EP 05: Why The Robots Aren't Coming
The Inference Podcast
https://inferencepodcast.uk/episodes/ep-05-why-the-robots-arent-coming

[0:50] Mian Khubaib Jim: Tom, every machine took job from human hands and gave an easier one. This time, humanoid shape robot is taking all of them. What happens to human hands now?

[1:00] Tom Lipinski: Are you asking me to play a scenario what happens when or if they take over, or are you asking me if it's possible for them to take over?

[1:09] Mian Khubaib Jim: Both scenarios.

[1:10] Tom Lipinski: Okay. Um, let's start with the fact that a lot of work that people do requires not just a lot of dexterity, but also a lot of sensory feedback, right? So your hand has seventeen thousand sensors, right? They detect not just movement or touch, but temperature, friction, you know, deflection, and so on, and so on. We can do things blindfolded with quite a lot of ease if we've done this before. Um, it's impossible at this stage to copy that, right? We don't even know how to digitize sensation like touch and so on. We've

[1:48] Tom Lipinski: digitized-- I mean, we digitized voice in the eighties, right? When telephony was expanding and so on, and so on. So the, current AI builds on a hell of a lot of work that was done decades ago. When it comes to sensory touch, it's, it's very early science. It's very blue sky. So I, I don't think you can replace most things that people do. You can, you can pick up a lot of jobs Around human activity where, which doesn't require, you know, a, a lot of, uh, sensory feedback. pick and place, machine tending, um, picking items from rigs and putting them into jigs or fixtures or emptying boxes into hoppers, let's say, in manufacturing. So that's, So sort of super boring, repetitive work. That can be automated, and to be fair, that should be automated because it's super, super, boring thing to do. If your life revolves around picking an item five thousand times a day and putting it somewhere else, you're gonna be asking yourself some hard

[2:48] Tom Lipinski: questions, right? So you do want to automate that, and that's-- You'll see that being automated over the next five years with, uh, humanoids or general purpose robots and, and other robots that sort of you know, mobits and, and so on. But when it comes to a more demanding work where you really need to use your hands and stretch things or or clip things and, and, and so on, that's-- you can't just do it on vision, so I can't see humanoids replacing this. So when it comes to manufacturing, what could possibly happen over the next ten years is about twenty-five, maybe thirty percent of work being automated with physical AI, while the rest of the work specializing a bit more around humans' best capabilities that robots cannot copy.

[3:38] Mian Khubaib Jim: Okay. One of your essays is that humanoids are just hardware. What do you mean by that?

[3:44] Tom Lipinski: Well, just look back at personal computers, right? When, when people started

[3:48] Tom Lipinski: building them in, in the '80s, um, or late '70s. And the general view in the space was, well, let's, let's build computers and, you know, um, enthusiasts can program them to do things, right? So at that point, computers literally had nothing on them, right? They were just hardware. And in order to make them provide outcome for a customer, provide any sort of value, you had to code it yourself, right? And, And, and humanoids and, and generally speaking, physical AI robots are not far from that point. And, um, even last week, um, Unitree admitted that, that the sort of the GPT moment for, for humanoids is very far away. Um, they're just hardware. You manufacture hardware. You can put firmware to, to control servos and, and motors and, and so on. and you could even train basic motion like, like mob-- you know- Moving forward or, or dancing,

[4:48] Tom Lipinski: right? But if you look at robots dancing, they, you know, if you walk closer to them, they'll knock your teeth out. There is no-- It's a one-way traffic, right? They're following motion planning with a little bit of, uh, responsiveness to gravity, but that's about it. They can't do anything, right? So vast, vast majority of robots that you see are pre-programmed or motion planned. There's no, there's no perception there. There is no responsiveness. There's no, you know, on-edge inference happening. So it-- they just, they're just pieces of hardware.

[5:23] Mian Khubaib Jim: Okay, then what is the use case for humanoid robots? Has anyone answered that?

[5:25] Tom Lipinski: How do I... The honest answer is I can't find any, right? If you look at-- I, I read a report from, um, a Ukrainian army, um, that came out about two weeks ago stating that humanoids fail dramatically in the field, right,

[5:45] Tom Lipinski: in trenches, that actually wheeled robots or tracked robots are way better. And that was probably the last place where I thought or the final, final chance for humanoids to, to prove that, you know, they could- Mm-hmm they could be useful because in the trenches, the environment is extremely Bipedal oriented. Uh, and if they fail there, um, and they get beaten by wheeled or tracked robots, even in the trenches, in mud, in mo- in warfare, um, I can't honestly think of a, of a useful case for them.

[6:20] Mian Khubaib Jim: Okay, so the most of the people who are criticizing the robot industry, they haven't built one, and your robot spent under, I think, four thousand British floors. What does it look like in real world?

[6:30] Tom Lipinski: So what, what... again, you can can look at it from both angles, right? When it comes to how to get most value out of robots and how to get most efficiency out of robots

[6:40] Tom Lipinski: and so on, you would probably build a robot for each and every use case, and they would be specialized for that particular use case, right? And if that use case enables you to deploy thousands and thousands of robots, then you will probably get to economies of scale, right? Which means that the robots will, ar- at some point, your manufacturing cost will arrive at a point that makes it economical to deploy, right? Um, and, and floor installation is a typical case where, you know, trying to get humanoids to crawl under the floor would be idiotic, right? There's not enough space. You know, you've got constraint. You know, you might have a space this, this shallow. You've got spaghetti of pipes and cables and, um, you know, rubble and so on. And you need to design and build something that's appropriate for the application, right? I, I think appropriate is the the right word that we're missing in robotics, right? A lot of companies chase the, the human form factor, which is not appropriate for the tasks that they're targeting. So, so, um, Q-Bot is, is a good

[7:40] Tom Lipinski: example. Um, it's done probably five and a half thousand floors by now. It's where robots are designed to work within the space that they're supposed to be deployed in, and therefore, it makes the job easier and, and, and, and you can then deliver an outcome. Uh, but there are lots of other use cases where specialized robots were designed for a particular task, right? Obviously, you've got people on the other side that will say, "If you're doing specialized robotics, you will never get to economies of scale. Therefore, your robots will always be super expensive. Therefore, you will never be able to deliver value which is competitive with, let's say, employing humans and so on." Um, and that really depends on, on on indivi- individual cases. But trying to deploy, you know, quarter of a million pounds worth of humanoid into, into a task where you would pay somebody thirty thousand dollars a year to do a job is, is not economically sane.

[8:36] Mian Khubaib Jim: No. let's let's pick a, a use case. Let's say it's construction, and

[8:36] Mian Khubaib Jim: construction sites are never same, right? Like the assembly line. So would they work for, for that case as well? well?

[8:36] Tom Lipinski: Let's

[8:36] Mian Khubaib Jim: say you have developed specific robots for construction work, but sites are different. Would it be able to adapt, or you have to teleoperate or something? Or retrain

[8:36] Tom Lipinski: The a- the answer is yes to all of the questions and and even though they were-- they, they sounded like they were, they were contradictory. First of all, you can't deploy without teleoperation, right? The, the use cases are so varied and different that in order to, to arrive at any sort of limited autonomy, you need to teleoperate a hell of a lot out of it, right? And think about h-how autonomy has worked so far, right? If you take a case of one of the few successful examples, when I say successful examples, Waymo. They've not arrived at full autonomy, but they've arrived at a high percentage of autonomy. It started by them driving cars and collecting data while driving cars, and then they were teleoperating cars and

[9:19] Tom Lipinski: collecting all the data, you know, external inputs, human reactions, and so on. And then they, on the basis of millions of hours of, of, of of data, they started fine-tuning the models to actually drive the cars around and get to north of ninety percent autonomy. Obviously, Waymo taxis are not fully autonomous. There is nothing that is fully autonomous. If you think about it, it does still require, um, human intervention when, um, taxis get lost and so on. They, they still have teams of teleoperators in, in Philippines, et cetera. But they managed to, over a period of twenty-two years, they managed to build a relatively robust autonomy, uh, operating in two dimensions with six outputs, right? But that's-- it's how it worked, right? When you're looking at claims that people make about, you know, zero-shot robotic deployment and so on, um, you haven't actually-- well, we haven't seen any of that deployed on, on, on real sites with any degree of reliability. So yes, you will have to teleoperate. Secondly, you asked

[10:19] Tom Lipinski: about robots in construction. You can design relatively Universal robots. And obviously the jury's out, right? There's some people will say, "Oh, it's a humanoid." If you look at all three, uh, which is a German startup that recently raised twenty-five mil, which is a former Arrival crew, they're going for a crab or spider shape, you know? And, and I, I can see the value in that from a perspective of stability. But at the same time, single arm probably not, not ideal for, for construction and, and, uh, you might, you might use different form factors. But if you look at the job, construction, right, where you're handling large, heavy goods, the number one challenge for you is to retain stability over a reasonably uneven ground, right? Bipedals, I mean, we, we've evolved to do it, but bipedal humanoids will not give you that. So you'd need something with either a, a, a large, uh, footprint, a wheelbase or track with low center

[11:19] Tom Lipinski: of gravity that can then handle objects and, you know, two, four, six arms, I don't know, that might handle objects. So you could have a more or less universal robotic construction, but that doesn't mean that people have succeeded, right? What, what you've seen in construction is specialization taking over. So companies like Monumental, the bricklaying robot, companies like Q-Bot, the insulating robot, um, a few other companies like, like Gravis, for example, the, the sort of earthmoving robots. They are moving forward because they're highly specialized rather than universal.

[11:52] Mian Khubaib Jim: No, even... So Tesla teleoperated their, demo and, uh, there's a- another, reward, twenty thousand dollars reward that ships with the person which looks from the camera, so like the eyes of the robot. W- what does that tell us about this industry at this stage?

[12:06] Tom Lipinski: Well, I think you're referring to 1X, um-

[12:09] Mian Khubaib Jim: Neo.

[12:09] Tom Lipinski: Yeah. Neo, Yeah Neo Robot. Um, which, yes, it's pretty much teleoperated.

[12:15] Mian Khubaib Jim: And someone controlling it from the-

[12:18] Tom Lipinski: Yeah, with a VR head- From the... US ... with, with a V- VR headset. Well, first of all, that is the only way you can do it. Yeah, you can't-- You-- There is no There are no models that can run robots in, in, in various locations, in homes, um, in unpredictable environments with, um, dogs, kids, mess, chaos, and, and and so on that can actually be autonomous. Um, you need human teleoperation. What-- I w- I would turn this around and I would say, A, you can't do it in any other way, but B, this is a large-scale data collection exercise. It's not a-- It's, it's, it's a deployment, but it's actually it's a data collection exercise dressed as a service. What, um, what uh, Neo is doing is just collecting indoor data, trying to get to a volume of data to build large enough data set that might

[13:15] Tom Lipinski: be useful to train, um, large foundational model or, or world action model or something that can then start eating away that teleoperator time and, and, and start developing some sort of autonomy. Um, I wouldn't expect it to be in the seventy or eighty percent range anytime soon, but, you know, with a large enough data set, you can start moving forward in that direction.

[13:39] Mian Khubaib Jim: No, when I watched our demo, I missed it, and then I I was watching interview and we-- then somebody pointed that it written it's non-- it's not autonomous.

[13:47] Tom Lipinski: Yes.

[13:47] Mian Khubaib Jim: Right? So when, when you were watching that demo or when you watch the demos, what, what, what, what do you see, like, as a robotic expert?

[13:51] Tom Lipinski: I I would, I would be looking at, you know, how well, you know, teleoperation is running. What's the sort of latency? Um, one of my, one of my, um, issues is the ergonomics of teleoperation. At the moment, teleoperating humanoids requires VR headsets, and which are famously nauseating, for example, and you might not be able to

[14:11] Tom Lipinski: teleoperate the robot for more than half an hour before you, you get, you know, disoriented or, or sick and so on. The technology is not yet there to build the entire infras- infrastructure behind the sort of the teleoperated deployment. So that's challenge number one. Ergonomics is, is challenge number two, the effectiveness of teleoperation. When you, when you're teleoperating with a VR headset and, and, and, you know, if you're not wearing an exoskeleton, for example, because if you're wearing an exoskeleton, your, your joint movements can be mimicked by the, the set of slave arms, you know, on a humanoid. But if you're not, then, then obviously there's a joint matching, mismatching. There's this what's called, uh, cross-embodiment training challenge and so on, um, where you might be-- and the teleoperator might be trying to do something, but the robot might not actually get there because of the, the physical constraint of the, of the... And, and so on. So there are lo- lots and lots of challenges. And even when, when you look at those demos, if you

[15:11] Tom Lipinski: think, "Oh, it's working because it's teleoperated," well, actually that's still a lucky one when it's all working W-w-when it's teleoperated. You still have quite a lot of challenges having that robust teleoperation infrastructure in place, the low latency, you know, uh, tele op setup and so on. It's, it's challenging to even teleoperate humanoids and, and forget about autonomy. Um, those demos are Well, I don't wanna get sued, but mostly smoke and mirrors

[15:42] Mian Khubaib Jim: Okay, then, on that note, Figures AI boss went to, uh, went on television and said there is no teleoperation. What would you tell him?

[15:53] Tom Lipinski: Okay, I'm gonna have to be careful here, right? But let's You won't get that statement from anybody who is a roboticist, from anybody who is an engineer, right? If you ask anybody who's been doing robotics for some time

[16:10] Tom Lipinski: and who understands how how robots work, right? They'll be a lot more cautious with statements they're making.

[16:20] Mian Khubaib Jim: Yeah.

[16:20] Tom Lipinski: Um, Brett is a marketing guy. Yeah. His entire life he's been selling things, and he's really good at this, right? He's succeeded in raising money for Figure 01 and so on. But there is more to, there's more to operation than, than teleoperation, right? You've got motion planning. Yeah. And, and, and robots can can follow sequences of motions without actually being teleoperated because they're literally following a sequence, you know, pre- pre-planned sequences of motions. Um, when it comes to autonomy, um, you know, like walking, gait, you know, maintaining stability and so on, well, that's, that's, that's edge computing working at its hardest, and that's still an extremely hard challenge. You could see that everyone's taken quite a few years to get to that point. Now they got to that point, um, great. The rest, you know,

[17:09] Tom Lipinski: Helix, which is what Figure 08 claims to be their model, which is, you know, zero shot, claims that, you know, oh, a robot knew what to do with ketchup even though ketchup was not in their pre-training. It's like you're you're training the robot on a kitchen data set, you know, that contained in pre-training, you know, millions of hours of kitchen data set. Maybe ketchup wasn't in the fine-tuning process, but it kind of had to be there in order for the data set to be large enough to do any reasonable training. So, you know, if you could, if you could train robots on, on half an hour worth of data, you know, you'd see robots everywhere around us doing things, right? Mm. The fact is we can't. So Even if if something's not teleoperated, it doesn't mean it's autonomous.

[17:56] Mian Khubaib Jim: Okay. Now, so where do you draw the line? Like, Q-Bots, uh, were also teleoperated. There, there is teleoperation as a tool versus a trick to sell the product. Where do you draw the line?

[18:07] Tom Lipinski: Okay. Um,

[18:08] Tom Lipinski: first of all, let's, let's start with the fact that I'm, I'm, I'm a founder of Q-Bot, but I no longer work at Q-Bot. I know. I left the company 18 months ago.

[18:14] Mian Khubaib Jim: We'll discuss

[18:14] Tom Lipinski: it later. Um, when it comes to, to that particular deployment,

[18:20] Mian Khubaib Jim: right?

[18:20] Tom Lipinski: right, Y- we combined autonomy with teleoperation. Yeah. We've obviously done everything through teleoperation to start with, then we introduced some motion planning, and then we started building autonomy when it comes to spray application. You still had to... E-every robot is teleoperated as it navigates, you know, the crawl space or the, the, the floor void, mainly because no two floor voids are the same, right? You've got a spaghetti of cables and pipes, and you've got rubble and all the other things. You have to-- You need a human to operate, to navigate around obstacles, to make sure that the robot doesn't, um, tie itself in knots because you're pulling quite a big umbilical that supplies the foam and so on. So it requires a certain operator skill and, and, and knowledge, and it also requires a lot of intervention. But then once you've locked in onto an area that you want to insulate, based

[19:07] Tom Lipinski: on on the data that's been collected, the robot can do that either autonomously or semi-autonomously or execute motion planning, um, which again, is not autonomy, but it's not a direct intervention by the teleoperator, right? So you've got, you've got a mixture of things that you can put together in order to deliver the outcome, um, to the customer. Um, but it's-- you can't do it without teleoperation, right? There's just not enough data to train successful navigation in extremely constrained and messy environments.

[19:35] Mian Khubaib Jim: So you mentioned about numbers like Figures, Figure AI is at 39 billion and other big companies, like almost every single humanoid company is over $10 billion pound, uh, $10 billion valuation. And, uh, do you think there is a reckoning coming uh, that, uh, that will pop this bubble? Like K Lab is gone, and then this Cartwheel Robotics, that's also out of business

[19:59] Tom Lipinski: Based on

[19:59] Mian Khubaib Jim: on the thesis that we have just done- Yes ... the data

[19:59] Mian Khubaib Jim: part.

[19:59] Tom Lipinski: So look, there are some humanoid companies that are quite clear that they're only making hardware and, you know, if you want to deploy those to do something, you need to do it yourself. So Unitree Well, Unitree is is a complex case because they make hardware, but also, but it's, it's a spy hardware, right? They deploy literally, you know, CCCP agents across the world, and your job is to put them in, in critical areas where they could spy and snoop on you. Um, but the fact is that Unitree is quite clear that they're-- they're just supplying you with hardware, right? And putting that into use, you know, w-whether it's teleoperation or, or model training, whatever, is somebody else's job, right? So I would say apart from the data collection side of things, which, you know, one could say it's a little bit untoward, I can see that there is a reasonable amount of honesty, you know, this is just hardware. You go and put it to use and work it out somehow. You've got companies like Agibot that will say, "Oh, we own the full stack. We're creating all this, you know, simulation

[20:58] Tom Lipinski: layer and training layer and everything else, and you can do all sorts of wonderful things with it." And then you go to s- you know, um, A-Agibot launched in London last month. You go to see demos and, you know, you watch it all fall apart. Um, and the only thing that works is robot dancing, and it's like it's one of those- you know, six months ago it was folding clothes and towels, and now it's dancing, you know, and I wonder what's gonna be in six months' time, you know, maybe robots singing or whatever. But none of these things are useful. Uh, when you-- when we looked at demos and, you know, this is like two days of demos when you're doing a product launch, right? You're gonna put your best foot forward. A robot that was supposed to be kicking ball into, into a net, you know, three meters away, um, kind of went on strike and folded itself into bits. Um, looked like somebody crying under the shower. And then robots that were supposedly mimicking production environments, uh, one was very heavily teleoperated because you could see that in uh, shaking arms that, you know. it's a teleop setup and it's quite a

[21:58] Tom Lipinski: rudimentary one. And the other one that was, I think, executing some motion planning kept bumping into different conveyor belts that were set up there and everything else, and kept moving things around because, you know, again, it was executing motion planning rather than being, you know, running perception and, you know, on, on, on edge inference and so on. So it was a l- I wouldn't say it was extremely disappointing because I didn't have high expectations, but it demonstrated that we can, we can put actuators and arms and bits of metal together, but we still can't make them do things.

[22:31] Mian Khubaib Jim: okay, we'll-- I will come back to this, but you mentioned Unitree. Unitree has made their, this hardware thing very cheap, like almost seventy percent cheap, and their margin is sixty percent up. Do you think West, West will be able to catch them up, or it's too late on the hardware part?

[22:45] Tom Lipinski: There are are various theories out there, and I had a few conversations with, with, um, with people about it. Some people would say

[22:56] Tom Lipinski: That It's all about the energy cost. That all the other costs can be matched, so whoever has the cheapest energy... Because you can, in theory, and it's a philosophical concept rather than a practical observation, in theory, you can make humanoids build humanoids. Yeah? I mean, far-fetched, but that's what a lot of people bet on, that we can have robots building robots, therefore labor cost don't matter, and the only cal- cost that matters is the energy cost. So whoever can drive the cost of energy to the lowest point will have the com- you know, the, the, the sort of economic advantage in, in building robots. At the same time I believe it's a very simplistic argument. If that was the case, then, you know, every country could make photovoltaics, and yet all the photovoltaic factories are in China, right? If that was the case, we could, we could... because photovoltaics are generally fully automated, you know,

[23:56] Tom Lipinski: gigafactories, right? So it's not as simple as that. Um, if you look at history, every technology wave followed a very similar trajectory, like personal computers were introduced and everybody in US was building PCs. You know, IBM was famously talking about all the money being, you know, being made in hardware. And you had all sorts of companies like, you know, Compaq and, and lots and lots, hundreds, thousands of companies set up in US to build personal computers. Um, very few of them remain, and the ones that do remain make computers in China, right? So the- once the Chinese have solved the manufacturing challenge, there was an extinction event in this space. And then you, you know, move a little bit forward and you can see internet. You know, everybody wanted to build the internet. Um, you had companies like Nortel, Alcatel, you know, Marconi and everybody else, you know, AT&T talking about building this internet infrastructure and,

[24:55] Tom Lipinski: and all the hardware behind the internet, right? There was a huge wave of like, we're gonna build hardware that runs the internet. We're gonna milk the internet, right? And then, you know, two trillion dollar crater later, yeah, none of them are around. Um, AT&T sort of is around because they're a monopoly, it's different. But companies that rushed into like, "Oh, we'll make stuff," yeah, ended up extinct because the Chinese have solved it. Same happens to photovoltaics in between twenty eleven and twenty thirteen, where Chinese have driven the price down ninety percent. I mean, we all benefited, but there's no manufacturing left in Europe and, and, and America. Um, similar thing is happening with cars.

[25:33] Mian Khubaib Jim: EVs?

[25:34] Tom Lipinski: Yes. I mean, Chinese make the cheapest... I mean, th-this is, this is the challenge. EVs are actually simple when you look at them, right? Mm-hmm. It's a couple of mo- couple of motors, batteries, and the chassis around it and a set of wheels. Um, they're absolutely slaughtering existing manufacturers. And if you think about automotive, it's highly automated, right? Mm-hmm. So how come they have such a large

[25:55] Tom Lipinski: advantage? So I, I think that the claim that the West can win in manufacturing humanoids because humanoids can make humanoids is unsubstantiated. And I, I, I highly doubt if, if that will work out. The way I see this is a similar sort of cycle repeating itself as, you know, one of, you know, as, as the three or four cycles before, where Chinese have solved humanoid manufacturing at cost and to the right quality. Having said that, they can make those things. They're still like computers without software, right? So somebody will have to put them to use.

[26:40] Mian Khubaib Jim: Now, why does this keep happening? And, and do you think is this the subsidies that are being given to these manufacturer in China or maybe West can win on the trust side, maybe? Trust and the brain, the AI, LLM, edge computing, and all that stuff

[26:54] Mian Khubaib Jim: on the software side

[26:54] Tom Lipinski: China is allocating 200 billion to scaling robotics, right? Um, and that goes right across the industry. And to be honest with you, I think that 200 billion figure is actually conservative because there's several, several actions are being implemented simultaneously. So first of all, there's a huge amount of assistance going into all the robotics companies, including humanoids, right? State subsidies and so on. At the same time... And, and it's not just financial. I mean, Unitree literally was given a factory by the, the Chinese state. Subsidies, yeah. "Here you go. Yeah, there's your factory. You just move it and make these things." So you've got, you've got, you've got subsidies, cash subsidies, but also lots of other assistance and so on. Uh, in addition to that, the financial ecosystem is being very heavily primed. So if you want to buy a humanoid, uh,

[27:52] Tom Lipinski: you're going to get a zero interest loan from the Chinese state immediately to procure them, right? So, so procurement of humanoids is subsidized, not just production of those. So, um, and it happens right across the economy. Um, exports are subsidized and so on and so on. So it's not just subsidies. It's, it's, it's pumping money through the entire ecosystem. Mm-hmm. Even though quite a few factories that that have tried humanoids have admitted that they're not particularly useful, Um, they don't do much, and when they do, they are a are a lot slower than, than humans and you're talking about, you know, moving crates around, which you could move with AMRs without a problem. So kind of putting them to, to the wrong use case as well. Um, but what you will have is this huge ecosystem, right? That's, that's been, that's been given momentum, right? By the state subsidy, just like, you know, production of silicone in,

[28:52] Tom Lipinski: in, in US was subsidized by the state, you know. The US, US US state subsidized quite a lot of technologies before, you know, before recently, and that's one of the reasons why they've become so successful. You, you get to a certain economy of scale, and at that point, you, you- unit economics make sense. People start buying and, and, and you get on with it. Um- And I I think that will continue. That China has enough money to throw at things, throw at problems. Whether humanoids solve any problems, that's a, that's a completely different question.

[29:18] Mian Khubaib Jim: No. okay, what, what do you think? What would be the valuation of this, this industry by twenty forty or something? Tr-

[29:27] Tom Lipinski: Tr--

[29:27] Mian Khubaib Jim: forty trillions, fifty trillion, hundred trillions. What would be the valuation?

[29:28] Tom Lipinski: I think we'll see, um, we'll see a bust before we see the boom, right? So right now we've got a hype cycle. We're probably somewhere towards or close to the top of the hype cycle, yeah? Which is y- the usual trajectory where, you know, there's new technology, there's this early promise, there's the huge hype. At

[29:46] Tom Lipinski: this point, all the VCs start pumping money into, into the business, and then there is this disillusionment, you know, and, um, valuations crash. You get a, a fair amount of extinction within the industry, and then somewhere towards the race to the bottom, a few use cases emerge, which I, I can't pinpoint at the moment, but there might be a few use cases that emerge, and they sort of stabilize the valuation. So the- I I think there'll be obviously more companies entering the market because, you know, when you, when you're pumping money into the market and with VCs you're parking-- pump-pumping money into the market, lots more companies get created to take advantage of it. Um, then there will be some sort of mass extinction event, um, driven by disillusionment. Uh, and then at some point after dust settles, we'll see key use cases emerge. Very unlikely bipedal, you know. I'm sure there'll be wheeled

[30:45] Tom Lipinski: humanoid-like robots, but they might not have heads, they might not, you know. So we might, we might broaden the definition, right? We might say, "Oh, that's also a humanoid. I know it's got wheels, and it's just got two arms on it, and nothing looks like a human. But because it's designed to do what human does, we're going to call that a humanoid." If you wanna bro-- if you broaden the definition, I will see, yes, there's huge, huge value in the market, right? There will be, I I wouldn't say forty trillion, but there will be, there will be, you know, trillion-plus market in, in, in, in deployment of robots to useful, useful tasks across various industries. But they are very unlikely to resemble humanoids that currently are getting all the investment.

[31:28] Mian Khubaib Jim: And how much China will tap on? Like, what percentage China will take?

[31:32] Tom Lipinski: I think China will produce vast majority of robots. Eighty

[31:37] Mian Khubaib Jim: percent?

[31:37] Tom Lipinski: I would say sixty to eighty percent, yes.

[31:40] Tom Lipinski: Yeah.

[31:40] Mian Khubaib Jim: Okay. So, uh, Britain, uh, this advanced research agency, they just gave fifty million pounds of, uh, subsidiary or investment for robot, uh, robot dexterity- Mm-hmm ... to solve this hand problem. As you mentioned, seventeen thousand receptors and, and, uh, our brain take, uh, less than a millisecond even to just to to make sure how much grip we need and all that stuff. Do you think that would give us an edge?

[32:00] Tom Lipinski: I think there is-

[32:00] Mian Khubaib Jim: Like whoever solved that dexterity first-

[32:02] Tom Lipinski: would

[32:02] Mian Khubaib Jim: have an edge

[32:03] Tom Lipinski: Look, Look,

[32:03] Mian Khubaib Jim: Look, I,

[32:03] Tom Lipinski: I'm not sure if there is one solving dexterity event. I think you, you might have several advances that have not yet occurred, um, that can help specialized robots, grippers, hands,

[32:14] Tom Lipinski: whatever, handle different objects and so on. You can have, for example, um, there is a space for soft robotics. Yeah. You've seen things like, uh, for example, an octopus-like, uh, grippers where they wrap around objects and so on. Right. So there's a... We'll see, we'll see emergence of, of various approaches to dexterity. So I don't think, I, I might be wrong, but I don't think we'll have an event where we can say there's a human hand and there's a robot hand, and they match one for one, and they-- the the the robot hand can do everything that the human hand does. We'll probably see various different types of grippers, you know, from soft to, to, to, to liquid-filled, to, to, to, to grippers that actually are responsive there. You know, they don't need any inference to, to sort of collect data and, you know, that, that might automatically react to

[33:14] Tom Lipinski: deflection in different ways. You've already seen, I mean, a lot of grippers on the, on the market, if you, if you sort of press them at the midpoint, will automat- will wrap around just because of the way they was structured, for example. So I think we'll see a lot happening on, on the gripper side, but I don't believe we'll-- it, it will be like, "Oh, look at it. This is the hand, and it does what human hand does." It will be a various array of, um, of different, um, technologies that can be applied to different applications.

[33:44] Mian Khubaib Jim: By the way, there is a perception out there, like, uh, and the, the people think that, like, the way GPT solved some use cases really well, like the content creation or the generation or all that stuff, there will-- a, a model will come out, and it will solve all these problem for robotics. You know, a bot will be able to take the decisions really fast because of the inference or the better model or something.

[34:06] Tom Lipinski: And

[34:06] Mian Khubaib Jim: we have been talking about data gap. What do you think? think?

[34:08] Tom Lipinski: Um, this is not a

[34:10] Tom Lipinski: linear progression that we're talking about here. Um, let's just take a step back and look at a large language models and how neural networks work in general. right? Um, when ChatGPT started being useful, it was when it was trained on, you know, vast amounts of data. Yeah. we're talking about the whole internet of data. And text is one-dimensional. Yeah. So we're talking about neural network that's been trained on vast and vast amounts of one-dimensional data in order to generate the next word and, and do it well. Okay? This is one-dimensional data. Now, when we talk about physical environment, not only are we talking about the, the three-dimensional data with with things like distances and, you know, and, and, and so on, with physics. Yeah. Where we have

[35:11] Tom Lipinski: motion and momentum, gravity and, and so on, With, uh, all the other sensory inputs. So one is, one is your- Yeah your sort of tactile, you know, um, uh, sensory feedback, but there, there's, you know, smells. There is, you know, different, you know, shades and so on. We're talking about an environment which is at least five orders of magnitude more complex. Yeah. 'Cause it's not only three dimensions in physics, but, you know, we haven't worked out how to convey, how to convert, you know, touch into, into, into digits yet. So at least five orders of magnitude more complex, right? And obviously, as you grow in the complexity of data, you sort of have to square the complexity of of training and, and the compute demand. So when it comes to arriving at the sort of ChatGPT moment for, for physical AI, like, like what Jensen Huang says, and Obviously, he would say that because

[36:10] Tom Lipinski: once we've arrived at, and if NVIDIA is selling chips to do that, they're probably worth, you know, ten quadrillions or something like that. Because the amount of computational power Is so vast that, A, we don't have space on Earth to do it with current models. B, we don't have the energy to, to run it using current models. Maybe quantum computing could help when it gets commercialized at scale. But the chat-- that, that, that developing that sort of ChatGPT level autonomy for physical AI, it's, it's so vastly more challenging than, than, than doing a decent neural network that can, you know, do your coding or whatever because it's been trained on internet data, that I just can't see us delivering that as humanity over the next ten, 15 years.

[37:04] Mian Khubaib Jim: You know, on that point, NVIDIA claims that they have, uh, you know, trained,

[37:04] Mian Khubaib Jim: nine months of whatever demonstration just in nine hours or ten or 20 hours with simulation type of thing, with synthetic data or whatever. What, what would you say about that? Do you think synthetic data or these synthetic demonstrations can help to train these? The,

[37:08] Tom Lipinski: The, yeah, the, The, synthetic data was the last year's hype, right? You know, when, when people thought that, "Oh, we can, we can, we can achieve all of that with synthetic data," and all we ended up with was the sim to real gap, You know, where things work in simulation, yeah, you know, trained on synthetic data, And then when you, when you try to, you know, when you upload the model onto e-even a set of arms, right, forget a robot that actually moves around and does something. It, it breaks down nine out of 10 cases, right? And still in ve-very controlled environments. We're talking about, you know, training a robot on synthetic data, two arms fixed to the desk, right? And you do-- you, you run through millions of hours of training. It works in the simulation every time. You upload it onto a robot, and even in

[37:51] Tom Lipinski: that very controlled environment of a small desk with few objects that it have to handle, you still break down nine out of 10 times, right? And that's ex- that's the environment that doesn't change, right? So So the synthetic data gave us a, a sim-to-real gap. Yeah, the fact-- the, this year's hype is, is paying people in third world countries to wear, you know, iPhone cameras or GoPros on their foreheads to record everything that they do. So that's the, that's the hype of 2026 and this assumption that that will solve everything. And I've seen company-- we-- you've-- we've seen companies raise, what, 50 mil uh, fundraisers, was it this week or last week, on the back of we collecting data from people that just go around their daily life wearing a camera, and then the promise from that we will build the data set to train robots, right? Um, that will fail. Um, and then the next year there'll be another hype, you know, are we gonna build it in, in so and so way? The fact is that people underestimate the

[38:51] Tom Lipinski: size of the challenge. Um, most, you know... And and it's, it's easy, you know. You read a paper at university and go like, "Oh, I'm-- I've got a brainwave. We're gonna solve this," you know? And people that have been doing sort of machine learning for years look at this and go like, "Hmm, Clearly you've never done it before, have you?" Um, none of these things will succeed, right? Robotics generally is, you know, is, is for, for, for, for, for masochistic people that wake up to, to fail repeatedly every single day and on occasion succeed. It's not something-- it's not a space where you can simultaneously solve everything and then have like zero shot, fully generalizable model purely because we would need to, you know, with current models, you know, and with current approaches, we'd probably need, uh, you know, an, an energy of the entire galaxy to run that, you know? So it's it's just not, it's not feasible.

[39:47] Mian Khubaib Jim: On that point, by the way, Stanford and NVIDIA, they had this

[39:49] Mian Khubaib Jim: collaboration where they gave five minutes of muscle memory and uh, three orders of magnitude more to, to a robot And, uh, I don't know, like, again, do you think that, that, that would make it faster

[39:58] Tom Lipinski: You could move technology forward in, in, in narrow specialized fields, right? For example, I think there is-- I think there are-- there, there is a lot to be achieved when you, when you're looking at grasping by actually creating, creating, by putting together, together grasping technologies that can react on, you know, immediately based on touch or based on sense without actually inference, for example. And, and can, you know, you touch something and it grabs it and, and so on. Um, there and there are lots of people researching, you know, fluid dynamics models trying to achieve that with soft robotics, et cetera. Um, there will be advances, but, but the field of-- what we're trying to

[40:41] Tom Lipinski: do To, to, to go back to first principles, we're trying to recreate a human being, yeah, that has evolved over millions of years through adaptation and, and so on, and and for all intents and purposes is more or less perfect, right? I can't think how it could be improved and so on. And we're trying to replicate that with, okay, billion-dollar budgets. That's nothing,

[41:09] Mian Khubaib Jim: yeah?

[41:09] Tom Lipinski: Over years. That's nowhere near close enough. So some companies will manage to solve a small proportion of the challenge somewhere, but it's not something that can actually just, like, happen from zero to one. It will take decades. And by the time we get there, it won't look like what we expect it to be. It will be several different strands specialized around different compute requirements, different task groups, you know, different environment, conditions. And, And, at that point,

[41:40] Tom Lipinski: we'll, we'll, we'll arrive at something that's useful in some cases.

[41:43] Mian Khubaib Jim: Hmm. Okay. On, on the data part, I think which humanoid company have the edge? Tesla maybe they can win when it comes to data collection and training. They're the most capable, maybe. No

[41:50] Tom Lipinski: I don't want to be badmouthing anybody. Um I I need to ponder So last year at Tesla AGM, it was January AGM, uh, Elon

[42:02] Mian Khubaib Jim: was January AGM, uh, Elon

[42:05] Tom Lipinski: Musk was asked about, you know, the challenges of of solving, you know, humanoids and so on. And obviously a lot of their valuation started riding on, on, on, on humanoids and so on.

[42:16] Mian Khubaib Jim: on. Um, and what Elon

[42:16] Tom Lipinski: and what Elon Musk said was making humanoids useful is a trillion-dollar challenge. Yeah. Or half a trillion dollar challenge.

[42:27] Tom Lipinski: It's in, it's in a, um... it's also in, in the same, in s- in, in the same AGM where he said that Tesla was gonna be a $10 trillion company, and everybody sort of latched on to that. But what I took out of it, it was that he expected to spend minimum five hundred billion, five hun- bi- with a B, right? Trying to train humanoids to be remotely useful at things. And he did say that it's like, you know, a thousand times more complex challenge than getting cars to drive, right? So he did say some sensible things. At the same time, I would still take that, I w- I would take that $500 billion cost as coming from the guy who said that we were gonna have Roadster, you know, four years ago and, and took deposits from people for that Roadster, which is still nowhere to be seen. So he's still extremely optimistic, right? And yet he puts a number at half a trillion challenge for his company to actually try to make them useful,

[43:27] Tom Lipinski: right? So- When Elon Musk says that it's gonna cost him at least half a trillion- you know, I would probably, you know, add a zero and say he's not wrong. Um, he's been not wrong on several fronts before. When you've got people saying, you know, "We'll solve-- We'll build this large foundational model that can generalize across all the human in- human tasks and so on, and we only need sort of 400 million for that," they are out by four orders of magnitude.

[44:04] Mian Khubaib Jim: Okay, what do you think about OneX hand? It went viral, 10 million views. Mm-hmm. And their, their claim was it's near human-level dexterity. And, uh, iRobot CEO Brooks said, "Mm, it's pure fantasies."

[44:18] Tom Lipinski: Well, first of all, I, I never shook that hand. So I don't know. You know, I've seen a, I've seen a video showing it that it can lift a hell of a lot. You know, payload is a big

[44:27] Tom Lipinski: thing. Uh, and obviously I've seen all the sort of dexterity, which is claims to to be beyond human because it can bend backwards and so on. So it's, it's a great piece of engineering. Um, can it last more than a week? That's... you know, if you think about something that had to become so complex, right, in order to to achieve something resembling, you know, human operation I'm, I'd, I'd be looking at it as a roboticist thinking, "I wonder what mean time between failure is," you know, and so on. But the fact is, I don't know if it works. I've seen the demo, you know, it looks, it looks great. I don't want to be going around and saying, "Oh, it doesn't work," because again, I don't know if it works, I don't know if it doesn't work. I can't say. It looks very complex. and I don't know if it's-- if it can survive more than a week in normal environment. I mean, the beauty about

[45:27] Tom Lipinski: human body is that we heal, you know? I don't know if that hand, you know, can lift 10 kilograms more than twice, for example. I just don't know. Yeah.

[45:37] Mian Khubaib Jim: Nvi-Nvidia just launched this Cosmo 3 Edge. It's a $250 board and, uh, two billion parameters for reasoning. Do you think that would help to make humanoids better? I, I

[45:50] Tom Lipinski: I don't think that that, that will... I- if you put a-- if you build a humanoid with that board on, you know, with that on board, that will, uh, work, make them better. You need a lot more compute on that front. But what it could do is help to stimulate innovation, which we need to even start tackling the physical AI autonomy challenge, right? We, we're talking about the challenge so vast that we, we don't even know at this stage

[46:18] Tom Lipinski: how we can begin to tackle that. So with, with that sort of cheap, cost-effective compute and those models that can be used by anybody, students and college kids and so on, we're making it easier for people to innovate and come up with ideas and hopefully build

[46:36] Mian Khubaib Jim: more

[46:36] Tom Lipinski: efficient models or different models that could-- that can start helping us to tackle that challenge.

[46:40] Mian Khubaib Jim: Okay. What-- everybody wants to buy the finished work, and, uh, we, we have seen software as a service. Do you see robot as a service? And, uh, if yes, what would be the commercial model? I

[46:52] Tom Lipinski: mean, I've been running robot as a service for over 10 years, right? And, you know, at Q-Bot, we've provided our partners with fully functioning robotic solution that contain not just hardware, but also, you know, software, TeleOps controls, uh, fleet management. And, and they were delivering outcomes to customers, charging a few thousand pounds, and we were taking, like,

[47:14] Tom Lipinski: 15, maybe 20% of it. So, um , in my view, that is the only way to commercialize robotics, right? Is to, is to have something like software as a service or even work as a service. So there, there is Sequoia recently, or Julian Beck from Sequoia, uh, published his view on On the next trillion-dollar company formations where he stated that the next generation of companies reaching trillion dollars will, will not be selling hardware or software. They will be, um, they'll be selling work. Mm-hmm. Yeah. And, and if you think about what you, what you said earlier about, you know, agents being able to do work, um, we'll see that service offering evolve to robots being able to do the work. And at that point, you're not-- it's not robots as a service. You're not

[48:14] Tom Lipinski: providing somebody with a robot and charging a service fee. You're actually providing somebody with an outcome. Yeah, so, so I see this as a model that, that will be extremely successful, um, where... And it's, if you look back, it, it's sort of how it all repeats itself o-over every, um, technology cycle. When, when we had personal computers, um, coming into, into a, into the market first, people focused on making them. But actually, the ones that enabled people to generate outcomes from them, you know, developing spreadsheets or word processing or, you know, graphic design and so on, were the ones that captured the most value, right? And, and I think with humanoids or with robots in general, you'll see a similar model emerge where

[49:03] Mian Khubaib Jim: will

[49:03] Tom Lipinski: hardware will commoditize because the Chinese will do at such a cost that, um, you know, unless you specialize in extremely, you, you'll have to just, you know, do it

[49:13] Tom Lipinski: at cost. Hardware would commoditize. I think models will commoditize as well. You know, every month you've got a new model coming out and so on. And I know people are putting a lot of money into models, but I think they will, uh, that they will commoditize. What-- What-- Where you will see the value being created is companies

[49:33] Mian Khubaib Jim: putting

[49:34] Tom Lipinski: all of that hardware and models and, and everything else and, and the infrastructure behind it to deliver value to customers or outcomes to customers. So customers won't be buying a robot or won't be buying a service. They'll just be buying work, right? So, so if you see... It's almost like an employment agency, right? You know, there, there'll be this model where Companies will evolve into managing fleets of robots, thousands of them, you know, to put them to particular tasks and literally charge like, you know, a

[50:06] Mian Khubaib Jim: human

[50:07] Tom Lipinski: salary, a fraction of human salary. I don't know, you know. You'll see that arrive at some sort of point. But it would be-- it, it will

[50:13] Tom Lipinski: be a fully managed solution, right? You know, it's not like you're buying a robot and then you're looking at guarantee what do I do if an arm breaks or what if I do if the software update fails? What about cybersecurity? How do I deal with that? There is so much around robotic deployment, you know, regulatory, safety, um, uh, cybersecurity, you know, and keeping it updated for the... If you, if you're looking at production, right, people change production sometimes every shift, yeah. There are new products coming out. So it's not like you can buy a robot, put it on a m- production line and go like, "Yeah, you do it." And so because you, you need to continuously adapt, it, it, it, it won't work as, as a capital expenditure. It can only work as a fully managed service where a startup business, a company delivers that service to a, to a manufacturing floor, but they're literally taking on the labor supply side of things, right? So they'll be like a, a labor force supplier that happens to use

[51:13] Tom Lipinski: robots in order to get the job done, right? But at the same time, manages everything from hardware, to models, to training, to telop support, to maintenance, to, you know, replacing batteries every now and again because they die over time, and so on and so on. So once you see those specialized companies emerge, I think we'll see a much, much, much quicker commercialization. But the problem with that, obviously, is that if you do that, you're taking on all the risk. At the moment, within the market, this, this, you know, everybody works in sort of consultative approach and, and so on. It's like, "Oh, you just, you know, we'll provide you with this autonomy service or whatever, but now you own this and you take the risk." Um, I think once, once new businesses or new entities have solved how to take on all of that risk, manage it, insure it, ensure the compliance, you know, and, and safety and everything else, and just deliver the outcome that people need, we'll see an explosion in, in adoption.

[52:10] Mian Khubaib Jim: On agents, on, on agents part, agents do take

[52:10] Mian Khubaib Jim: jobs from humans. And, uh, do you think robots will be taking jobs from humans or they are just filling the jobs that people don't want in the first place? I

[52:10] Tom Lipinski: I I think, I think you need to consider the scale, right? Um, I mean, you've got companies like Artisan, for example, in US, you know, famously advertising, you know, "Don't hire humans, hire agents," and so on. Um- Mm. Um, now when it comes to human jobs, let's say, let's say look at, you know, what I focus on. It, it's manual repetitive tasks in manufacturing, right? If we say we're looking at a factory that employs 800,000 people, so it's a, it's a reasonable medium-sized factory, right? And you look at, okay, well- We can automate these tasks and maybe deploy twenty, 30 robots within a year, right? To do those tasks and, you know, ideally one task, so you don't actually

[52:10] Tom Lipinski: multiply the training. Well, that factory will lose around 200 people per year, and will have to rehire about 200 people per year. Yeah, it just happens, right? It's the normal turnover. Some, some factories are 30% turnover per year, right? So if it's 1,000 people, they will lose 300 people, you know, each and every year, and they will have to go out and search, recruit, you know, train, onboard, vet, all, do all those things, and then people leave still, right? So if, if an average factory loses two to 300 people, uh, per year, there is a room for two to 300 robots to be integrated into the factory without a single job loss. Yeah? And then with people moving to tasks that are a lot more dexterous, that require a lot more improvisation and so on, I've, I've, I've gone around factories where people are keen to automate. You know, we've gone around the whole factory and we said, "There's not one job that we could tackle here." Right?

[52:10] Tom Lipinski: Because they're extremely dexterous. They-- Especially where, where people operate like wiring, you know, consumer units for power stations. Right? Takes 110 hours for consumer units, right? For, for big fuse, uh, for, for like fuse boards, right?

[52:10] Mian Khubaib Jim: People

[52:11] Tom Lipinski: would love to automate that, but you're actually working with tips of your fingers for most of the time. You don't even see what you're doing, right? Where you're putting, pulling wires behind the bus bars and, and so on and, you know, putting them into those, you know, different devices like circuit breakers and so on. You can't possibly, you know, deploy a robot to do that, whether it was dexterous or grippers or whatever. You just... It's too hard, right? So we've looked at lots of jobs where we're just like, "Can't be done. Not, not in the next five years, maybe not in the next ten." Yeah? And probably beyond that. So what you'll see is the sort of the, the boring, mundane jobs being automated with robots, yeah? Whether it's humanoids or mobile

[52:52] Tom Lipinski: manipulators or whatever. The ones that you kind of, you don't wanna do in the first place. And, and to be honest with you, I, I've got-- The way I see this is

[53:02] Mian Khubaib Jim: The

[53:02] Tom Lipinski: The fact that you have a shitty job is keeping you in a shitty job. You know why? Your boss will not let you progress, you know, do some training, move on to a better job because they don't have anybody to do your shitty job.

[53:18] Tom Lipinski: Yeah? So it's actually you being inversely oppressed, right? Because of the shittiness of your job and nobody wanting to replace you when you move on, you are literally constrained to that one job, right? The moment we can automate those repetitive shitty jobs, right, we're actually freeing up people to move up the food chain, so to speak, get some training, you know, do something more demanding, earn more money, right, have a more rewarding career, right? But at the moment, because we've got all of those

[53:50] Tom Lipinski: repetitive shitty jobs that need doing and we keep putting people to do it, we kind of have to keep them there, which, which is literally like, like an inverse oppression.

[53:59] Mian Khubaib Jim: Hmm. Okay. Do you think is there any physical job which, which is safe for, let's say, for next 20 years and, uh, any job which will be gone less in three or, or five years? I

[54:08] Tom Lipinski: I I mean, if you consider driving a physical job- Mm-hmm would obviously with- That is

[54:14] Mian Khubaib Jim: that w- obviously with- That is a physical

[54:14] Tom Lipinski: job. Yeah. Yes. Th- that's-- That will be mostly automated over the next 10 to 20 years. I mean, we can see that happening now, right? It's, it's all... I mean, it'll take time. It needs a lot of data, and we'll have all sorts of issues that will set back or, you know, taxi autonomy, you know, and, and other driver autonomy. But-

[54:28] Mian Khubaib Jim: but

[54:28] Tom Lipinski: I mean, we've automa-- I mean, y-you look at DLR, you know, that's been... You know, trains have been running autonomously for years, right? Tube, not so much. Not because there isn't a technology to do it,

[54:38] Tom Lipinski: but there is a human resistance, right? So moving things around, cars, vans, trucks, you know, trolleys, forklifts, taxis, trains, and so on, will be automated to a much larger extent as-- than it is now. So it's definitely that, right? If you're, if you're an Uber driver, you know, learn a new skill.

[54:58] Mian Khubaib Jim: Hmm, okay. By the way, const-- when, uh, uh, as we are talking about physical jobs, construction is the hardest one, right? This is where you went. And why we are not solving these type of hard jobs first rather than, you know, laundry folding? I-- What's the point of

[55:14] Tom Lipinski: you

[55:15] Mian Khubaib Jim: know-

[55:16] Tom Lipinski: Well,

[55:16] Mian Khubaib Jim: it's a

[55:17] Tom Lipinski: a lot easier to, to run a laundry folding demo than to run a, you know, plasterboard fixing demo, right? If you, if you think about laundry folding is... I, I don't wanna take away from achievements of laundry folding, but we're looking at a soft

[55:36] Tom Lipinski: deformable item placed on a hard surface that you can grasp in several places. Yeah. And as you move it, it moves with you, right? And you fold it, it's compliant, right? So there are several ways to succeed- Mm-hmm ... folding laundry, right? When it comes to hard jobs like construction, you're not working on the tabletop in a controlled environment with something which is very compliant and will just follow your movement, right? You're on construction environments which are messy, um, which are demanding, uneven, with different lighting conditions, and so on. And on top of that, there's quite often only one way to successfully execute the task, right? If you're fixing a plasterboard to the wall, you if you're three, four millimeters off, off, off target, you're off target. You failed, right? So, um, that's number one, right? It's, it's, it's a, it's, it's much easier to run a demo folding towels or clothes or, or whatever. Number two is

[56:38] Tom Lipinski: construction tasks are extremely hard. Yeah.

[56:41] Mian Khubaib Jim: They are indeed.

[56:43] Tom Lipinski: And, and then

[56:43] Mian Khubaib Jim: you,

[56:43] Tom Lipinski: you can see, you know, I can, I can pick up an impact driver and without looking at it, I can adjust it to sit in my hand better. Yeah. If I pick it up s- off center, if it's facing the wrong way, I pick it up and then just immediately adjust it because I can... I know how, you know, which way I wanna face it and so on. Right? I can walk up the ladder with five screws in my mouth, you know, and, and then, you know, holding something with one arm. Yeah. And then- Yeah ... execute the tasks and so on. It's,

[57:16] Mian Khubaib Jim: It's

[57:16] Tom Lipinski: not something-- I mean, just walking up the ladder while holding something and, and re- you- It's almost like it's, it's, it's acrobatically impossible for robots to do, right? Before you even say how hard is the job, right? It's even getting in the position with the job.

[57:34] Tom Lipinski: It's, it's like it's, it's a different level of complexity, right? It's-- Where I can see robots making inroads and go... You've obviously got robots in construction now, right? You've got, you've got Monumental that recently raised thirty-two mil, uh, includ- from Khosla, um, and amongst others. Um, and you've got, um, other companies that have been relatively successful. Obviously, um, you've got Gravis, which has been moving earth around, but you've got, you've got other companies, the former sort of Waymo people going into construction, again, moving earth around, right? Um, there are are areas where you can deploy in construction if you narrow them enough. Yeah. And, and I think there is more room, uh, when it comes to offsite construction, where you've got high- highly repeatable items moving down the con- production line, you know, boxes where you need to, you know, put insulation in between the studs or, or maybe pull cables through. But actually, it's all rinse and repeat, so there's

[58:34] Tom Lipinski: no-- You're literally following a plan, right? Within a very well-defined environment, you know, 3D accurate environment where you know exactly where things are and so on. So you could, you-- Uh, uh, we'll see growing adoption of, of, of, of, of robotics in, in offsite where

[58:52] Mian Khubaib Jim: environments are

[58:52] Tom Lipinski: are semi-structured, right? Rather than onsite. And then we'll probably see some adoption onsite when it comes to

[59:01] Mian Khubaib Jim: Tasks

[59:01] Tom Lipinski: that not necessarily replace anybody, but make people's work more efficient, right? So let's say 40% of time on, on, on site is spent on not actually doing anything. You're, you're walking around looking for your tools, or you're looking for stuff that's supposed to have arrived, but, you know, it says there that it's arrived, but you don't know where. You l-- you know. So for example, moving goods around on sites, I can see that being gradually automated with, you know... I mean, it could be armed, m- AMR robots, or it could just be AMRs with bushels on them, you know- Mm-hmm ... going around construction sites, delivering stuff, you know, to

[59:34] Tom Lipinski: location. So we'll, we'll see, we'll see these sort of companies emerge trying to do this, probably larger sites adopting that. But when it comes to actually humans, you know, holding power tools and making change with those power tools,

[59:51] Mian Khubaib Jim: that's,

[59:52] Tom Lipinski: I would say, you know, you're not gonna see much, much happening there within the next 10 years.

[59:56] Mian Khubaib Jim: Maybe, maybe the robot can have six arms, you know?

[1:00:00] Tom Lipinski: Yeah. Um, I mean, again, I would always go back to, to trying to make sure that whatever solution that you're coming up with is appropriate for the challenge, right? So it could be that the six-arm robot is perfect for fixing plasterboard, but it might also be that the pla- plasterboard eight by four size is suboptimal for robotic application. You might have larger or smaller sheets emerge that work better with, you know, with robots and so on.

[1:00:28] Mian Khubaib Jim: In

[1:00:28] Tom Lipinski: construction, we've built things on, you know, the whole supply chain

[1:00:32] Tom Lipinski: is built around a a human being able to lift, you know, a bag of 25 kilos more or less maximum, human being able to hold something which is four feet, you know, four foot across and, and so on. Um, maybe we'll see more adoption when we, we start rethinking that. You know, what if, what if we didn't have a human form? You know, what if we could have any form? Would we then... You know, how would we design the components that go into construction, you know? Some companies are trying to do that. Um, All3 is trying to Redesign everything vertically in construction, right? So the way buildings are designed, the way the buildings are manufactured, and then the way the buildings are put together. Um, it's, it's a huge challenge, yeah. I, I sympathize with,

[1:01:26] Mian Khubaib Jim: with

[1:01:26] Tom Lipinski: some of the approach because, like, well, if you want to put buildings together with robots, then you probably

[1:01:31] Tom Lipinski: have to make them different. And if you're making them different, then you probably have to design them different. So you kind of see like, okay, it's sort of you've-- you- you're forced into being fully vertically integrated. Yeah. Having said that, how much money is required to, to build that fully integrated company that can design, manufacture, and build buildings and do all of that within unit economics that actually make sense, um, is... I don't have an answer for that. You know, it-- They've-- I've... I know that they've raised, I think, uh, twenty-five million recently, but, you know, it's, uh... I think you've seen Katerra in US. They, they, they were valued at two billion at some point, and that went badly wrong. So automation in construction, innovation in construction, robotics in construction are very big challenges. Not many companies have succeeded. Few, few have, like Skanska can actually build off-site and, and make money on that. But most of the companies that have

[1:02:31] Tom Lipinski: tried sort of failed. So-- And in this case, we're not just looking at building off-site, but we're looking at designing and then manufacturing with robots. So you might need not twenty-five mil, but twenty-five billion to, to sort of get to that point, right? But as I I said earlier on, if you redesign the process to suit the ro-robots, whatever they might be, you might have a, a higher chance of success.

[1:02:56] Mian Khubaib Jim: Hmm. Okay. Uh, you mentioned something about Ukraine war before, and, uh, the cheap drones became the deadliest weapon-

[1:03:04] Tom Lipinski: Mm-hmm ...

[1:03:04] Mian Khubaib Jim: in that war. And, uh, do you see like humanoids like Legged or Wheelie or whatever carrying a weapon soon?

[1:03:13] Tom Lipinski: Well, first

[1:03:13] Mian Khubaib Jim: of all- At scale Yeah,

[1:03:14] Tom Lipinski: yeah

[1:03:14] Mian Khubaib Jim: Not just one or two, like, you know

[1:03:14] Tom Lipinski: So as a, as a As an engineer and roboticist and a designer and lots of other things that I've done, I actually have an issue with The idea of, of A, building a humanoid

[1:03:26] Tom Lipinski: and B, getting that humanoid to carry something that was designed for humans because we can't change our actuators or end effectors, right? Yeah. So you, you, you're adding the complexity that's unnecessary, right? If you, if you're asking a humanoid to grasp something like an impact driver, right? You're making that job 10 times harder than if it could just have an end effector that would serve that function, right? So when you say humanoids carrying weapons, I, I probably don't see that as a sensible thing because it's just, you know... I mean, it has to grasp, it has to point in the right direction, it has to, you know, find a trigger, and so on. But there will be-- there already are robots, you know, that move around battlefields that can fire weapons and can do so autonomously. You know, it's like robots can drop obviously, you know, grenades on people and so on. But there is an emergence of those lots of tracked vehicles. Um, and famous case probably

[1:04:26] Tom Lipinski: came out last month, um, where a single robot held a village for like six weeks or something, you know, against the Russians just because it was, it was able to point a gun and, and, and, you know, and, and, and shoot in a direction, right? So you see the adoption. Uh, but like I said earlier on, Ukrainians have tried, you know, humanoids because, you know, bipedals can navigate, you know, rough terrain and, and so on, and, and clearly stated that that has failed, that they can't match tracked or wheeled vehicles. So yes, autonomous robots deployed in, in, in, in battlefields? Absolutely. Will there be humanoids?

[1:05:08] Mian Khubaib Jim: Maybe Maybe arms, maybe arms will be replaced with built-in machine guns or something like that, you know, heavy stuff. Anyways, uh, what, what is your personal red line on this? Would you build one?

[1:05:20] Tom Lipinski: I'm not into building humanoids, no

[1:05:23] Mian Khubaib Jim: No, no.

[1:05:26] Mian Khubaib Jim: robots with weapons or weapon robots, whatever you wanna call them.

[1:05:31] Tom Lipinski: It's not the space that I focus on, right? So, so I focus on, on industrialization, right? So my, my... Let's, let's start from philosophical approach, right? The way I see robots really making a positive impact on, on, on our society is if they enable us to re-industrialize, uh, our Western civilization, right? We, we have sort of forgotten how to make things, and we've outsourced all that thing-making to others, and we've become a lot poorer in effect, or less capable to respond to crisis and, and geopolitical events, and so on. I believe that deploying robots at scale, we can call them humanoids. I don't, you know- Yeah I don't see that being humanoid. It will be, um, moving robots with arms-

[1:06:17] Mian Khubaib Jim: Yeah ...

[1:06:17] Tom Lipinski: um, that can, you know, inference on edge. Um, I believe that deploying robots at scale in manufacturing

[1:06:24] Tom Lipinski: will enable us to rebuild our industries, right, and bring our supply chains back home and, and enable production at almost like the Chinese cost, right? We're flexible, scalable, cost-effective and, and, you know, and by let's say ten X-ing manufacturing, we can three X the jobs even though we've automated a lot of it. Because just, I mean, if you look at history, this is what happened when, you know, textiles have grown in the UK. You know, the most of... You know, yes, per meter of fabric, you know, made,

[1:06:58] Mian Khubaib Jim: E-

[1:06:58] Tom Lipinski: the the UK ended up or Great Britain ended up employing ten thousand fewer spinsters, spinsters, right? But at the same time, the production exploded, and then the entire employment in the enabling industry exploded, rails and, you know, canals and everything else, building the infrastructure to, to carry those things around, and so on. I think deploying robots in at scale, yeah, in, in, in manufacturing will enable us to re-industrialize in a similar

[1:07:24] Tom Lipinski: fashion, right? Just rebuilding infrastructure, enabling us to make things locally, rebuilding the use-- you know, the, the, the skill set that we've lost because we sort of outsourced all the manufacturing, and so on. And it will make for much stronger, much more equal, um, economies, and I think that's where they can have the biggest impact. Um, will these countries need to be defended against external threats? But, but I'm sure they will, and there will be other people focusing on that space. But I think the biggest societal impact you can have when you re-industrialize a country.

[1:07:58] Mian Khubaib Jim: Hmm. Okay.

[1:07:58] Tom Lipinski: So

[1:07:58] Mian Khubaib Jim: would Would you like, uh-- do you see this, the like industrialization, I, I see that point. But robots, the reason these humanoids are very fascinating and, and the demo and the laundry, all that stuff, because people think one day these robots can babysit or can take care of the, your, you know, senior citizens or all that stuff. Would you, would you let that, a humanoid or a robot or whatever

[1:08:16] Mian Khubaib Jim: pick up that sort of, you know, chores for you?

[1:08:16] Tom Lipinski: What, what, what, what we will see is the emergence of, of... You, you might not even call them robots at some stages. You will see, let's say, intelligent furniture, you know, or whatever. Oh, amen. That's it. They- Think a- think about it. We, we get so obsessed by trying to copy a person, yeah? Whereas actually, let's say you look at assisted living, yeah? What you've got when it comes to assisted living is improving mobility, improving independence, improving access to certain goods and services for people who are no longer as able as they used to be, yeah? Do you need to duplicate a person and have another large object, hundred kilos, 80 kilos heavy, you know, that can-- that has safety issues, that can fail, that can fall on people or whatever? Or would you rather say we are now, bec- thanks to robotics, yeah, we are surrounded

[1:09:08] Tom Lipinski: by intelligent buildings, intelligent furniture that literally does equivalent job for you, but it can m- help people move around because you sit... Your chair moves where you need to, you know. Your, your, your transfer to the bed might actually be like you just get, you know, handled by a chair into the bed. Your bathing, you know, might have intelligent bathroom that might do all that in- instead of you trying to do that, you know, it might actually get you washed in, in an effective way. You don't need to introduce this big lump of metal, yeah, which is bipedal, which has 90-minute battery time, which is... You don't even know how to shut it down safely, right? Or do you say we now have assisted living intelligence that has-- that, that the building can take care of, you know, an older-- elderly in this assisted living setup, and actually it's based around mobile furniture. It's based around... You know, because what you're trying to

[1:10:08] Tom Lipinski: solve, I mean, y- again, it's about being appropriate, right? You're solving mobility issues, you're solving access issues, you're solving supervision issues.

[1:10:17] Mian Khubaib Jim: We

[1:10:17] Tom Lipinski: We can deliver that, you know, with, with just improving objects around you, you know, and, you know, your chair. We have moving chairs. They're called wheelchairs. They're robotic wheelchairs and so on. If you draw a line of these things getting more capable, you know, they could do that one function. Again, w- yes, we're talking about specialization, but that's the only way to do something really, really, really well, is when you, when you focus on specialization. So I believe that there will be huge impact, yeah, on elderly and so on, but it m- might not take the form that we all seem to be sort of somehow hoodwinked into believing.

[1:10:46] Mian Khubaib Jim: And I think IKEA, IKEA got a nice idea, intelligent furniture. Okay, uh, so, uh, there, there is this, uh, this story like in Japan, in Osaka, uh, mall.

[1:10:55] Mian Khubaib Jim: So there is... There was this wheelie robot and, uh, a group of children came together and started kicking it, and now the researcher has made it like if you see childrens, you know, run from that place. Something like this. It was a very funny story. Do you... Why-- You think when we see a robot with wheels or humanoid shape or anything like that, and it was just a simple robot, you know, picking up stuff out that, uh-

[1:11:18] Tom Lipinski: time. Uh,

[1:11:18] Mian Khubaib Jim: W- do we need to teach our children or do we need to, you know, learn about how to interact with them or, or something?

[1:11:26] Tom Lipinski: I, I think-

[1:11:27] Mian Khubaib Jim: And they really enjoyed it, by the way.

[1:11:29] Tom Lipinski: Yeah. So I mean, um, have you seen Chappie, the, the South African movie about the robot that gains intelligence, you know? It's-- That robot gets a lot of abuse as well, right? There, there's, there's a quite heartbreaking scene when the robot gets beaten up badly and, and, and you feel for it, right? Um, to be honest with you, right, th- this, this subject matter that you've just raised, um,

[1:11:57] Tom Lipinski: deserves, um, university degree courses to look into and so on. It's a very deep subject, right? And you could, you could look at it from different perspectives. You could say, "We need to educate people to interact with robots better." Or you can say, "Well, actually, maybe if the robot had a form factor that was cuddly or inviting or,

[1:12:19] Mian Khubaib Jim: or

[1:12:20] Tom Lipinski: it invoked pleasant feelings, then they wouldn't get abused," you know, or whatever. So we, we've looked at this. We've looked at form factors, right? Is there a form factor that makes a robot more likely to be abused than not, right? Um, is there a, is there an appearance that we could give robot? Or is there maybe a way of, you know, it's like, "Well, if you tip me over, I will be sad," right? You know, it's like rather than, you know, it's an object, right? So and you obviously have-- There are lots of examples, you know, from, let's say, you know, Disney and so on, when you see robots that are sort of cuddly or, you know, Baymax

[1:12:56] Tom Lipinski: from Big Hero 6 or, or Bernie from, from WALL-E, you know. And there are lots of robots out there that are just like oh so cute that you probably wouldn't abuse them, right? But at the same time, how does that sort of human-robot interaction, you know, what, what governs it, right? What psychological issues or, or, or motivations are behind it? It's a huge, it's a huge subject area, I think that deserves a lot more research, you know, by literally universities focusing the entire fields of, of research into that. But I, I see this as, you know, in a simplistic way. You can have a robot that looks in a way that you just don't wanna- Mm ... don't wanna hurt it because it will be like, hmm, it looks so cuddly and harmless or whatever. Or you could have a robot that looks like, you know, "Don't mess with me." Mm. Yeah. And then you're like, "Yeah, I'm not gonna, I'm not gonna tip that one over." "Looks like it could beat me up." Um, you could have very strict regulations. You

[1:13:56] Tom Lipinski: know, if you vandalize a robot, there is, you know, you paying a fine or going to prison or whatever. Although we're probably moving away from the whole prison thing at the moment, but that's a different subject. Yeah. Mm-hmm. Um, so various ways of-- You could, you could deal with it in various ways. I think it will-- It's At the moment, we don't know what makes people beat up robots, right? Mm. But, you know, there are, there, there are teenagers and, and, and, and others that vandalize trains and, you know, and, and, and other things. So it's, it maybe it's a people thing rather than a robot thing.

[1:14:31] Mian Khubaib Jim: Mm. Okay. Okay. Uh, there is this interesting thing from Bill Gates. He says that any job that a robot, humanoid or whatever you are gonna call it, is taking, there must be a tax on it. Let's say it was a sorting job or something, there must be a robot tax. What do you think about that?

[1:14:45] Tom Lipinski: Well, it will happen. I mean, you know, if you think about the government, especially our government, right, the, the one thing that

[1:14:54] Tom Lipinski: they, you know, that keeps them awake at night is, is trying to devise new ways of taxing people. Yeah? And businesses and so on, right? So, so let's say we, we provide a labor service to a customer, right? And we supply labor to, to, to a, to a manufacturing site. It's 100 robots doing jobs, right? Obviously, when we invoice monthly, we, we put VAT on that. Yeah? So there's a, there's, there's a 20% VAT on that as well immediately. If you were an employee, there will be a, there would be a PAYE tax, you know, and pension and so on. Robots might not need pensions, but I can see government introducing law that, you know, where robots are deployed, and then you have to kind of work out, are we talking about fixed robots, production robots, arms, or, you know. Is a robotic arm a robot, or is, does it have to be, be a f- two armed robot? Does it have to be mobile? Does it need to have inference, you know, or whatever. But I can see government coming

[1:15:54] Tom Lipinski: out with all sorts of ways to tax it, um, Because that's what governments do, you know. And obviously, I've got my own very strong feelings about it, um, which probably not subject of this podcast. But it will, it will come, and then the business will come out with various ways of trying to go around it so they can actually operate.

[1:16:10] Mian Khubaib Jim: Hmm. Okay. So let's suppose... No, not let's suppose it, that it looks true. Brain is built in America and hardware is built in China. Where would we stand in 2040, like Britain, United Kingdom?

[1:16:24] Tom Lipinski: Some say we'll be a third world country. Um, look, the,

[1:16:29] Mian Khubaib Jim: the

[1:16:30] Tom Lipinski: the UK government has a habit of announcing tiny,

[1:16:35] Mian Khubaib Jim: tiny

[1:16:36] Tom Lipinski: initiatives with huge fanfare and then failing to deliver them, right? Even the whole data center

[1:16:47] Mian Khubaib Jim: adoption

[1:16:47] Tom Lipinski: in the the UK, right, with great government announcements is completely

[1:16:55] Mian Khubaib Jim: on

[1:16:55] Tom Lipinski: hold because data centers can't connect to the grid. They can't secure the power. They can't, you know. And the announcements, I mean, the, the, the, the, the local authority announcements were like, "Oh, we're building this huge 70 megawatt, um, so s- uh, yeah, 70 megawatt data center." But the planning condition says that you need to source your own supply. You can't draw it from the local grid or whatever. In the UK, the UK, you wait eight years for grid connection. You, you can build an offshore power station, an, an offshore wind turbine, you know, field, and then have it sit around for four years because the grid connection isn't there, right? So we are so bad at infrastructure that we can't even deliver those tiny, tiny, tiny announcements about, you know, a 14 billion spent on data centers. Even those things can't

[1:17:47] Tom Lipinski: happen because we are incapable of integrating that into our power grid. And at the same time, you've got China that's pumping 200 billion into, you know, one industry, right? So

[1:17:56] Mian Khubaib Jim: And building artificial sun to have unlimited energy and, uh, and so many other things. Yeah. Okay, uh, so w-what would you recommend to a 16-year-old right now to learn for next five years?

[1:18:13] Tom Lipinski: Physics, engineering, these are, these are the, um,

[1:18:18] Mian Khubaib Jim: hardcore

[1:18:18] Tom Lipinski: skills that will, will ensure that you actually do have work, right? And, and I've seen this even in some-- There's a internally famous case. When you're a roboticist, you laugh about it. You know, there's a, there's a company building humanoids on wheels, and they, they put a prototype together and then set it for production. They've completed the design and, and, and, you know,

[1:18:44] Tom Lipinski: they've really targeted production at scale and built all the, you know, supply chain around it and, and, and tooling and so on. And

[1:18:52] Mian Khubaib Jim: When it

[1:18:52] Tom Lipinski: it came to, like demonstrating it, the moment it fully erected itself, it tipped over,

[1:18:59] Tom Lipinski: because it was, you know, nobody thought of physics and gravity when they were designing it. So even in current fields that actually work in robotics, you know, you can see that, that physics and, and engineering, um, leave a lot to be desired. So hard sciences, uh, especially around engineering, you know, mechanical engineering, design engineering and, and physics, I believe are, are still gonna hold up strong. I wouldn't talk about software because we've got no idea where that will land, but it looks like, you know, anybody can code now. Yeah. The, you know, lots of people have different opinions. But when it comes to hard engineering subjects, I think they, they have, they have a great future.

[1:19:40] Mian Khubaib Jim: So, uh, let's talk about you. This, uh, a young man in Poland building sites. It's you.

[1:19:47] Tom Lipinski: And,

[1:19:47] Mian Khubaib Jim: uh, how

[1:19:47] Tom Lipinski: did

[1:19:47] Mian Khubaib Jim: you end up becoming a serial entrepreneur? What brought you to Britain?

[1:19:51] Tom Lipinski: Um, it's--

[1:19:51] Mian Khubaib Jim: I

[1:19:51] Tom Lipinski: I think it's an addiction. I think some people are just relentless builders, and this, this has always been me. I, you know, as a kid, I had this project of building a submarine. Thank God I never got it to test it- ... 'cause it wouldn't have ended up well. I was halfway through building my own car, uh, as a teenager. I, I filed my first patent at the age of 13. Um, it made into some sort of, you know, national magazine called Młoda Technika. Um, and, and I've just-- I could never stop. I've always had this,

[1:20:26] Mian Khubaib Jim: this

[1:20:27] Tom Lipinski: mind that, you know, this is wrong, there's a better way of doing it and, and, and so on. And quite

[1:20:32] Tom Lipinski: often I, I managed to, to prove myself right. So when, when, you know, when COVID started, I, I kinda had this Beautiful Mind moment when I was like, "They're all wrong. It's airborne, and it's because of X." And I did a, a deep dive and spent few good months, you know, researching the subject, published academic papers, won awards for best contribution to science and, and, and, uh, and art of building service engineering, got, got referenced in government guidance afterwards, and so on. So I've always had this very critical eye to, to like how things are set up and whether they work in the way that, whether that's an optimal way of working and so on. And I've always had this relentless

[1:21:19] Mian Khubaib Jim: Addiction

[1:21:19] Tom Lipinski: to b-build things, right? And I think that the challenge which, you know, over time you realize, uh, as you grow older, is putting that to best use. You know, focusing that, that, that, that

[1:21:33] Tom Lipinski: energy and this, this addiction to build on, on things that can actually have the biggest impact, right? On, on, on society, you know, where you can build a big scalable business and so on. So, um, that's that. But I, I... When it comes to Poland, I, I, I sort of grew up in Poland to an extent, and then at some point I was like, "I'm done with that." And, and literally just packed my stuff, got on a bus with 100 quid in my pocket in 1994 and, and arrived in the UK. And I had-- I didn't know anybody in the UK. I didn't speak English. and it was just one of those, you just put yourself out there and, and things will work out. So I guess that in itself is quite entrepreneurial. If you've got somebody who can just randomly, you know, move somewhere else without any links of family or friends, uh, and then just get at it, it's... It, It, it, it probably demonstrates that sort of

[1:22:33] Tom Lipinski: will to risk everything just for the sake of, you know, learning more. Hmm. Okay.

[1:22:35] Mian Khubaib Jim: So this is a bit personal question. Okay. Uh, two years back, you were r-removed as a CEO and be- and, uh, was given CTO title and then in between Christmas and New Year of 2024, you resigned when everybody was doing, you know, partying and stuff. And w- the thing is, y- as an entrepreneur, especially young entrepreneurs, they never see this coming. And, uh, would you like to share that story? Or what are the lessons?

[1:22:55] Tom Lipinski: I would- Yes. So,

[1:22:55] Mian Khubaib Jim: it

[1:22:55] Tom Lipinski: it happens to a lot of people, you know, when they've built a company and then they part ways under whatever circumstances where things haven't worked out. if that is to be a cautionary tale, I would

[1:23:14] Mian Khubaib Jim: Wind the

[1:23:14] Tom Lipinski: the the

[1:23:14] Tom Lipinski: clock back 10 years, at which point we actually lost the company. We just didn't know about it, right? We, we were raising- Who is we? Me and my co-founders, right? So it was, um... it-- I founded Q-Bot with, uh, with an Imperial College professor. Um, we were sort of, um, squatting at the, uh, Skempton building, eh, at Imperial. Um, and then another co-founder joined us, um, who did-- who went through Imperial as well, and we were fundraising,

[1:23:48] Mian Khubaib Jim: and

[1:23:48] Tom Lipinski: we were were building a robotics company in 2012, right? It's crazy. Like, nobody would invest in robotics then. It was not, not, not a sexy subject. and we've come across one investor that was prepared to put some money in and put forward a term sheet that our lawyers told us not to sign. Yeah. that term sheet had extremely controlling conditions to it, right? And the lawyers, the lawyers said-- uh, uh, made a simple

[1:24:14] Tom Lipinski: observation, like, "If you sign these terms, you are not going to be in control of this company." Yeah?

[1:24:21] Mian Khubaib Jim: And

[1:24:23] Tom Lipinski: in the end, we did sign those terms, right? We took some money, and it wasn't a lot of money. It was like 50 grand, right? Mm. But we had, you know, Innovate UK funding lined up, and we just needed to show something and so on. And that- that-

[1:24:36] Mian Khubaib Jim: That-

[1:24:36] Tom Lipinski: That- That's

[1:24:36] Mian Khubaib Jim: That's 2012, right?

[1:24:36] Tom Lipinski: That was 2014. 2014. Oh, yeah. So we founded the company... So 2014, we took that investment. And what that term sheet meant, and that's the, you know, for young entrepreneurs, if you want to learn something, was that the investors that came in at that time with 50 grand. I mean, we've raised over 10 million of investment since. Nobody came on those terms, right? So the, the veto rights that handed over to that investor, the control rights

[1:24:54] Mian Khubaib Jim: that

[1:24:55] Tom Lipinski: we've handed over to that investor, the control rights that we handed over to that investor meant that we could not control the

[1:25:01] Tom Lipinski: strategy of the business. We could c- not control the, the, you know, the execution, the markets, whatever. We were not in charge, right? So those investors could place, you know, could decide, "We'll just put a CEO here. We get rid of the CEO, you know, we'll put another CEO in place, you know," and then three months later, another CEO. You know, we've gone through... At one point, we've gone through a succession of three CEOs in a year, right? And, and we were just like passengers watching all this unfold because we were not in control. Right? So

[1:25:33] Mian Khubaib Jim: the...

[1:25:33] Tom Lipinski: the... I don't wanna dive deep on, into, into the details, but actually you can link it all back to that moment where we handed over control to the company we haven't built yet to somebody for 50 grand. Yeah? And we were na- never able to regain control of that. And, you know, at some... At that time you might think, "Well, but obviously we both want the same thing." Yeah? You know, "We want this company to be a success. Our investors want this company to be a success,

[1:26:01] Tom Lipinski: therefore our objectives are aligned." Yeah? Well, not necessarily, right? You might want to be building a sustainable business, whereas your investor might be looking at, "How do I represent the highest value of this company on my assets under management- Yes ... you know, at the highest possible level, so my--

[1:26:23] Mian Khubaib Jim: So

[1:26:24] Tom Lipinski: I, I can be seen as successful or what- or whatever." And they may... They can make strategic decisions based on their asset under management valuation rather than the sustainability of the company, right? And, and at that point, you might completely be trying to achieve opposite things, right? So, so bear that in mind that

[1:26:48] Mian Khubaib Jim: if

[1:26:48] Tom Lipinski: you're building something, right, there might be a sacrifice in not taking money from some people

[1:26:58] Mian Khubaib Jim: That

[1:26:58] Tom Lipinski: That make you suffer a bit more. Yeah, and obviously when you've

[1:27:01] Tom Lipinski: got-- when you're looking at that check, you're just like, "I just want that money." Yeah. Mm. "I just want that money. I really want that money." But actually, rushing here without good legal advice, which we had, we just ignored it. Yeah, without good legal advice. Other things, you know, that you could look at, which is, you know, participating rights, you know, ratchets, and so on. It's really easy to, to write yourself in a situation where, A, you're not in control of your business, or when you exit for tens of millions, you might actually not see a penny. That's another thing that, that I've seen a lot of founders fall into, the trap of, you know, two X, you know, participating rights and so on, where, you know, you take ten million, but actually those investors now need to get the first twenty million from your sale. And then you take that money twice, and then you, you know, you get to sell for, I don't know, forty-eight. And before you know, you know, you get twenty-five thousand for, you know, ten years of building. Mm. It's, it's a

[1:28:01] Tom Lipinski: cautionary tale. For us, it was to control the company.

[1:28:03] Mian Khubaib Jim: Hmm. Okay. Okay, who, who else left? Like you, you left the company, and the co-founders as well, or are they still there?

[1:28:09] Tom Lipinski: The, the co-founder that I mentioned, uh,

[1:28:13] Mian Khubaib Jim: exited

[1:28:13] Tom Lipinski: before me.

[1:28:14] Tom Lipinski: Like a couple of years before me. Uh, that was also very rough and turbulent.

[1:28:19] Mian Khubaib Jim: Uh, one more question. Why, why did you choose that period? What were you thinking between the Christmas and the New Year?

[1:28:25] Tom Lipinski: I didn't choose.

[1:28:28] Tom Lipinski: It was, uh, it was a very tense situation. Um, we all knew... I mean, to be honest with you, it's not just me. A majority of Q-Bot employees left around the same time.

[1:28:40] Tom Lipinski: Okay. Yeah. So we've, we've-- and they've joined us. They, they, they-- We're working together at AIRA now. We, you know, um, I didn't- We can move on from it ... I, I didn't choose the time.

[1:28:40] Mian Khubaib Jim: It's fine. It's fine. Okay. These are just some quick question, one line or one word.

[1:28:44] Mian Khubaib Jim: Uh, space,

[1:28:44] Mian Khubaib Jim: data centers, genius or distraction?

[1:28:48] Tom Lipinski: I think if we want to solve physical AI

[1:28:51] Mian Khubaib Jim: AI

[1:28:52] Tom Lipinski: autonomy over the next sort of twenty, thirty years, we will need to carpet the solar system with, with AI, you know, with data centers in space. Because I just can't see us- Yeah ... unless quantum computing solves it all. But I believe that we will need that compute. The funny thing is, if you look back we've always been skeptical of the amount of compute that we were gonna need next year or the year after or the year after that.

[1:29:24] Mian Khubaib Jim: We've

[1:29:24] Tom Lipinski: always grown into it. Yeah, so I, I, I think, I think it's a shape of things to come. Um, and it might be the most efficient way to do it.

[1:29:36] Mian Khubaib Jim: Okay. AI bubble pop which year or never?

[1:29:39] Tom Lipinski: There will be a huge

[1:29:44] Tom Lipinski: reckoning in generative AI space when

[1:29:49] Mian Khubaib Jim: companies

[1:29:50] Tom Lipinski: like OpenAI and, um...

[1:29:52] Mian Khubaib Jim: Wait,

[1:29:52] Tom Lipinski: Anthropic. Jesus. Come

[1:29:52] Mian Khubaib Jim: on, don't forget him.

[1:29:52] Tom Lipinski: Yeah. Okay, so there will be a reckoning in generative AI space when OpenAI and Anthropic start raising prices to s- you know, justify their IPOs and, and, and, and, and start going for profit, right?

[1:30:04] Mian Khubaib Jim: It's getting cheaper.

[1:30:06] Tom Lipinski: It's going to get more expensive, in my view.

[1:30:10] Tom Lipinski: Um, and we'll see a lot of wrapper companies fold. Yeah, so I think there will be, there'll, there'll be several waves of AI bust, right? Or bubble bursting.

[1:30:22] Mian Khubaib Jim: Okay, so next few questions are like, like quick questions.

[1:30:22] Mian Khubaib Jim: Okay. So AGI, which year or never?

[1:30:30] Tom Lipinski: 2030s

[1:30:30] Mian Khubaib Jim: 2030s Tesla Optimus, theater or a product?

[1:30:40] Tom Lipinski: What's my insurance? Um

[1:30:42] Mian Khubaib Jim: I'd

[1:30:45] Tom Lipinski: I'd rather not say

[1:30:47] Mian Khubaib Jim: Okay. Figure AI, $39 billion valuation.

[1:30:53] Tom Lipinski: valuation. What do

[1:30:53] Mian Khubaib Jim: do you think?

[1:30:56] Tom Lipinski: Theranos.

[1:30:56] Mian Khubaib Jim: China robot war on the hardware specifically.

[1:30:59] Tom Lipinski: I think China will control most of the robot manufacturing

[1:31:03] Mian Khubaib Jim: Okay. Your robots, how many starts by 2030?

[1:31:09] Tom Lipinski: Hundreds.

[1:31:10] Tom Lipinski: 10,000 robots deployed

[1:31:11] Mian Khubaib Jim: Britain's first trillion-dollar company

[1:31:17] Tom Lipinski: Let

[1:31:17] Mian Khubaib Jim: Let me make it more specific. In your lifetime, yes or no?

[1:31:20] Tom Lipinski: Yes.

[1:31:22] Mian Khubaib Jim: Self-driving car across London tomorrow, yes or no? There's

[1:31:27] Tom Lipinski: lots of Waymos

[1:31:28] Tom Lipinski: in London already, just not self-driving.

[1:31:30] Mian Khubaib Jim: Mm-hmm. Okay, Poland gave you work eth- uh, work ethic and Britain gave you the company. Fail, uh, yes or no?

[1:31:30] Tom Lipinski: I have a very Eastern European approach to things, yes.

[1:31:30] Mian Khubaib Jim: Hmm. Best decision

[1:31:30] Tom Lipinski: I would say proposing to my current wife Current

[1:31:30] Mian Khubaib Jim: wife. Okay. Uh, last thing that scared you

[1:31:30] Tom Lipinski: Last thing that scared me

[1:31:30] Mian Khubaib Jim: Yeah. In terms of business, AI or something It

[1:31:30] Tom Lipinski: It might come across strange, but I am actually scared of the government

[1:31:30] Mian Khubaib Jim: Okay. Okay, last robot demo, humanoid or whatever, that amazed you. You said, "Okay, they got it

[1:31:30] Mian Khubaib Jim: right."

[1:31:30] Tom Lipinski: I think Dex are onto something.

[1:31:30] Mian Khubaib Jim: Mm-hmm. Okay. By 2030, what would be obvious that people are not looking at right now now

[1:31:30] Tom Lipinski: or

[1:31:30] Mian Khubaib Jim: missing right now?

[1:31:30] Tom Lipinski: Um, where the most value is created in the robotic space.

[1:31:30] Mian Khubaib Jim: this is the last question. Uh, this is Inference Podcast. That is in your company exists, my company exists is because of the inference. Uh, so I want you to give me your inference, one specific prediction about ro- robot industry, and where do you see your AIRA company that's in 2035?

[1:31:48] Mian Khubaib Jim: So

[1:31:49] Tom Lipinski: prediction when it comes to what's useful on the inference side, I would say small specialized models. Um, I believe they'll be a lot more successful than sort of

[1:31:58] Tom Lipinski: large foundational models. Um,

[1:32:03] Mian Khubaib Jim: where

[1:32:03] Tom Lipinski: we see-- where I see AIRA in 2035 is helping to

[1:32:10] Mian Khubaib Jim: reshore

[1:32:10] Tom Lipinski: reshore manufacturing into the UK by changing the way we make things.

[1:32:17] Mian Khubaib Jim: Okay, quick follow-up question. Let's say you have sold your company for £100 million. You exited, whatever. What would you build next? What is the next opportunity you see that somebody can tap on?

[1:32:18] Tom Lipinski: I will only know that when it occurs

[1:32:18] Mian Khubaib Jim: Okay then. Thank you.

[1:32:18] Tom Lipinski: Thank you.