In the popular imagination, Japan is almost synonymous with robots.

While Japan once dominated cutting-edge robotics, over the past decade she has fallen further and further behind the US and China.

Today we sit down with Chiamin Lai of Firstlight Capital, who believes that Japan might just regain that leadership. We talk about the unique opportunity and advantage Japan has in the deployment of practical physical AI, the enterprise culture that is holding it back, and what a handful of innovators are doing about it today.

It’s a great conversation, and I think you’ll enjoy it.

Podcast: Play in new window | Download | Embed

Subscribe: Apple Podcasts | RSS | More

Show Notes

  • How starting startups in Japan has changed over the past 20 years — especially for foreigners
  • How Japan’s labor shortage is driving the adoption of physical AI
  • The biggest problem in integrating GenAI and robotics
  • The best use cases for physical AI today and why healthcare is not one of them
  • How secrecy is holding back AI innovation
  • What keeps Japanese enterprise from embracing open innovation
  • Can Japan’s VC ecosystem afford to fund AI in the era of massive funding rounds
  • Why physical AI companies should not create their own hardware
  • Why Japanese startups should not look to hardware for competitive advantage
  • The importance of industry cooperation and why it’s critical for Japan’s AI success
  • What physical AI will look like in Japan in five years

Links from the Founder

Transcript

Welcome to Disrupting Japan, Straight Talk from Japan’s most innovative founders and VCs.

I’m Tim Romero and thanks for joining me.

Japan has always had a special and very positive relationship with robots from Astro Boy and Doraemon in the fifties and sixties, to Sony’s Asimo in the 2000s to SoftBanks Pepper in the 2010s. It has always felt like Japan was set to create and then to lead a humanoid robot revolution.

But that didn’t happen.

In fact, today, Japan seems to be far behind both China and the US in the development of not just humanoid robots, but intelligent robots in general.

Well, today we sit down with Chiamin Lai partner at Firstlight Capital, to discuss how that came to be and what we can do about it. Now, Chiamin’s investment interests are deeply focused on physical AI and specifically physical AI startups in Japan. And she remains optimistic about the future of AI and robotics in Japan.

We talk about the market and the financial structures pushing Japan to adopt meaningful physical AI before the rest of the world. The technology and social challenges of trying to use AI and robotics and healthcare, and some really great advice for physical AI startups that are planning to raise money.

But, you know, Chiamin tells that story much better than I can. So, let’s get right to the interview.

Interview

Tim: So, we’re sitting here with Chiamin Lai, the general partner at Firstlight Capital, and a director at Japan Venture Capital Association. So, thanks for sitting down with me.

Chiamin: Thank you, Tim.

Tim: Before joining Firstlight, you worked in startups and investing in Japan and in China and in the US but you’ve had ties to Japan for quite a while, haven’t you?

Chiamin: Yeah, I was born Taiwan, but then I came here when I was teenager, and after that I received education here. I also work in Japan, but then later to Europe and then came back. So I can say this is like my hometown in the way. I have more friends, more connection, and my family here. So yeah, some of my friends said, you are more Japanese than we are. Sometimes I agree.

Tim: Yeah, I know the feeling. I’ve been here over 30 years myself. Yeah, it kind of sneaks up on you. And Japan is a very comfortable place to live once you kind of get used to it all.

Chiamin: Yeah. But I would say it actually changed a lot for the past 20 years or 30 years. When I came, Japan is not that open up. Like people sometimes complain about they have a hard time finding apartment and so on. I’m like, okay when I came it was worse.

Tim: Yeah, that’s for sure.

Chiamin: Yeah. Finding a part-time job, finding a job was not that easy at that time because we still have a lot of population. They don’t really need a foreigner to work for their company.

Tim: Well, I think that’s one of the biggest changes is so when I started my first startups back in the dotcom era, a big part of it was that there weren’t a lot of options open to foreigners in Japan. Having a regular career track job was exceptionally rare, and now it’s almost kind of flipped.

Chiamin: Yeah. Yeah. I agree. I think it’s good for the country. I think both you and I, we stay here for a long time, so we have a deep understanding about this country and a lot of foreigner like us I think we all wish that we can contribute somehow to this society because it’s a good country to live. That’s also one of the reason why, even though I left for few years and I decided to come back to Japan and to do some contribution, and that’s one of the reason I’m here.

Tim: Well, let’s talk about that because first slide invests fairly broadly in pre-seed, up to seed, but your own focus and your own passion seems to be in physical AI. So what is physical AI and why is it important now?

Chiamin: So as a fund, our investment thesis is how can we actually solve the demographic challenge here in Japan. And if you want to solve that problem, of course you can use AI, we can use software, we can use automation. A lot of solution out there. So for my investment, I also of course invest in AI company, SaaS company. But one of the reason why I have heavily looking into physical AI is because of my background. Before I came back after COVID, I work in China for seven years. So I actually witnessed the heavy growth in China for entrepreneurship from 2011 to 2019. And then I also was working for a DCM, which is a Silicon Valley venture capital. I also was seeing how the Silicon Valley startup was doing. When I came back to this industry, I look in Japan, I think one thing that a lot of entrepreneurs forgot is what is the strength about this society and what is the heavily problem that you solve. For the past five years, SaaS has become a very common platform or common tool for a lot of office worker. But what we’re looking into this social problem right now, we need to urgently solve the problem on the ground, which means that essential worker. If you look at the restaurant, if you look at the construction, we are heavily lacking people, but we don’t really have a solution to solve. So, let me come back to your question of physical AI. My definition of physical AI is how can we embed artificial intelligence on the actual physical operation?

Tim: I mean, there’s a lot in there. So let me try to peel that back a bit. So, could physical AI just be IoT rebranded with some AI, or is it something fundamentally different?

Chiamin: I think right now, if you look at the VC, a lot of investment tied to humanoid robotics. If you look at recent fundraise with figure AI, with all those very interesting robotics company, humanoid company, right now, I think US, they are aiming for that. For me, I think it could be robotics, it could be using hardware to embed artificial intelligence. And to go back to you saying, is that a rebounding of IoT? I disagree. Because what we are talking about is how do you distinguish gen AI and AI, that’s the discussion we normally have.

Tim: Well, I mean, I want to get into that just nailing down what physical AI is before we get into gen AI. So there’s AI can be applied to industrial processes, to robotics control, to like I mentioned IoT with a bit of AI shoved in there. But physical AI, does it require robotics? Does it require AI actually controlling something in the environment?

Chiamin: I would say physical AI would be the enabler for automation in terms of action.

Tim: So, does smart sensors count as physical AI?

Chiamin: No. No, I don’t count smart sensors as physical AI, it’s automation of the procedure or process on the ground. So, let me give you a very easy imagination for physical AI on the construction side. So, if you are building the house here in Japan today, you need to have director checking the progress. So, making sure your house can be built and also making sure it’s on time and on quality. Today, how do they do that? Is they will need to visit the site. They will need to use the measure to really measure exactly the length and so on. But Japan, today, we are lacking a lot of experience monitor person because they are retiring. So, what’s happening today is the undergrads come to onsite, he’s going to check in your house. Are you going to be comfortable for not really experienced person to making sure all the construction is correct. That’s going to happen in 10 years, or even in five years. Now, when we say physical AI is if you have a tool there that has intelligence that can actually check what is going on the ground, using the AI to understand your status as well as the lengths because you need to measure what is going on, if that can be automated. And then you can have an agent to tell you, okay, in this case, this is what you need to do.

Tim: So, the physical AI is AI that is interacting with the physical world in some…

Chiamin: Yeah. And then it can generate action. And that’s why I think why people say robotics right now, because you need somebody to do the actual physical world.

Tim: Well, yeah. In most cases there would be some sort of an interaction component.

Chiamin: Right, exactly. Is that there needs to be action there.

Tim: So talking about AGI, I’ve got this kind of working theory that even though we don’t quite know what it is yet, that if it emerges, it’s going to come from this kind of physical AI because this is the only chance that AI has to interact directly with the physical world and respond to stimulus. And that’s like where our intelligence evolve.

Chiamin: Exactly. Exactly. I think that’s why I’m very excited looking to this space. If we really want to solve the social issue here in Japan, I think if we want to keep the economy size as it is today, we cannot run this economy with half of a population. I mean, we can have chat GPT to help office worker to be more efficient, but who is going to do the groundwork?

Tim: Well, I think Japan really is a fantastic market for this because of the declining population. But also culturally, Japan is very friendly towards automation.

Chiamin: Yeah. Japan is actually the first country actually launched human lawyer robots 20 years. Hondas…

Tim: Yes, yes. Asimo

Chiamin: Asimo. Yeah, exactly. But at that time, there wasn’t a gen AI. So, there’s very limitation in terms of use case in terms of IT intelligence.

Tim: Well, let’s bounce back to gen AI here. So, like Invidia recently talked about generative physical AI. Now help me square this circle because let’s call it traditional AI, which has been involved with robotics since the very beginning. So generative AI is really good at taking large training sets and generating output that is reasonably close to what you expect. And the more traditional robotics AI has been smaller training sets that produce very exact precise output. So, why is LLMs and Gen AI appropriate for this kind of physical AI? Is it a replacement? Is it going to enable new functionality?

Chiamin: The industrial robots, they are more rule-based robots. So, you need to have a preset there. But then physical AI right now is the brain itself is similar to old Chat GPT you can understand that way. So, it means that you don’t really need to tell the person exactly everything. Just like what we learn on our brain. We probably have some basic, but we learn from ourself. And that’s what the physical AI trying to do achieve in terms of the brain. That’s the first things…

Tim: Let me push that a bit because LLMs, the inputs and outputs of inference very much look like intelligence, but they don’t learn after they’ve been trained. So, how much of this do you find is a metaphor and how much do you think is genuine intelligence?

Chiamin: We need to separate gen AI and physical AI? So, what is difficult right now for physical AI is actually not the brain itself. The difficult part for physical AI is in the action level. Meaning if you want to actually make the robotics work, you need the robotics to really understand what’s the environment as well as the physics. And today we don’t have enough data for robotics to understand physics or to understand the space. So, what most of the startup right now they are struggling is not in the brain. It’s how to connect the brain with the arm. With the space, the distance, if you want to hold something, the strength, the power of the catch will be different. And that is the difficult part. How are you going to control the hardware? The reason why it’s difficult for physical AI or robotics AI Company to emerge without large capital is you need to have hardware to really train what’s going on together with the robotics LLM. So, it’s a combination with hardware and software.

Tim: Well, that makes sense. I mean, software always moves so much faster than hardware. I mean, you can iterate on software constantly and cheaply and iterating a hardware is capital and time intensive. Is that one of your focuses now trying to find startups that are innovating in that coordination, in that interaction component?

Chiamin: I think we have a lot of opportunity here in Japan. One thing is, if you look at Chat GPT or gen AI product, most of those model are studied by a lot of large dataset. What is missing today in the physical AI world is nobody owns the data. Nobody owns the physical data, which means it doesn’t matter if you are startup in China or United States or Japan. Every single startup needs to figure out a way how to capture the data.

Tim: And how reusable is that data? I mean language data, papers, blogs, whatever. As long as it’s in the same human language, it’s readable, digestible by everyone. But this physical interaction, how much of the data from a Toshiba robot would be applicable?

Chiamin: It’s not even those data for those physical AI robotics company. They’re trying to create a generative robotics. So, it means that they want the robotics to know how to clean, how to fold the clothes, how to walk. So today, every startup doing differently. So for example, if you look at the figure AI, they have their own data training center. So, they will have remote simulation. They have somebody wear VR and then control the robotics remotely. And then that the robotics learn by capture the 3D data. And then if in China they will create a factory, they’ll have like 200 people just creating the data to fit into the robotics. And right now, today in in Japan, we cannot do states or in China. So, they are a lot of associations. So for example robot AI Association, they are trying to work with industry people to say, hey, can we share the data so we can use this data to create fundamental robotics language model.

Tim: And I guess this brings us back to that idea of that human form factor in robotics. So, we talked about figure AI and everyone right now seems to be focused on the human form factor. So, is this the way to go? Because I understand in one sense it would allow a certain amount of universal training data to be shared. On the other hand, like historically, this has not been a good approach. We’ve always made tools whether they’re robots or handheld tools to optimize it for a specific task. So, Elon Musk, I think just a couple weeks ago announced that in the future, 80% of Tesla’s value is going to come from these optimist robots. And figure AI’s got a market cap of, what is it, 39 billion which is pretty impressive since they’ve kind of shipped a beta robot or two. Is this the way to go?

Chiamin: As the Japanese startup, I don’t recommend we go for the similar path, as you just mentioned, Japan is in the good phase right now is we have case that we need to solve. What figure AI and now are doing, they are looking for maybe 10 years or 15 years from now that you will have robotics going to your houses, but that require huge amount of capital, that require the top tier talent to be able to work on that. If you look in Japan right now, my own thesis is we need to be more vertical. If you are vertical focused, then the amount of data you need is not as much as what they are doing right now. What you need to capture is industrial specific problem and what kind of data that you can store, what you can capture, what you can work with to be able to create your small robotics language model or your small VOM VOA and so on. And that is the way I will suggest to a lot of entrepreneurs looking to into.

Tim: Well, in Japan, historically has been very strong in that area. They’re still one of the world leaders in industrial robotics. But that the human form factor, I’m just curious on your take be, how much of it is this capturing the imagination versus a practical application? Because we were talking about the Asimo earlier and that was, I don’t know, mid late nineties. When did they put that out?

Chiamin: I think it was early 20. Well, but maybe 30, 20, 30 years ago.

Tim: It was a while ago.

Chiamin: It was while ago. I don’t exactly remember the day.

Tim: But we can go back to science fiction in the sixties and this was always, Robbie robot. This has always been kind of the dream, the aspiration, but even looking five, 10 years from where we are now, do you think it’s going to remain this aspirational dream?

Chiamin: I think the aspirational dream will remain there. But today in the reality, there’s already a lot of different type of form is working. You have the dot form, you have those, how do you say, the track, so that’s already on place right now, but I think why people want to go for humanoid is because the technology is raging to the point that you can reduce the production cost. And after LLM, what will be another, AI related, exciting topic that most of entrepreneur will get excited besides space in a way, then that will be okay. Maybe 10 years from now, we do have robotics working inside of a house and to really helping us. But my personal view, even 10 years from now, I don’t think it’s going to be replacing human, it’s going to working together with human in the space. For example, if you need to have a lifting or those hot summer in a way, you need somebody to really work on the construction. Maybe you can have a robotics to work 24 hours, but you still need humans permission. You probably also need to have people actually stand next to the robotics to really work together.

Tim: Well, I guess that is the trade off the human form factor gives you this theoretical flexibility that it could do anything a human could, but the optimized form factor gives this reliability efficiency more affordable.

Chiamin: Exactly. More affordable. Yeah.

Tim: Let’s get back to the innovation in a declining population and how physical AI plays into this. So in Japan, what are the best target markets, the target applications for that now?

Chiamin: Right now I look a lot in construction and also I think there will be a potential in logistic because you need to have people really on the ground to do things. For example, if it’s a logistic warehouse for picking and packing, I mean, some of those small thing, you can’t really ask the automated warehouse to do so. You still need to have a people doing that work. Construction is very obvious right now. A lot of construction side, they are locking people and it’s delaying. This will be, I think the first two industry. I think we do have a lot of potential there. And it’s easier for robotics to work with, but I think for long term, can we really apply robotics nursing home care, that’s the area. I think for the past 20 years, a lot of people are trying to do this and it never really succeed.

Tim: Yes. It is always like the first choice maybe because it’s just such an obvious problem people want to apply robotics to it. But yeah, as you say, it’s been 20 years, people have been throwing robots at elder care and it’s not sticking. What’s the problem? What keeps it from being adopted in elder care or in medicine more generally?

Chiamin: I think it’s urge of an emergency. People 20 years ago saying, we need to do something. But my take is we still have a lot of population that is working in that industry. But right now, if you compare the number of people need care compared to those people who want to get into that industry. You probably already know right now in the nursing care, we do have a lot of foreigner working in the industry because we don’t have enough people.

Tim: Particularly a lot of Filipinos have come in.

Chiamin: Exactly. And one of the biggest reason is those care company, they cannot offer high salary because of the healthcare system that we do have here in Japan. So, unless the regulation will change to say, hey, you can have a freedom to offer high salary that will be kind of create different dynamics. But under the current healthcare system, the revenue is kept. And if you’re going to hire people, it’s not going to work.

Tim: Well. That would make it ripe for automation.

Chiamin: Exactly. So, I think there’s a tipping point there.

Tim: Well, yes, but I mean, if you look at other, like surgical robots have made progress in Japan. Those have been adopted there, there definitely is a shortage of staff. I don’t know. I mean, to me, I think the fundamental problem with healthcare is that there is a certain human element that is at least as important, as the provision of physical care.

Chiamin: I think for your point, for example, if you think about surgical, it’s very easy to become automated because it is in the constrained environment. The procedure, you can actually design that very well. If you think about nursing care, it’s in the environment like your house and the problem, it really differ depending on the person. So, we don’t have yet that flexibility for the robotics to do that.

Tim: That’s a good point. It’s not just the human connection. There’s a real training data problem.

Chiamin: Yeah, exactly. You have a training data problem. And that’s why I think right now, all those human lawyers startup are trying to capture the data because you don’t really have the data. You don’t really have the trained situation there.

Tim: You know, it’s an area that I would love to see developed and if it gets developed anywhere, it’s going to get developed in Japan first.

Chiamin: I hope so. I really hope so.

Tim: I just think culturally right. It’s just Japanese people are just more culturally comfortable with robotics and automation than almost any other country I’ve…

Chiamin: However, Japanese is not comfortable of discussing the data. They can actually accept the robotic in-house, but then they need to accept that you need to give certain data for the robotics society.

Tim: I hope this works out at some point. Like I said, we’ve been different startups and different large industries have been working on it steadily for 20 years with sadly minimal commercial progress.

Chiamin: Yes. Yes. I think right now METI is recognized this needs, so they are trying to create some project to give the budget and so on, which is good. But I still believe that it’s not one company’s action. If we really want to move this forward, we probably need to have one team thinking in the way that okay, for example, assuming you and I, we are a competitor, it’s very easy to say, Hey, Tim, you work with startup A I’m going to work with startup B because we don’t want to. But I think that’s not going to create something. The important thing is how can we work together, give our data, but then when I say give the data, it doesn’t mean that we’re going to give our secret source as a data. It’s the 80% of the fundamental data. The more you’re sharing, the more you can actually create something that’s useful and then you get, ah, on top of it.

Tim: Yeah. I think you’re absolutely right. It is that sharing. And if you look at like significant innovation. There is always this cluster of innovators who are sharing ideas and a little data, whether it is, I mean, if you’re in generative AI, you’ve got to be in San Francisco today because there’s all these people sharing crazy ideas. Hardware, there’s a few clusters in China. Yes, absolutely. People sharing ideas and just bouncing problems off each other and the pace moves so fast. And I think there needs to be that ecosystem in Tokyo of a lot of different startups sharing ideas and bouncing ideas off each other.

Chiamin: I agree. I agree. I think the startup itself that we are sharing the data and the shared idea, but I think if you look at the B2B, and if you look in vertical, I think we do need to work closely with a large corporation. And then when you want to work with a large corporation, then you will really, depending on whether they are willing to do that together with their competitor or with startup, that’s one thing that I hope that it can move faster.

Tim: Yeah. I mean, I think so if you compare Japanese enterprise’s willingness to work with startups and embrace open innovation in general, they’re much more open now than they were 10 years, 20 years ago. But do you see Japanese enterprise opening up to being able to work in an open manner with each other in the way we’re describing? Or do you think this is something that the startups kind of have to do on their own?

Chiamin: I think it’s both. Or sometimes I think it’s even government need to do some push. I, I think one thing is whether the management there really feel there’s urgency there. In certain industry, they’re really waking up because they can see what’s happening right now, every single day. But in some industry, they might think, oh, okay, we still have a few years that that’s figured out. But once thing I noticed is that when a lot of a Japanese large corporation person thinking that, oh, we still have time, it doesn’t mean you, you have time.

Tim: I think that is so true. Japanese industry can be incredibly innovative. I mean, we saw it in the seventies and eighties, but also like when the digital camera came out, everyone knew, okay, Kodak and Fujifilm, their days were numbered, and Kodak basically failed. It was run into bank despite the fact that they basically invented the digital camera. Fujifilm, however reinvented itself and pivoted it moved into industrial chemicals and coatings and films. And so Japanese enterprise, even big enterprise can be very innovative, but they will exhaust every other possibility first before innovation.

Chiamin: I also think that if you look at the innovation history when Japan was shining, a lot of them is in manufacturing in the hardware. When the internet came in 1998, after that sort of everything slowly switched to software. And because of a bubble burst, Japan actually didn’t invest a lot in the software engineering. A lot of the software engineering, they are in the hardware component software. It’s not in with so-called AI or computer software engineer, that’s why I’m also excited to physical AI is because at least if we can do this well, combining the accumulation of knowledge in the hardware, but accepting the software AI talent and then put them together as one team, maybe we still have a chance to win.

Tim: I think so. It really does seem to be playing to Japan’s strengths and Japanese software programmers and software competency has done a lot of catching up over the last 20 years. And so Japan already has a lot of like industrial robots, but also in like agriculture, whether it’s planting rice and picking strawberries, there’s a lot of automation being deployed in a lot of experimentation. So it makes sense that this is a natural market. So, we’ve talked about healthcare, industrial robotics. What about consumer facing physical AI?

Chiamin: So far we don’t look at consumer facing AI but I think it’s a huge potential there. Consumer AI product right now is either coming from United States or China, and I personally wish that we do have a core team working on that, but I think it will be hard to catch up unless the core team has experience or creating a consumer product either in United States or China.

Tim: So, there’s a lot of really promising physical AI tech being developed in Japan. But are Japanese startups really reaching the global markets? I see a lot of talk about the advances in AI and the advances in robotics coming out of the US and out of China and even living in Japan. I don’t hear a lot about Japanese AI. Is this just a perception and marketing problem, or does Japan have a lot of catching up to do?

Chiamin: I think both for physical AI related startup, because they’re doing a lot of B2B business, they’re initial focus will probably be in Japan. But then if you look at United States or China, if you compare, it’s about the speed of the startup and also the capital that you can raise or not. One, it’s about the focus. Two is whether if they want to go global, we probably need to raise a lot of capital, but if you want to raise the capital, you probably need to build your story that is attractive enough for a global investor.

Tim: Yeah, I agree. I think robotics and AI, I don’t think there really are local markets for this. There really is just one global market and hardware in general, but robotics specifically require, I mean, it’s very capital intensive. So, looking at VCs in Japan compared to China, compared to the US, can physical AI startups get enough funding in Japan to mature and compete?

Chiamin: I think it could, because if you look at the recent train, every single VC is heavily looking deep tech. And we do have JAFCO they have a huge amount of fund size out there. So I think it’s easier than before. But can Japan VC provide exactly amount, just like a figure AI that I would say’s probably not possible.

Tim: Yeah. There are some outliers there.

Chiamin: There are some outliers.

Tim: But none, I mean, and okay, fine. If you have to raise a $2 billion round, you go to the US and that’s fine. But working to get a company to that stage, there’s a lot of seed money in Japan. You’re a seed fund yourself. It’s good the government’s supporting. But what I see is one advantage that Japan might have is that at least US VCs are very hesitant to fund hardware startups particularly early stage hardware startups. Is that something that’s different in Japan, or?

Chiamin: I think it used to be like that and they it’s probably still like that a little bit, but maybe I’m biased. When I become the GP of Firstlight Capital in 2022, I already invest a lot of hardware related company and at that time, not so many VC are looking into hardware. But this year, today, there’s a lot of noise in physical AI, in deep tech in hardware. And I think people transform their attention to be, how can we build a startup that can really create its mode or that can go global? Also because AI SaaS, everybody can invest.

Tim: Yeah. But you know VCs are trend followers. And 12 months ago everyone was a defense investor. And now everyone’s a robotics investor. I mean, how much of this do you think is going to stick because it requires some deep expertise to invest in this area.

Chiamin: Exactly. Well, I don’t know how long it’ll stick. My personal belief is that this is not a training investment you can do. Once you invest, this is going to be minimum five years to reach to the point. But I also don’t think that it’s a good strategy to create your own hardware in Japan. I don’t think hardware is necessary. You can always buy hardware from China. The important thing is the software inside of hardware. And also the important thing is whether you have your own model that created in a specific sector.

Tim: But doesn’t that kind of go against what you were saying before, it was Japan’s unique strength. If we’re assuming the hardware’s coming from China, then we’re back to a pure software play. And does Japan have unique advantages in a pure software?

Chiamin: When I say you don’t do mass production, it means that even you buy hardware, you still need to do tweaking. You still need to adjustment, you still need to modify a lot of the things inside. I think that Japan has its own strength, but if you think about you want to build a factory here in Japan and then make mass production, that will require a lot of capital intensive investment. To echo back to what you said, if we are early stage investor, are we going to invest in the stuff and say, Hey, I’m going to create a hardware and with the hardware is so expensive, then it will be difficult. But if you can actually build a product that in day one, you really don’t need to have that heavy, expensive robot, but you are building something that you can just deploy in the human way robots, I think it makes sense.

Tim: That’s the problem all hardware startups face is that, but it seems with Japan having such a strength in industrial robotics and robotics in general, and with the enterprises being more open to open innovation, that there’s a real chance for these startups to partner.

Chiamin: Yes, yes, of course. But I will have to add that. If I look at this industry, you’ll see there’s a two camp. One camp is from emerge from AI. It’s the same in China or in United States. It’s those people who work in the gen AI industry. There’s another camp that is those people who work in hardware, 20 years in the industrial robots. Those people, they are not overlap with the gene AI people and then this camp is more from hardware. And then say, we probably need to have gen AI, which camp will come out the good result, we don’t know.

Tim: Well, going back to the innovator’s dilemma, we predict the new LLM new approach is going to be the one that wins out in the long run.

Chiamin: Exactly. But then if these two are not working together, if they are divided, this is not going to go anywhere.

Tim: That’s a really good point. Yeah.

Chiamin: And what I’m really afraid is that I hope this two camp can hold their hand and then try to figure out something together.

Tim: Well, with recently the US becoming more isolationists with a lot of these tariffs being imposed on China, do you think this is an opportunity for Japanese manufacturing, Japanese robotics? Or do you think this is more of a temporary situation?

Chiamin: I think it’s opportunity, but also it’s such opportunity for China. So whether we can run fast enough to do that. What I learned when I work in China is once they have a constraint, they are much, much more innovative.

Tim: Oh yeah, for sure. We saw that with deep seek. Just incredibly clever. Well listen, Chiamin, before I let you go, I want to ask you to pull out your crystal ball gaze into a minute and tell me what you think physical AI in Japan is going to look like in three to five years. What are we going to see?

Chiamin: Wow, it’s really hard question. I think the bad case that I don’t want to see is five years from now, we do have humanoid robotics working in a lot of different kind of industry, but they made it outside of Japan. The worst case is what we don’t utilize Japan’s knowledge capture but it gets utilized outside of Japan, not as the Japanese assets. That’s the thing I don’t want to see.

Tim: Well, I think that’s a real concern because right now that kind of looks like where we’re headed, because in that humanoid form factor, the US and China’s quite a bit ahead of Japan.
Chiamin: Right. That’s the thing I don’t want to see. So, what I really would love to see is at least in one industry, for example, construction or manufacturer, we do have our local brand that is working inside of the factory with the construction side and then is exporting outside of Japan. I’m not talking about the hardware needs to be made in Japan. Just like we have a lot of things made in China, but that’s controlled by the Japanese company. I think that’s okay. But the fundamental thing is the brain, the controller, the action, the intelligence itself, generated…

Tim: The real value add.

Chiamin: The real value added, is based on the Japanese philosophy in a way.

Tim: Do you think that breakthrough is going to come through the humanoid form factor or do you think it’ll be more dedicated?

Chiamin: If I’m going to project three to five, it’s not going to be a human factor. I think it’s going to be other form, but really solving the problem

Tim: Well, I think Japan is well suited to do that. Excellent. Well, I will get back to you in like three years and we’ll see how that gets in.

Chiamin: Sure. Hopefully. Hopefully I was correct.

Tim: Well, Chiamin, thank you so much for sitting down. This is really fun.
Chiamin: Thank you. Thank you.

Outtro

And we are back.

I have always been a student of the history of the future, and one of the things that is never changing is that new technology is always promising to let us create automatons in our image.

In the 1700s, electricity’s ability to cause dead muscle tissue to twitch and to contract resulted in many declaring it to be the spark of life. It kicked off a wave of research in human and animal automatons. That was the basis for Mary Shelley’s Frankenstein.

In the late 1800s, entrepreneurs promised to use this new technology of steam power to develop human-like robots that would perform human tasks and investors rushed into finance them. The steam man of Newark drew widespread acclaim and praise from the media of the day when it was first announced.

In the 50s, the new technology was computers, which we were told worked just like the human brain. And that was finally going to give us our humanoid robots. And you know, that vision that was laid out in the fifties and sixties, that still largely defines how we see humanoid robots today. But today, of course, it’s LLMs, which we are told work just like the human brain that is finally going to give us our humanoid robots.

It all sounds very familiar, but you know, I’m going to go against all historical precedent and say that this time it looks like we really are going to get functional humanoid robots.

And yes, yes, everyone going back to the 1700s has always thought that way, but sometimes this time really is different.

Unfortunately, we won’t see widespread adoption of humanoid robots once they arrive. Not for technological reasons, but for economic ones. Chiamin and I talked about the human form factor versus functional form factors. Forms designed for a specific purpose, whether that be cleaning a house or assembling a smartphone. And even in the long term economic forces will ensure that functional form factors will win out over humanoid form factors.

And here’s why.

Now it is easy and frankly somewhat lazy to look at a task and say, I wish a robot could do that. And then imagine a robot that looks and acts kind of like a human being doing that task, but that’s not how technology moves forward. That has never been how technology moves forward.

Whatever the task in question or the form factor of the machine, it takes time and money to train a robot to do that task. And once that task is automated, there is immediate pressure to optimize it. Other processes are designed around it. The robot itself is made less complex to decrease costs and increase efficiency. I mean, you don’t give a robot two arms when one will do. You don’t make it bipedal when you can put it on wheels or on a track.

Humanoid robots will be pushed out of every market almost as soon as that market opens up.

The counter argument to this is that humanoid robots will be able to perform many different kinds of tasks with the same hardware. Now, this might be true, but it’s not really economically viable. It still takes extensive training for a robot to become proficient in any given task. And once humanoid robots open the market and the market begins to grow, then others will take it from them with less expensive, more efficient custom form factors.

Using technology design for specific tasks rather than general purpose tools is always how it plays out from factory automation to, well, all the way back to our first stone tools.

The technology needed for humanoid robots is pretty solid now, but the business model is still pretty shaky. And yet this time does seem different. We are probably going to be interacting with real humanoid robots in the next five or 10 years.

But, you know, some part of me kind of hopes we won’t be. Looking back over the past 400 years, the dream of creating machines that looked and act like us and the scientific and creative and even artistic efforts that striving for that creation has involved, has probably advanced mankind far more than actual working humanoid robots ever will.

 

 

If you want to talk more about physical AI in Japan, Chiamin and I would love to talk with you. So come by disruptingjapan.com/show244 and let’s talk about it. And if you enjoy disrupting Japan, share a link online or just, you know, tell people about it. Disrupting Japan is free forever and letting people know about is the absolute best way you can support the podcast.

But most of all, thanks for listening. And thank you for letting people interested in Japanese startups and VCs know about the show.

I’m Tim Romero and thanks for listening to Disrupting Japan.