not going to take off like that because you can't speed up time and evaluation in the real world takes time. Right. Uh analogy sometimes I give is imagine let's say Tesla ship the next FSD version, let's say FSD 14. And let's say within a day they have a million miles collected using fsd. And let's say there's no crash or disengagements. The question is, does Tesla or anyone in the world know whether they truly solve level five? You don't really know because what if in the next month there's a crash? So what I started to realize is that in the physical world, or embodied AI robotics AI, the only verifiable fact is when there's a real world failure. So it's very pretty paradoxical, right? So the only positive thing you can get is when things go to shit. Um, so in a way the better that your AI model, the longer it will take before you see some real shit goes down. And so it will take you longer to verify whether you have truly have a model. Again because of all these dynamics, I do think that it will take a long time actually to it will feel a lot longer. Um, but I do think they will come also rather soon for maybe smaller use cases like folding of clothes and whatnot.

Speaker B: And briefly you mentioned um, the maintenance costs, which that to me is shocking. You said, for example we have a $30,000 unitree robot. Uh, but you said a dollar of capex might cost you a dollar of maintenance per year. What is that? Are these replacement parts or what is this maintenance going towards?

Speaker A: Yeah, it's mainly the human, right? Because these robots are going to break all the time M and they are also in general not safe near untrained people. So you oftentimes basically have a human there babysitting it. Right. Um, I mean look at a humanoid, um, basically you can think of it like an electric scooter that is one meter above ground. Imagine electric scooter dropping on your feet one meter from above. You know, is this roughly that. So it will hurt and uh, it may be extremely dangerous if you have kids or like really elderly people. Right. And also the way especially I mean since we are talking about humanoid, like a lot of the. If you look closely at a lot of these, um, humanoid manufacturer today, like the way they structure the mechanical design is that they basically have actuators at the joints. Now if you look at human beings or any animals for that matter, the actuators are in your core, right? So in your bicep the joints are where you have basically are pivot points, so you just have tendons. And so these are all cable driven, right? So it matters because if your actuators is here then you're relying on really good actuators and it's amazing what the actuators can do this but they are actually creating extremely big um, um. Basically angular momentum. So for them to stop they need to really just stop. But if you get hit by it, it's not as compliant if you get hit by an arm like that. Because uh, if you do the so called the full angular momentum calculation you realize it's actually a lot less. Um. But I, so I would say like it's, it's therefore it's, it's, it's actually not really safe. If you get hit by this guy humanoid you definitely get a bruise. Right. Um, I do think like for example 1x I think maybe Optimus from uh texter they are starting to do maybe tendon based um along some joints. But you can imagine is harder to maintain and harder to designed for and harder to manufacture. Um, yeah. So whereas if it's uh actuators at the joints, if it fail you just take it out and then put it in. Right. Because everything is just static part. Right. So it's very easy to maintain and therefore cheap to scale up. So I think it's a trade off. And so um. But yeah, I mean so I think safety is another thing as well. Um, yeah. So uh, along the path to having good humanoids that let's say help you in the home will definitely have humanoid that accidentally maybe even kill somebody. I think it will be like that.

Speaker B: Right.

Speaker A: And then there will be some backlash. There'll be a lot of safety issues. Not uh, to mention all the privacy consideration. Um, people complain about having um, uh, like webcams that ah, you know like uh, all this, all this from Google or Nest. Um and so now instead of static camera you have a camera that's stick onto something that can also move. Right. So it can do stuff, you know. Uh, and so uh, there's a lot of things that on the path there, there will be a lot of I think compliance, privacy, safety issues that we still need to iron out. Um, and I suspect different society, like different cities, different countries, uh, will have different ways of treating these. Yeah.

Speaker B: So overall humans are going to be in a loop a lot longer than I expected. But for you it makes total sense and there's safety issues and efficiency gains and whatnot. So is that why you decided also to start creating Bit robot which is essentially the kind of the crowdsourcing uh platform that does it plug into Frodobots or how does Bitrobot fit within this ecosystem that you're creating.

Speaker A: Yeah. So I started having this idea for Bitrobot about a year ago. Um, um, because at the time, basically, obviously we've been doing cyber robot and we started um, collaborating with a lot of researchers and I started to realize, oh no, in fact they started to realize, oh, besides our robot, can you do some other robot? And, and then one of them asked me to take a look at this project that do like robotic arm and I said, oh, okay, maybe I can do it. So it started that way. Uh, and then I started to realize maybe this cyber robot thing that we're doing that eventually, I mean I wanted to have a crypto incentive to collect this cyber robot data set. Let's say it's in a way maybe a specialized case of a much more generalized case where again you can use crypto incentives to get many different types of resources and do it a lot more efficiently than a big centralized entity like the Google and the openais of the world. Um, I think there's some things that I think they will do very well. But there are some things that for example, I can't imagine OpenAI or Google Go and try to have tiny cyber robots running around in hundreds of cities around the world. I think the privacy backlash will be too big for them probably. But for tiny weird crypto project, it's completely okay because anyway it's all cross source, it's all permissionless. Right. Um, um, but yeah, I started to realize you don't need to just limit to cyborg robot. You can do in fact all kinds of robotic embodiment. It uh, could be humanoid, it could be surgical robots, it could be drones and whatnot. And then on top of that if you start thinking like that, you can in fact go beyond just so called, let's say real world data collection. You can in fact, for example, um, I can use critical incentives, incentives to get a lot of GPU to help you train. Let's uh, say world model. World models is on the rage today. And world model in a way is a data driven simulation engine that can help you. Um, maybe it's a better simulation engine, but basically give you synthetic data that you can use to train robots. So this is still data collection, but it's data collection in a simulation environment. But to do that you basically need a lot of gpu. And I think Kutu can definitely get you gpus. And I would say uh, around that time I started looking into Bittensor and I would say the reason why we call it Bitrobot, honestly I want to Kind of pay homage to the bitrater Bittensor folks. Because I would say before Bittensor I wouldn't have believed that it's possible to do a network of subnets. Um, I think other projects have tried to do it, but uh, I think Bittensor was the one that convinced me that it's possible. Right. Um, so honestly I also kind of went down the rabbit hole a little bit. I did seriously consider maybe doing a subnet on Bittensor. Um, but what gave me the push to do Bitrobot as a separate thing was I think one which is that I believe embodied AI will be so big, uh, it deserves its own Bittensor. Right. Uh, so that's one. The other part of it is I just find that it's a little bit too difficult to compare. Let's say there are subnets that do protein folding, right. On Bittensor there's another subnet to do rom, let's say something to do RM M. And then if you have another robotic thing as another subnet, I just find that all of them are very important, but they're all completely different. And so when it comes time to distribute the token, it's almost like so different, qualitatively different, that there's no basis of judgment between them. Whereas if you set up uh, Bittensor like thing that's basically a big robot is going to be a network of subnets. Right. Um, even though you could have one subnet, let's say doing cyber robot stuff, another one doing humanoid fried and egg let's say, or maybe another one doing world models for self driving, let's say, even though they are all different but because it's still in an embodied AI space. Um, I think you can at least compare them, um, not Apple, it's got to be green apple to red Apple kind of comparison I think. And part of that belief actually comes from also ah, um, I would say this line of research called cross Embodiment. Um, so it's a pretty hot research area the last 12 years. In fact the best paper that came from the biggest robot conference last year from the DeepMind folks is about exactly this. Right. So the idea for Cross Embodiment is that it's possible for you to train a more generalizable and in a way a more performance AI model and embodied AI model if you have data set, let's say coming from uh, robotic arm as well as, let's say cyber robot. So the most dramatic paper I've seen that they kind of showcase this is uh, I think there's another Paper from Berkeley, I can't remember the paper name but basically these guys, um, train, um, ah, a model that can fly a drone using data from a cyber robot as well as robotic um, arm. It's crazy, right? Obviously it's a very pretty bad model. But somehow this model that fly a drone is able to fly a drone. So I think the intuition there is that ultimately all these robots, they all operate on earth and everything is subject to the same gravity. Uh, physical law is the same everywhere. Right. And so there's a fair bit of uh, positive uh, learning transfer, let's say whether you are gathering learnings from a humanoid form factor or you're learning from cyber robot. So just now I mentioned about why solving cyborg robot, like having cyborg robot data is important and I mentioned humanoid because even though the data you collect from this small little bugger, actually this data set I believe can be transferred quite easily to humanoid. Um, if you take that view, therefore you can argue that there are all these different subnets. Eventually the whole is going to be bigger than the sum of parts and there's a lot of synergy between them. Um, and so yeah, so I would say the whole cross embodiment line of research added a bit more like, I would say scientific rigor to my thinking about, okay, why something like Bitrobot may make sense. Um, and so yeah, it's a very ambitious project I do think, trying to use crypto incentive to gather all sorts of resources, um, to make a meaningful uh, push to embody AI. Um, I think crypto has a chance to do something meaningful. Um, but we do have not a very big window to do that I think in the next few years because the web two guys, clearly the big ones are uh, the startups. The big startups are raising hundreds of millions. Um, the big boys. OpenAI, Google, uh, DeepMind, Gemini Robotics, Nvidia. There are huge resources being poured in, but I do think there are some actually not like a number of things that I think just it's way better to do it distributed to crowdsource basically. Um, so I think that's where crypto can really move the needle.

Speaker B: So this is a bad analogy, but would you say that if you had to summarize what you're roughly building is scale AI for robotics, because scale AI is doing the data labeling before they were doing it with humans and then over time they used AI to help label the data, they still probably use some humans. M. Is that kind of a rough analogy? It's like, hey, we're going to get this really valuable data and help all these companies with their data needs, um, by using humans and humans combined with AI to collect all this very valuable real world data. Is that like a okay analogy or not really.

Speaker A: Yeah, I think that will capture maybe half the story, you know, because scale, scale, I do data only, more or less. And obviously Bit Robot will have a bunch of subnets collecting data. Uh, uh, it could be real world data, teleoperation data, but like I mentioned, it could be synthetic data, it could also be video data. But uh, beyond data collection, I think there will be subnets that actually try to build, build the models themselves. So again I mentioned about that competition that we co organized with the DeepMind folks. If you look at the competition, which is going to be a subnet by the way, um, what happens in that competition is that we have a bunch of researchers from all these universities, they take some data coming from, let's say Subnet one, which have all this cyber robot data, but then this researcher take those data, we give them a bunch of compute, again using crypto incentives. At the end of the day, what is the outcome of that competition is a bunch of AR models. And so that actually will, I think should just be. If we wanted 1000 Axis, we definitely want to do that for that competition. How do we give these researchers a lot more data, a lot more compute and a lot more robots to test in the real world. Right, but this subnet, the output is not data. The subnet is actually the final thing, the AI model. And so I think there will be subnets in Bitrobot that actually get to the final thing. Um, and I would argue actually the more obviously I started collecting data as a thing. I do take the view now that I think in a way, data, ah, if you have proprietary data and variable data that's useful for research, it gives you a seat at the table, let's say. But um, in terms of value capture is very hard to do, technically speaking. Um, and also actually economically speaking, you probably don't deserve. It will be very hard to capture a decent, um, slice of the pie in a way. If you look at what OpenAI pays Reddit, I think they pay them 50 bucks a year or something like that. It's peanuts compared to OpenAI. But I would argue Reddit is probably one of the best data set out there on the Internet. Um, so as a proprietary data provider, Reddit gets a tiny amount. So I think that the biggest value capture will be done by those who do the final thing. Because what ultimately everyone wants is the final AI model, the real physical, artificial physical intelligence. Right, the uh, embodied AI. Um, if you can do that then then it's very easy to monetize. You just sell transformer tokens, um, because that's the final thing people want. Right. Um, and I think if Bitrobot, some subnets are able to do that then hopefully uh, Bitrobot as a network can capture some of the value and then obviously that value gets transferred back to the token holders. Um, yeah, so I would say scale AI only due to data, but I think Bitrobot will have something that do more than just data.

Speaker B: So can you give me the roadmap for. You know, obviously it's going to be some, some guesstimation here, but the roadmap for the next decade of robotics, you know, like let's say like, oh, we're going to have um, you know, delivery wheeled vehicles in three years or you know like, like fully autonomous or whatever. Even though it's kind of already happening today. But. And uh, then five years we're going to have humanoids and then in seven years we're going to have uh, drones doing delivery. Like can you do your best guess, give us kind of the 10 year roadmap for robotics. I know it's a big question, but what do you think?

Speaker A: Yeah, okay, maybe famous last words for me, but I'm slightly less optimistic than some of the other guys timeline. Um, I think home humanoid is probably the biggest going to be probably a trillion dollar thing like what Elon Musk kept saying. Uh, I think to do it at the price point that the household will want to buy one as if you're buying a washing machine and have it safe enough that you and then also have it do enough things to justify the price. Um, I'm guessing it's five to 10 years. I don't know whether uh, it's far enough and I mean it could be five years if certain things start to come through, but it also could be ten years. Um, um, I mean I keep coming back to the self driving because you know like I think we are right about at the curse where self driving really starting to work, starting to scale. Um, I sat for the first time in a self driving car in 2015 so I got the chance to set in. This is like before Waymo, right? So uh, I got a chance. A buddy of mine from Google let me uh, go for a joyride. And this is like, I think they call it Google Brain. It's this circular thing and I was just mind blown, right? I was Driving around in Palo Alto, um, autonomously. And I thought, oh gosh man, this is going to be everywhere in the next few years. And of course like 10 years later we're still talking about it. And so I think, I suspect it might be like that. Um, it's like we are so close, but then the closer we are, we still realize that it's like the next nine still matters. Um, and not to forget, like in a lot of places human labor is quite cheap. So um, of course there, I mean there are also places where like you know, uh, people are running into uh, labor shortage all the time. Right. You know, you know, so it's ah, um, yeah, I guess, I mean, yeah, maybe five to 10 years. Yeah. But maybe some of these like more specific like uh, things like I said, like a bi. Manual arm, um, that is folding clothes. I think that's something like two years away. So I think we're quite close on those. And if you ask me this about the folding clothes thing, one year ago I will have guessed five years. But clearly my timeline has shortened because I've seen stuff and I'm quite convinced it can be done in the next two years. Um, so I might be wrong again on the big humanoid thing. Uh, uh, yeah, um, I mean for now for sure, I won't put the humanoids anywhere near my kids or my parents. Yeah, it's not safe. Um, yeah.

Speaker B: What do you think about drone delivery? Is that, is the primary thing, stopping drone delivery? Is that regulation or do you think there's still a big safety and kind of intelligence issue around drones?

Speaker A: Yeah, I think delivery, you're talking about last mile delivery, right? Like the sidewalk, right? Um, mhm, good question. Uh, I think to replace the doordash guy or the Amazon delivery guy, maybe at least another five years or so.

Speaker B: Wow. Same as humanoids.

Speaker A: At least another five years. Yeah, yeah, at least another five years. I mean in some sense, uh, I don't know whether it's overkill to get a humanoid to do the last mile delivery. Maybe the right form factor is still something like a box, like the cyber robot thing that you've seen around. Um, but then that is still a bit of a question. I mean like I said, the hardware, they can't climb stairs, well, they can't press the lift, they can't go escalator. So um, yeah, and also the AI stuff I think is non trivial. Like I say, I do believe sidewalk navigation is actually harder than self driving on the road. So um, yeah, I mean to my knowledge most of these sidewalk Robot startups, they actually, some of them maybe will claim that they are level four, but I mean, if I'm not wrong, most of them are definitely have tablet operators in the loop, right? Very much so. And so, um, I think you cannot get to the unit economics where you can compete against a human, um, until you truly don't have humans in the loop. And That's a level five. Right. And the gulf between level two and level five could be another 10 years. Yeah. So I think maybe some robot is even harder. Yeah. So it's at least another five years also, in my view.

Speaker B: So this is a naive question, but, you know, I, I am seeing the rapid increase of, I think ChatGpt or I forget what it is. Sam Altman was saying that intelligence and AI models is, is increasing, uh, 300% per year. Right. So just like, you know, compounding at just an absurd rate, you know, log scale, as you mentioned before, um, you know that that's kind of the true test of, of, of, of, you know, progress right there. And, you know, couldn't we have a scenario where like, in two years we have just, uh. Even a sidewalk robot is just unbelievably smart. Right. It just, it just is as smart as a human. So it can navigate the dog and go around the rock and whatever and avoid the people that are trying to attack it or whatnot. And wouldn't that accelerate the timelines massively if we can get very cheap, uh, very intel, you know, highly intelligent computers to fit inside of these bots.

Speaker A: Yeah. So what you're describing is basically the pace of improvement for AI that is in a digital realm. And I definitely think that we will have AGI by that definition in the next two, three years. So, yeah, I do believe we will have AGI before we have embodied AI, I guess. Embodied AGI. Yeah. Uh, I think the last remaining jobs for humans are probably hairdresser, plumber, athletes, uh, chef, I think. Chef, yeah. Extremely hard to do what a chef can do. Yeah.

Speaker B: But like, so, for example, if I. A drone, if we give it AGI level intelligence in this drone's body, if you will, and, um, it can see everything and can kind of navigate. Couldn't it just navigate all the issues and pretend that there was a parachute or whatever, so it's very safe or whatnot? But couldn't it just kind of overcome a lot of these physical AI issues that we're talking about because it's as smart as a person or you're like, no, literally the limiting factor is just like the data Set like it just can't get that smart in the physical world because we're not training in the physical world yet on a mass scale.

Speaker A: Yeah, I think the iteration speed is very different for robotics. Um, especially if you consider the last step, which is evaluation. Um, you know, when OpenAI or Deep Seq, they open a model or llama, you realize that the first thing they show you is a bunch of benchmarks. And obviously there's no perfect benchmarks, but they get those benchmarks probably within an hour or less. So even though there's no perfect benchmark in that world, but they are never, ah, um, a bottleneck in the iteration speed. But again, what I mentioned about the 1 million miles thing for Tesla, how do you know whether you build a very good embodied AI model unless you truly test it in the real world? You cannot speed up time. Uh, so there will be this big unknown period where you are just waiting to see things fail. Um, and because you can't speed up time in the real world, you would just have a very, very long period where you have to kind of just lock in more evidence that this won't fail. This won't fail. If you notice the iteration cycle for let's say Tesla, there's nothing to get longer and longer. A little bit like that. That's only my suspicion. My suspicion is they just need more miles to see the next failure case beyond a certain point. I think these models mainly learn from failures, not so called positive examples. And when, uh, the failure cases are so far in between, you just need take more time and you can't speed that up. Um, now the other thing I also want to point out is that sometimes I wonder, um, whether we are in a local minima where right now the way we are scaling towards, let's say intelligence is through scaling through data and compute, right? Um, so everyone understood that. Now but if you think about, okay, I call this the mosquito intelligence. So if you look at a mosquito, right, its power consumption is probably 0.1 watt or less than that. Um, I don't know whether you try to kill mosquito before, it's damn hard to kill one. So these guys are damn agile, which means they are damn good physical intelligence. So how can something that consumes so little power, um, I don't know, it seems to me that this guy, I mean you can say that, oh, he's got hundreds of millions of, uh, experience that basically learn from that. But the fact is that a mosquito will not remember at all what the generations did, right? And yet very soon, once the Guy become mosquito. He can fly around and figure things out and then basically be really good at it. So it seems to me there's a very different paradigm that has maybe nothing to do with scaling. And there are some research that's kind of in the um, robotics world going towards this line. So, um, there's this group that I really like from Imperial College, um, Edward John's group. Um, I interviewed him recently on our own podcast called Robo Paper. But basically his group focused a lot on this concept of in context learning, basically small model and you basically instead of loading it with a ton of data, you just teach it while, give it some demo while it's doing it and then we just correct it. Which is kind of like what humans can do, right? You donate a million examples before you know how to pass your mouse from left right hand to left hand. Um, uh, I don't know, maybe that's another big element here. Surprise down the road where someone figure out another way to scale. M maybe scale is not the word to use, it's just a different algorithm breakthrough, something to do with maybe offline reinforcement learning. Um, I suspect, if I were to guess, um, they're basically way more data efficient than what is now. Um, then actually we are not talking about having a lot of compute and having a lot of data is literally just how the animal kingdom does it. Um, I think that's the other thing that sometimes I wonder, I don't know. Uh, but at least we don't have the algorithm yet. So if we don't have the algorithm breakthrough then we are still on this path of more data, more compute,

Speaker B: use

Speaker A: crypto to get thousands of people around the world, contribute and build towards that.

Speaker B: Yeah, but, so another naive question here, but couldn't I make a digital twin of the roads? Uh, let's pretend I'm Tesla and I want to accelerate my self driving learning. Couldn't I make a digital twin of the roads? And of course it's not going to be exactly one to one, but it'll be pretty good simulation. And couldn't I just make a million iterations of that and have the full self driving train in real time or no, the digital twin technology is not even remotely good enough or remotely close enough to the actual world.

Speaker A: Yeah, I think simulation there are a few issues. Uh, well, I must say that I think simulation, especially world models will have a big role to play for robotics. I do think so, but it's not going to solve everything. And there are a couple of issues with simulation. So, um, there are Some tasks where simulation will fail pretty miserably. Like where simulation has done well is usually where, um, they call it local motion, where let's say you've seen all this crazy humanoid doing dancing and all this, right? But if you look at really what the task at hand, the environment is very static. There's no other human or robots in the environment. Uh, and that's where simulation do very well. Simulation in my view will completely not work for navigation because the difficult part is not simulating. I mean this is the easiest robot in the world. It only has 2 degrees of free freedom, right? Go straight, go back, turn left and right. That's it. So the difficult in simulation is not simulating what you do as a robot. The difficult part is simulating the environment. But it becomes very chicken and ache, right? If you have not solved level 5 autonomy, how do you simulate the behavior of the next car of the pedestrian who are all level five beings, right? So that's one big chicken thing, right? So on some tasks I don't think simulation alone can solve it. Um, I think there's a big question now, maybe very, uh, good data driven world models can solve it potentially. Um, I think it's still very much tbd. Um, the reason I'm still a little bit, I don't think it alone can solve it is because, yes, in simulation, in theory, you can speed up time like you said, one billion times. Let's say the thing though, the thought experiment I would run through usually is like this, right? So imagine, let's say you ask a simulation engine to imagine, uh, a robotic arm that's frying an egg, okay? Now just think of all the different permutations you can potentially have of a robotic arm that's frying egg. Every egg looks similar, but then every egg is unique, right? Um, not only that, I mean the, the, the frying pan, you know, the lighting in the environment, you know, fire versus electric, like this. If you consider all these variables, the search space or exploration space, that means potential alternate worlds you can simulate is nearly infinite now. So what if you can speed up time by 1 billion x? Question is, is this a good sample size of the entire alternate universe that you can have? I mean in statistics, right? Basically you just sample and you take a sample. You know, it's definitely a small sample of the overall population. But if you more or less know that this is a normal distribution and whatnot, that you can tell something about the overall population. What I'm not so sure about uh, simulation is that what if the world is Just not normal distribution. It's just so many different edge cases potential. Right. And so what if you can 1 billion x time, you're still barely pushing on a string. Maybe it's not enough. Whereas if you collect things in the real world, at least it's real. Right? So it would directly translate, you know, for sure it's real. Um, um, it's not clear to me whether you 1 billion x time you can actually cover enough of the search space. Right. Maybe the search rate is so big then. The last thing I would say is that when you talk about simulation, at least as of today, we are very, in my view, very bottlenecked by the talents who know how to run simulation. Well, my own guess is that there are only a few thousand people, or maybe less than a thousand people in the world who knows how to run it. Well, a lot of people can do simulation. When I say run it well, that means run it at scale without wasting your gpu. Because if you don't know what you're doing, you can simulate whatever you want to simulate. But you always have this, what they call sim to real gap. Because whatever you simulate doesn't give a wide enough distribution that mirrors what may happen in the real world. And if it doesn't mirror enough, catch all the edge cases that may happen in the real world, then you have wasted a ton of gpu. Right. And so to do simulation, well, you are very bottlenecked by those basically unicorn researchers who know what to do when given a big cluster of H1 hundreds, let's say. And we're also very bottlenecked by those guys, I think. Um, yeah, of course Nvidia will want to tell everyone, hey, simulation soft everyone, because they have all the compute and they also got the talent to run the compute, to run the simulation. M uh, but yeah, so I think to solve robotics you would need simulation. You probably also need video data, but you also need teleop data in the real world. Right. They all have their own different trade offs. Um, and yeah, you will need all three, I think.

Speaker B: Which industries are you thinking will be transformed by robotics first? And of course there's manufacturing, which I think that there's already heavy use of robotics in manufacturing. But I guess I'm talking about more of like the level above that. Not, not just like one task all day, all day. More about this, uh, the ability to do different tasks. Kind of like that last mile m delivery that we're talking about.

Speaker A: Yeah, I imagine it would be easier to be done in manufacturing. I suppose there are still Human jobs, physical human jobs that are in manufacturing plant. Um, because once you're in a manufacturing settings then safety is less of a concern for sure. Um, so it'll be easier to deploy them. Um, but I'm not sure, I mean I've never done manufacturing that level kind of automation. I mean we do manufacture like toys today, right? Um, um, it's not very clear to me, uh, whether that's the biggest one. I do think the biggest one is humanoids at homes or in, let's say urban settings. I think that is the biggest one. Probably a trillion dollar at least kind of. Uh.

Speaker B: Which is the first though?

Speaker A: Mhm. Sorry.

Speaker B: Because humanoids, uh, I agree are the ultimate kind of end goal. But which it sounds like that's kind of a little bit farther away. Which industries are going to be transformed first by robotics besides manufacturing?

Speaker A: Yeah, I think manufacturing first, but also some more like limited. I mean I keep mentioning about folding because I personally absolutely hate folding clothes, but I have to do it. Uh, so um, I will probably be one of the early buyer. I don't mind paying thousands of dollars for a machine that can fold clothes for me. Um, and I'm okay with 95% success rate because I just hate it so much. So that would be, I think this kind that will be faster. Um, um, but that is still a fairly restricted like single use case kind of robots, right? Um, yeah. Um, but I think the biggest again is a humanoid form factor. I do believe that. But we are quite far, I would

Speaker B: say for the folks that are listening and that are bullish on the robot revolution. What businesses do you think that they could start in order to ride the wave of robotics or kind of be involved? But they're like, listen, I'm not going to try to out compete Google or Microsoft or whatever on these robotics, but I do want to ride this wave and be involved. Have you thought about any business opportunities that are tangential? Like maybe. Oh, I'm going to train to become um, a technician for humanoids and other robotics and just kind of be able to fix them and set up a repair shop or like what are these business opportunities in your mind that you've uh, pondered?

Speaker A: Yeah, I think that's a great question. So I think that's why we want to set up beatrobot. Um, it's to allow, let's say again that gamer in Philippines to feel that he has um, a slice in a pie. Um, I think a lot of them maybe grew up watching Star wars or whatnot and love robots, but obviously don't have the opportunity to become a researchers and whatnot. Um, but like I said, I do very much believe we need a lot of humans in the loop and so could be humans who help to look after, do basic maintenance, charge the robot, that kind of stuff, deploy them. Basically. Um, we need a lot of these uh, humans to do it. We also need probably humans to do teleoperation. Right? Um, yeah. And I think, um, I would say these are all very active ways to so called participate. Um, but of course for those who are intellectually curious about the space, um, I do want to shield my own podcast called Robo Papers. So I do this weekly podcast uh, with this friend, um, of mine called Chris Paston. So Chris leads the AI efforts at Agility, uh, another humanoid company, uh, in us but we basically every year we interview authors from basically some of the top robotic AI papers. Um, pretty technical, but because I'm a layman, um, more layman I would say definitely compared to Chris. So I try to ask all the dumb questions, but I would say it's still not for the faint of heart. It's quite technical sometimes. But uh, if you really want to deep dive, I think Robo Papers, uh, that podcast is actually really good. I do think as a resource we have usually some of the best paper that get published. Usually within a few weeks we'll get them on the podcast. Um, recently like cvpr, um, the biggest computer vision conference, the best paper there. Um, yeah, the author was on our podcast last month. And then there was another best paper recently that was also on our podcast. I think we were quite lucky to get good researchers. So I mean for those, if you are interested, can listen into those. Um, um, honestly I try to keep up with the research. It helps me a lot. I mean I'm geeky that way and like I said, I read a lot of papers. Uh, I think it forces me to understand maybe, let's say 50% of the paper, maybe in the past I only understand 20 to 30%. Um, um, yeah, I do think that, I mean obviously again I'm biased, but seeing where uh, AI is advancing in the digital realm, I think once we hit real AGI in the digital realm, I imagine the only investable and worthwhile things for humans to do is in robotics. So if you take that view, right. I mean, what's that to do? If you have AGI in the digital realm, the only thing left is in the physical world where we still have a bit of an edge. And so if you take that view then actually it's not too late to invest in robotics. It might be the only last things that humans can still contribute because the AI still suck it, like still cannot fry an egg.

Speaker B: All right, well, maybe my next after, uh, Sir Fermion and this podcast, I'll start training to be a plumber or a chef. Those seem safe for the next few years, so that'll be my target. Um, all right, so you're reading all these papers and talking to all these incredible people at Robo. The podcast is Robo papers, right? Yeah, Robo papers. And what are some, like, underrated breakthroughs or kind of like, you know, cool trends or. Because you're on the bleeding edge. Right. So what are some things that are really exciting for you that you've learned that these, these researchers, uh, are, or, you know, founders are creating that has you super pumped up because, you know, we're talking in, uh, this podcast is of course, everything you're building with Frodobots and Bit Robot and everything else. And it's very exciting, but you're also tampering the expectation. You're like, listen, like, we're not going to get humanoids tomorrow. It's going to take some years. So in terms of these breakthroughs that you're reading about, anything that you're like, oh, my gosh, this is incredible. Or is it still pretty tampered and kind of, you know, more of a linear progress?

Speaker A: Yeah, I would say, I mean, for robotics is harder to get the wow factor, unlike the AI in the digital realm. Like, you know, because the pace of iteration is so fast, robotics is just slow. And you look at it, it's like, huh, you know, it can fall across. And you think this is like state of the art. Right. You know, but honestly, that's where like, uh, robotics still is today. Um, but there are a level of trends on the research world that I think is very interesting. Like, world model, like I mentioned, is a big one. Um, so world model is, in a way, uh, still a simulation, but it's a very different way to think about simulation. It's very data driven, kind of like simulation. Right. So, um, I think world models, it's going to play a big part in robotics and there's a lot of top researchers working on it now. Um, I think cross embodiment, like I mentioned earlier. Um, how do you take learnings from, let's say humanoid and then apply it to cyber robot and vice versa? I think things like that. Right. And so I think that's because if that research start to really pay dividends, then in theory we need less data, we just need more diverse data. Right. I think that's very. And um, that's why people are excited to work on it. Right. Um, uh, I think that there are like I mentioned the group from Imperial College that incontext learning. I personally think it's my favorite paper from last year actually. So uh, the paper has got instant policy from Edward John's group. Uh, so definitely worth if not read the paper then maybe check out the podcast. Um, and then I would say some of the big ones, um, you know, uh, basically what the foundational model can do out of the box. I'm talking about the Gemini, the um, um, GPT4.0 kind of like model that is multimodal, um, that are starting to be trained also using some robotic data. So I believe the Gemini guys, the way they think about it is they just treat robotic data as another modality other than text and videos and images. Um, and so the more they start to rope in robotic as a first class citizen along text and image, the more I think you get a lot out from just a huge foundation model, just from the Geminis of the world, basically a couple of huge foundational model. And so um, yeah, I think those will be really interesting to see how they can perform because right now uh, in some ways you can say that they are behind a more bespoke model. But on the other hand they are extremely generalizable um, because they do have the Internet scale data from the text and the images and the videos. Right. And so yeah, um, but like I mentioned earlier, I still at the back of my head I'm thinking hey, you know the mosquito never have, don't need the amount of power, don't need that amount of computer, probably have never had that kind of memory, probably have an extremely short term memory. How can it fly so well? Um, uh, yeah. So you have to wonder, are we missing something very, very fundamental?

Speaker B: There's got to be a better algo out there somewhere. There's some God tier algorithm uh, that will help us so one day we'll find it. Well Michael, this has been incredible and uh, thank you for going over time. I appreciate that. It's been an incredible conversation.

Speaker A: Yeah, thanks so much for having me man. It's been a great chat.

Speaker B: Yeah, of course you're going to have to come back on again in some months because I feel like uh, even though you see the pace isn't going that rapidly, maybe for you because you're in it every day, for me it seems like the pace of robotics has been absolutely insane over the past year and change. So, uh, very excited to have you on again in the future.