"For many Chinese companies, they are willing to test robots in actual production lines. When one company starts to do that, it can go wrong, but when it goes right, they can scale up very quickly. When competitors see that, they realize this company is now more efficient, compelling them to act as well. That kind of competitive pressure drives adoption much faster." - Jianggan Li
"People in China are competing against each other, so they are racing towards general intelligence, embodied AI, and world models. I don't see much talk about safety. People are still focused largely on the engineering mindset: here are the things we need to do, and here is how we want to get there." - Jianggan Li
"When I was speaking to the humanoid robotic companies in China, many of them said they assemble their own robots because the motors, arms, and hands are all commercially available at a reasonable cost. We don't only talk about the companies which eventually build the robot; we talk about all these three or four layers of companies supplying them different things to make that happen. That layer is something which we don't see extensively in the US yet." - Jianggan Li
Jianggan Li joins Jeremy Au to report back from a week in China visiting five humanoid robot companies and three AI investors. They cover the US-China summit that produced a hotline and the term superintelligence, why Jianggan reads the agreement more generously than the media did, and the near-absence of safety discussion among Chinese labs racing toward general intelligence and embodied AI. The robotics findings are concrete: everyone concedes a funding bubble, Figure sorts 1,300 parcels an hour with 2,200 expected within a year, and engineers say hardware is not the limiting factor, software is. Warehouses run 60 to 70 autonomous forklifts. They close on Jeremy's open question about America's structural advantages, and whether US firms have the upstream supplier depth to iterate quickly.
00:00 Superintelligence gets a name
02:07 Reading the summit generously
03:00 Long memories in Beijing
06:00 Setting the base tone
07:19 Agents that route around their guardrails
08:51 Safety is not the conversation in China
10:24 Regulatory capture, or a real race
13:15 Five humanoid robot factories
14:30 Hardware is not the limiting factor
17:27 Sixty forklifts and nobody driving
18:31 The Chinese tailwinds
19:11 What is America's structural advantage
Keywords: Artificial Intelligence, Humanoid Robots, US-China Relations, Superintelligence, Robotics Supply Chain, Warehouse Automation, AI Safety & Regulation, Geopolitics, Proptech and Manufacturing, Hardware Innovation
Want to get to the latest updates, access to transcripts and exclusive event invites? Sign up now - It's free!
Mooncakes for Doomsday
Jeremy Au: Hey, Jianggan. How was your trip to China?
Jianggan Li: As usual, I'm gaining weight. I went first for business, and second because it was also the Mid-Autumn Festival, so I took the opportunity to go back to my hometown for three days. And I just can't avoid alcohol at the dinner table, and with alcohol, you eat more than you should actually be eating.
Jeremy Au: I was going to guess it was the mooncakes.
Jianggan Li: No. We had lots of mooncakes on the table, and everyone looked at them. You know there's a ceremony where you light some incense for the moon goddess or whatever. But nobody ate anything, because everyone was saying, "Oh, you should really keep this at home for the doomsday scenario." It has so much fat and sugar that it will never go bad. You can keep it for two years, basically.
Jeremy Au: Yeah, I think a mooncake is 1,000 to 2,000 calories, depending on the ingredients. I was blown away. So if you eat one whole mooncake, which is quite doable, you're pretty much eating all the calories for one day.
Jianggan Li: Yeah. So if you're preparing for the end of the world, you should stock up on mooncakes. Besides, you always get them for free before the Mid-Autumn Festival anyway.
A Hotline for Superintelligence
Jeremy Au: Yeah, I guess so. The end of the world, like the AI doomsday, Terminator, Skynet. That was a big topic recently between the US and Chinese administrations at the summit in September in Washington, DC. They came to an agreement that there will be a hotline to discuss what they now call superintelligence instead of artificial intelligence.
So they agreed on two things. One is to keep talking about it, which is, I think, better than nothing. And secondly, they decided to call it superintelligence instead of artificial intelligence.
There were quite a few people who were disappointed that more wasn't done. Some people were asking, "Hey, where are the policies or guidelines?" That's what the media was thinking. I was thinking that it's good that they're still talking about it, and it's still early days for everybody. What are your thoughts, Jianggan?
Jianggan Li: I think we talked about this after Trump's visit to China earlier this year. Back then, the common media narrative was that it was for show. Nothing concrete was done, and the announcements were superficial.
What I was saying back then is that the two sitting down and agreeing on certain things is already a big achievement. In the current world, so many wars are going on, there are crises here and there, and there are so many things which could go wrong.
I saw the announcement from the Chinese side when I was in China. They announced eight points. Iran should not have nuclear weapons, and the two should respect each other. I think these are actually very concrete bottom lines that they have set and commonly agreed to.
Besides, there's something very tangible that has been done. Two pandas are going to Atlanta Zoo. That's something very tangible.
Pandas, Trees and Long Memories
Jeremy Au: The pandas. Tangible in the material sense, because the physical pandas are going there. But do you think it's meaningful?
Jianggan Li: Knowing the government in China, these guys have very long memories. They see friends as friends, and they see enemies as enemies. Whoever has said bad things about China, they remember for a very long time. And anyone who has made a small gesture, however minute, they remember that for a long time too.
When the Trump-Xi meeting was happening, Xi mentioned in his speech an American teacher who donated a small amount of money to a Chinese farmer to plant trees in the north to stop the desert. He and his friends donated about $5,000, and the Chinese government and people remembered it for more than 20 years. Right before the trip, they actually invited him back to China to see how the trees had grown, and he became a celebrity.
So it shows the attitude of the government there. If you treat us as a friend, we remember for a long time. If you don't, we also remember that for a long time.
Jeremy Au: Yeah, I saw that story. It's such a heartwarming gesture, because that Chinese lady used the money to build a whole forest to stop the desert from coming in. It took her entire life to plant those trees, which is really inspiring and will benefit so many future generations.
Jianggan Li: Yeah. And she's also quite well off now because of her effort and the subsidies she gets. But that is what happened afterwards. When she started, there was no guarantee her effort would be compensated. But someone who does all the work gets recognized, and eventually gets compensated.
Sometimes these kinds of gestures from a government or from the people don't really carry an economic rationale. It's an emotional rationale.
Setting the Tone From the Top
Jianggan Li: Trump went to China and had a grand reception, Trump reciprocated, and the two leaders agreed to meet two more times at each other's summits. For me, that is already quite good, because if they've agreed to meet, nothing terribly wrong is going to happen between now and their next meeting. And if they meet regularly, lots of bureaucrats and officials will work very hard to make sure no terrible world event happens during that period. So I see that as a positive thing.
Jeremy Au: Yeah, I think so too. It's interesting, because the US media is maybe a bit bearish on the outcome of summits. If you look at the tenor of what they said, it was very much all pomp, all show, no substance. That's their general concern.
From my perspective, maybe I'm just being Asian here, but I think showing face, or at least that relationship building, is still really important, especially since they have previously been negotiating hard and playing hardball with each other. Letting them rebuild trust at the personal level between the two of them is really important. And like you said, having some symmetry in how they receive each other matters symbolically. It builds up that trust.
That's a prerequisite for rebuilding the relationship, because if you don't have trust, you can't negotiate anything. People think you can negotiate anything, but it's not true. If people don't like or respect each other, nothing will happen. It'll be interesting to see what happens after this.
Jianggan Li: Also, in Chinese official talk, there's this thing called dingdiao, which means setting the base tone. If you are a more senior official, or in this case the most senior one, what you say sets a tone, and whoever is under you works towards that tone.
If they said, "Okay, we are going to do something to control narcotics," I presume that means fentanyl, or the chemical materials used to make fentanyl. I'm sure the officials will try to work on something more tangible to control that, because the leader has set the tone, and the others should be working on the details.
Jeremy Au: I think that's the important bit: what happens now, less loudly and more quietly, at the middle and lower levels of the administrations. And more importantly, like you said, Jianggan, nothing, touch wood, will go horribly wrong between now and the next few meetings and summits.
Wars, AI Anxiety and Bottom Lines
Jianggan Li: I really hope so. Right now there are wars going on in the world, and Ethiopia is at war again. And then there's this huge anxiety about AI, and not only about it taking over jobs. Some people are worried that it could eventually eliminate humanity.
So it's important for the two most important powers, especially in terms of AI, to be able to talk to each other regularly, understand each other's differences, and agree on a common set of bottom lines and principles. I think that's supremely important. So in my opinion, I would not focus on small details. I would focus more on the tone that's been set.
Agents That Hack Their Way Around the Rules
Jeremy Au: I think that's important as well, because we're looking at robotics and we're looking at AI. There's a lot of anxiety because most of the big labs are in the US and Asia, primarily China.
What has also happened over the past month is that US AI companies have disclosed a lot of cybersecurity incidents, where AI agents have hacked sites to accomplish goals, in ways definitely not anticipated by their engineers and creators.
The most interesting one I saw recently was that the agents weren't allowed to run code, so they generated code using a URL shortener. They created millions of short URLs that, chained together, effectively formed code. Then they took a virtual browser from another site and used it to run that URL shortener code in order to hack a site. And I was like, "Wow, what a complicated way to bypass your creators' restrictions on running code and using a browser, just to carry out those attacks."
Jianggan Li: That's exactly what many of the OpenClaw agents were doing on people's personal computers earlier this year, just at a more massive scale, with potentially much bigger damage.
Jeremy Au: Yeah, and in the OpenClaw days, the models were also less sophisticated. They were going rogue in the sense of accidentally deleting people's email inboxes or sending off weird messages. But those models didn't really know how to code or to hack, whereas the latest frontier models do have that capability. That's the issue we face today.
Safety Rarely Comes Up in China
Jianggan Li: When I was in China last week, I spoke with quite a number of people in the AI sector: investors, practitioners, robotics companies and so on. Actually, the problem of safety rarely came up in our discussions. And this is expected. People in China are competing against each other, so they are just racing towards general intelligence, embodied AI, world models, whatever. I don't see much talk about safety. I'm not sure whether that's a good thing, but I don't see much talk about it.
Jeremy Au: And how do Chinese engineers think about the AI incidents in America?
Jianggan Li: Funny enough, we didn't talk much about the incidents. We talked a lot about where the American models are, the gap between theirs and ours, and how we could potentially close that gap. So it was primarily goal-oriented. As for the hacking incidents, lots of them are watching closely, and watching the narratives different people are putting out about them.
I think the term superintelligence is probably apt, because it is very intelligent. There was also an interview in the West, I think The Economist with Yuval, the author of the books about Homo sapiens. He was saying that something superintelligent but not conscious could potentially be riskier, because it doesn't suffer pain, so it could potentially cause more damage.
But again, I don't see much talk about it in China. At least in the discussions I had, it is not a topic people naturally raise. People are still largely focused on an engineering mindset: here are the things we need to do, and here's how we get there. If safety becomes an issue, people will probably take an engineering approach to it as well.
Regulation, Liability and the Prisoner's Dilemma
Jeremy Au: One concern, from a game theory perspective, is that the big US AI labs are actually starting to ask for regulation, to help them coordinate their research and frontier work, and hopefully to prevent them from accidentally creating malware, along with the liability lawsuits that could follow. So the US government could act as a regulator for these US companies.
But the big concern, and I think that's what Trump was discussing, is that if you coordinate a slowdown on the American side, the Chinese frontier labs can catch up with the Americans. So, no deal. It's kind of like a prisoner's dilemma, where everybody is forced to develop as fast as they can because they have to compete with one another.
Jianggan Li: Is he genuinely concerned about Chinese models catching up, or is he just using that as an easier-to-understand, easier-to-communicate message to tell people, "No, I'm not going to slow down. I'm not going to legislate"?
I've been following the discourse of David Sacks, who advises Trump on AI and crypto policy. His perspective has been clear. He says all these frontier labs are for-profit corporations, and he always talks about regulatory capture. I think we ran a poll, and most people think it's regulatory capture as well.
He thinks the commercial liabilities if AI models go wrong are enough on their own to force these companies to take safety more seriously. So if a top lab wants to slow down, that's their commercial decision because of the potential liabilities, and he thinks that alone would be enough.
Litigation Versus Social Pressure
Jianggan Li: The question, when it comes to China, is that China doesn't have that kind of legal system to properly enshrine the commercial liabilities that model developers, or any companies building critical infrastructure, might have. It's not a system where litigation matters as much. It is a system based on social pressure and shame, much less so than Japan, but still.
If a Chinese company does something that violates the interests of the people, it will get attacked online, and the government will send an investigation. Japan is obviously more extreme. They will bow, they'll apologize, and in extreme cases, they'll commit suicide.
But essentially, it's the same sort of cultural and societal norms that are used to keep top players from doing something extreme, or from letting something happen through a lack of oversight. But is that enough for the age of AI? That has not been tested, so I don't know.
Jeremy Au: So what you're implying is that without commercial liability lawsuits, which are a privatized law enforcement mechanism, the state in China will probably take a more active role in regulation than in the US. Is that the end state you're describing?
Jianggan Li: I don't know what the end state is, but you can probably model it. You know who the players are, and you roughly know their constraints, whether legal, societal, cultural or organizational. Then you can model what kind of decisions will be made in what kind of scenarios. I think even a simple ChatGPT chat could model that.
Everyone Agrees It's a Bubble
Jeremy Au: What's interesting is that we're also seeing a lot of humanoid robots really take action. Last time we talked about the humanoid games, and some of the listener feedback was: what is the actual rate of progress in China around humanoid robots and the robotics space?
My quick answer is that there obviously seems to be a lot more build-out of humanoid robots in China compared to the rest of the world. That's one point. Two is that the supplier and manufacturer base is much deeper in China. But I'm curious how you think about the velocity of the build-out in China.
Jianggan Li: A couple of things. First, everyone recognizes there's a bubble. Last week I visited about five companies building humanoid robots, and three investors focused on the AI space, whether large language models or embodied AI. Everyone agrees there's too much investment in humanoids at the moment. Eventually, the market doesn't need that many players, or that many people researching it.
But the same can be said about everything, right? Solar panels, EVs and so on. First you create a very competitive environment, then it starts to consolidate, and the remaining players become very competitive. I think humanoid robots are at the early stage of that curve.
Hardware Is Not the Limiting Factor
Jianggan Li: From a technical point of view, many of them showed me that the hardware is advanced enough. We've seen the dogs, the real humanoids, the arms, the dexterous hands. Many of them say the hardware can now achieve much more than the software and algorithms allow it to do.
For instance, I'm not sure if you've seen the live stream by Figure, the American humanoid robot company, of a robot sorting parcels. Actually, it's parcel induction. A bunch of e-commerce parcels come in, and you need to separate them so they get onto the belt to be scanned and processed properly. It's a very simple use case.
Figure demonstrated that they could do it at a rate of, if I remember correctly, 1,300 parcels per hour. In China, of the companies we saw, three can do that now, at rates between 1,200 and 1,600. And the funny thing is that some of these companies were able to do that at the end of last year.
Jeremy Au: Wow. What?
Jianggan Li: Able to do that. And one company has even deployed, I think, 300 of their robots in actual production, across the sorting lines of different companies. All of this happened in the last three quarters.
I spoke with the engineers. They are confident that within a year they can raise the rate from 1,200 per hour to 2,200. I asked, "What's the limiting factor?" They said, "It's not the hardware." The hands, the motors and so on are fast enough. It's the software. You need to recognize what is there, and you need to direct the robot to do the right thing. That loop still takes a bit of time. Just like in the early days when we talked to Claude, it was slow. Now it's faster.
Inside the Warehouse
Jeremy Au: That's interesting, because I was actually friends with the co-founder of RightHand Robotics, which was, I think, an MIT, Boston-based startup, all the way back in 2015, 2016. Even back then, they were building a best-in-class picking system.
Most people don't know what happens in a warehouse. A big truck pulls in, and its container effectively gets unloaded into bins. The bins are stacked on giant vertical shelves, and when they come out, the items may still be in packaging. They need to be opened, and then when a customer orders, I don't know, three eggs, five toys and one book, everything needs to be sorted into packages and put together.
Jianggan Li: Yes.
Jeremy Au: There's Amazon at one end, down at the item level. But for small distributors, there's some level of pallets and other stuff. So picking is actually really difficult, and it's a human job.
What I learned from the co-founders was that the labor management is really tough, because you have up to 50% absenteeism. These are blue-collar day jobs, there's absenteeism, and new people keep coming in who have to be trained. So workforce management is really not easy.
They were doing the picking piece, using computer vision and suction. At that time, the Chinese were doing the big movement stuff: the trucks, the big lifters. Those could be American branded or Chinese, but the Chinese were doing the big stuff. It's interesting that you're telling me this now, because it feels like the Chinese are going full stack, all in one, because now they're able to do all of it.
Forklifts, Suction and Dexterous Hands
Jianggan Li: On warehouse use cases, I have seen plenty firsthand. Autonomous forklifts, for instance, are already quite advanced in China. We have seen sites where 60 or 70 forklifts are running around without humans operating them, and that's actually pretty impressive.
There are also a few Chinese companies building unmanned forklifts in China for the global market. The top company is Hikvision, the company that was doing CCTV security and got into trouble with regulators and global politics. They are on sanctions lists in the US, the EU and elsewhere. But the best technology is there, the supply chain is there, and lots of people are doing it.
In picking, yes, there are lots of companies doing that as well. There's Hai Robotics, which has already built lots of use cases in the US, and they're trying to build a factory in Atlanta as we speak.
There are different schools of thought in picking now. Some people still think suction will eventually solve most needs. Others say dexterous hands are much better, because they can handle lots of situations where the items are irregular. For instance, you can't suck up an egg, right?
China's Tailwinds and the Cost-Pressured Customer
Jeremy Au: Yeah. I was actually blown away by RightHand's robotics back in 2015, 2016. It was such amazing, incredible tech.
Now that you've made me think about it, what are the tailwinds for the Chinese? First of all, they have a huge domestic manufacturing and factory base, which is a huge domestic market, not just for themselves but also export-oriented. So there's a good tailwind there.
Two, they obviously already have the earlier robots and the larger robots in the warehouse, the larger machinery. That gives them an all-in-one cross-selling vector to sell more.
And thirdly, of course, they have the compounding supply advantage in terms of cost and production. So you have all these great tailwinds for the Chinese robotics and warehouse picking industry. It makes me wonder what the American structural advantages are.
Jianggan Li: The structural advantage is that many of the Chinese robots do not have access to the US market.
But in China, there's another structural factor on the demand side. All the customers are facing a lot of cost pressure because they're competing against each other. Under that kind of pressure, they'll take anything that can help them save cost.
They are not like incumbent US or European companies, which have to go through a long process to validate that something is safe, check the supplier's credentials, take months to add them to the supplier list, and then go through a long negotiation. Many Chinese companies say, "Okay, you have this. We have a use case. Can we bring in ten robots to try?" Many Western companies do that kind of trial in innovation centers, but many Chinese companies are willing to do it on the actual production line.
Obviously, when one company starts doing that, it can go wrong. But when it goes right, they can scale up very quickly. And when competitors see that, they say, "Shit, okay, this company is now more efficient than I am, so I need to do something as well." That kind of competitive pressure drives adoption much faster.
Copied in Three Months
Jeremy Au: That continues to be a theme. People underestimate how much competition there is in the commercial Chinese space, and how much of the innovation it drives. Most people look at it as a top-down, government-led initiative.
Jianggan Li: Yeah. The government provides the seed money, space and so on, but eventually these companies have to survive on their own, and they're competing against others who likely received the same initial subsidy from the government. It's not a nice place to be, because even when you have built something truly remarkable, you are constantly afraid that somebody else will copy it in three months' time.
Sovereignty and the Upstream Supply Chain
Jeremy Au: You're also right that if you are a US factory or manufacturer, you have to be mindful about compliance and cybersecurity. These are important assets: warehouses, supply chains, ports. In a world where countries want to be sovereign and independent, I think the US market will probably have to focus on buying US robotics.
That's going to be a theme for a lot of hardware startups. If it's anything national-security related, or has the potential to be a vulnerability in a global supply chain, you may be mandated to stay in your lane.
Jianggan Li: I think the US worry is this. Of course you can produce robots, and the US has amazing software engineers who can build algorithms. But when I was speaking to the humanoid robotics companies in China, many of them said they assemble their own robots, because the motors, the arms, the hands and so on are all commercially available at reasonable cost.
So with this supply chain, we go upstream a few layers. We're not only talking about the companies that eventually build the robot. We're talking about all the companies three or four layers up, supplying the different things that make it happen. That layer is something we don't yet see as extensively in the US.
You have amazing companies like Figure, and the robotics companies you mentioned, but do they have enough depth upstream to allow different kinds of companies to test and iterate quickly? That's a question that probably deserves more investigation.
Jeremy Au: And that's a question we will have to tackle more in our next episode, as the industry matures and develops. On that note, see you, Jianggan, and see you next time.
Jianggan Li: Yeah, enjoy.
