Katherine Forrest: Yeah, you know, so first of all, I want to do two things. One, I want to mention the name of the report that you had just referenced, which was the May 2025 CSET report, also from the same Georgetown Center for Security and Emerging Technology. And it's called, “Wuhan's AI Development: China's Alternative Springboard to Artificial General Intelligence.” And that's again, May 2025. So I wanna sort of make sure we have that. And then I wanna do something else, which is talk a little bit about the connection between embodiment and AGI. ’Cause you know, I'm a big fan of this and we've talked about Jan LeCun on a prior episode and world models and there's a relationship here. So there is a theory within the AI research community that artificial general intelligence will take some form of embodiment to actually be fully realized. Why? Because AI otherwise is learning through words, through pictures, through different kinds of multimodal data, but it's not learning necessarily the laws of physics as humans do through moving in the world. And so there's a concept that embodied AI, which would be, let's just say, all kinds of robotics, and there are lots and lots of different kinds of robotics. It's not just humanoid robotics or quadruped robotics, and we did an episode on robotics, but it's lots of different devices that move and experience the world in a real way. And the concept is that if you've got embodied AI that pushes over a... vase, which I call a vase and I don't call it a voss. I just never have. Okay. I just want to say that. And it pushes over a vase and realizes that the vase, if it's made out of a breakable material like ceramic or glass, is going to break. That that is then a learned moment. So rather than just hearing or reading or watching it on a movie that something breaks, sort of actually experiencing the physics of it is a different kind of experience that then gets brought into the information repository of whatever robotic embodiment that is and into the AI that is within that robotic embodiment. And there is a different kind of learned behavior. So, you know, we know that China is very interested in embodiment and there has been, you know, a fair amount of speculation about what they're doing about it.
Scott Caravello: Right, and so then turning back to that first CSET report that we mentioned, and it explains that compared to the US's very LLM-centric approach to AGI, China has made a broader bet, and that includes, but isn't limited to, embodied AI. But the report is looking at embodied AI and looking at statements from researchers in China, along with measures and initiatives that are being taken at the national and local government levels, to show just how widely shared that view is, that embodied AI is a key potential path to explore on the road to AGI. So, I think, I do just want to sort of draw that out a bit more, which is just that it's still one of a number of bets that China is placing on AGI, but it's an important one that they're taking seriously.
Katherine Forrest: Yeah, and that's one that actually, you know, some others, as we'd mentioned before, like Jan LeCun and some others are actually also thinking that there needs to be a kind of a way in which a world model gets embedded into AI in order for it to fully experience, through a world model, AGI. Now, one thing I do want to say, as a little aside, is that there are very, very likely many different ways to actually reach AGI. And when you reach AGI, there's going to be a spectrum of where you are on that, sort of, road of AGI. The road could be a very long road, and you could be at the first point on that road where you could hit AGI and not be deeply into AGI. You might not be as far along as another AI model. So not all AI models may reach AGI at the same time, and they may not all reach it in the same way. So it's possible that you don't need embodiment for AGI. That's one of the bets that many of the high capability LLM developers are making, although some of them also have various kinds of embodiments. We know that there are world models which now are emergent capabilities within even LLMs. So I don't want to suggest through this discussion today that embodiment has to be the way to get to AGI. So let's talk about, you know, the way in which the LLMs do it now.
Scott Caravello: Sure. So I can dive into that a little bit quickly, Katherine, which is just that language for LLMs is a depiction of the world and it's a representation. And the idea that you scale it up and you increase the reasoning, the abstraction is going to help you get to AGI. And then for embodied AI, on the other hand, the idea is that human intelligence is a product of interaction between the brain, the body and the environment. And the physical element is a key component of our intelligence. And so the idea is that it's therefore going to be a part of achieving artificial general intelligence that approximates our abilities. But, so, like you mentioned, Katherine, though, it's not like LLMs, you know, have no role—that LLMs and world models need to be integrated into a physical component to get to AGI. But, again, that is just one school of thought.
Katherine Forrest: Yeah, you know, it's really it's a philosophical debate and one of the things that I say in my forthcoming book on superintelligence…
Scott Caravello: There we go!
Katherine Forrest: … called, “Of Another Mind,” soon to be out for pre-order, uh, one of the things that I say is that, you know, we humans in different parts of the world and on our earth learn an awful lot about the way the world works in terms of physics. And in terms of, for instance, the way that, you know, gravity plays with waves to make the waves in our ocean and the waves in the air, that there are lots of things that we can learn about that we can't see even. And so that's one of the reasons why I, personally, think that embodiment is not necessarily the be all and the end all. But, in any event, moving back to the concept that CSET is talking about, because it's very, very important. So, the CSET report tracks initiatives from various levels of the government of China—what's clear is that China is taking this incredibly seriously. It's taking embodiment and the relationship of embodiment to AGI very seriously. And one example in the CSET report is, at the municipal level, there's a three-year plan from a Beijing administrative body to facilitate, you know, this particular technology's development, but also that the Central Committee of the Chinese Communist Party has identified embodied AI as a target for accelerated development. So there's a lot of money, a lot of research going into this.
Scott Caravello: Right, and that target was set at the fourth plenum in October, which was focused on drafting the country's 15th five-year plan. So as CSET notes, inclusion of embodied AI really signals its importance in central leadership in China.
Katherine Forrest: Right, and there are many others too, which we don't have time to get into, but the report identifies the numerous research and innovation centers around China that are working on embodied AI and particularly forms of embodiment that would fall under what we would call the rubric of robotics. And it highlights how in higher education in China, embodiment has been embraced as part of AI technical development, and there are eight, now, top universities that have announced plans to offer a major in that.
Scott Caravello: Yeah, and, you know, that gives me another idea for an episode, Katherine—which is discussing how AI is being folded, more broadly, into higher education in the United States—because I’ve literally no idea if there are similar programs in the United States that are focused, you know, on sub-sectors of the technology.
Katherine Forrest: I am gonna write that down!
Scott Caravello: Ok—great, great. And, going back to the report, it underscores the breadth of the country’s approach—from industry to research to education to those government planning and initiatives—but, I think, with that, we can segway into the other CSET report that we had mentioned, which, again, is titled, “Wuhan’s AI Development: China’s Alternative Springboard to Artificial General Intelligence,” and, again, that’s from May 2025, but that describes a project taking place in Wuhan to roll out what’s called a large-scale social simulator, and infuse AI throughout the city, and that’s being done by bringing together the Chinese Academy of Sciences’ Institute of Automation, Huawei, the technology company, and Peking University. And, so, according to CSET’s report, that social simulator will, “simulate multi-scale, physical-world, and social complex systems,” on the path to becoming the operating model of a co-existing smart society, and a tool for social governance.
Katherine Forrest: Yeah, it's interesting. I–you know, you have to, sort of, wonder what all that means and how it's been going. So, while that initiative doesn't seem to be about embodiment in a robotic sense, it is about infusing AI into a real environment throughout a city and learning, based upon interactions, how AI learns from that environment and how humans, actually, may interact and learn from the AI. So, it's really, uh, if we can get more information on it as it's gone through some sort of evolution and a timeline over there, it'll be very interesting to look at.
Scott Caravello: Right, and, so, there is an AGI angle here, and, so, I don’t want to miss that, but the idea is that the rollout of this type of system, which is taking data from all over the city, will give AI opportunities to become more intelligent and, thereby, promote its evolution into AGI. The data that the platform is getting from the city is, you know, real data that’s grounded in the physical world. And, so, the idea of being that contrasts with the language we’re talking about, that’s a depiction of the physical world. And, so, this also just highlights what we were saying earlier… that the country is really diversifying its bets in pursuit of AGI.
Katherine Forrest: Yeah, and, you know, one thing that's worth touching on, which is really sort of an aside, but important, is that there have been reports at the end of 2025, in major U.S. established news outlets, that China has constructed the biggest power grid ever seen, generating twice as much electricity as the United States—at least in a certain area. And so this means, for instance, that China data centers could be supplied in different ways and their, of course, cost structure is always different from what we have in the United States.
Scott Caravello: Yeah, and they are anticipated to spend more than half a trillion on their grid over–over the next five years. And, then, in 2028, China’s expected to operationalize a nationally interconnected grid of data centers, which is, you know, referred to as a national cloud.
Katherine Forrest: Yeah, you know, this is what OpenAI in the past has called an electron advantage, that computing is simply cheaper and more widely available in China compared to the U.S. because our power grid is aging. And there's been actually a lot of media coverage of that, you know, in the recent, uh, last couple of weeks.
Scott Caravello: And, so, Katherine, since you like how I summed it up so much last time… I’m gonna go for it again.
Katherine Forrest: Yeah, sum it up again! Take us home, Scott!
Scott Caravello: So, China’s placing bets on different approaches to AGI, including embodiment, and this large scale deployment of AI throughout society, and, just as in the United States, the infrastructure is an essential part of the story, too. And, China’s buildout of its power grid and data centers are a key piece of that puzzle.
Katherine Forrest: The data centers and power grids, energy, water, the resources, the rare earth minerals, these are all essential components that sometimes we lose sight of, but they are the core of the whole thing. So, that's all we've got time for today. I'm Katherine Forrest.
Scott Caravello: And I’m Scott Caravello—don’t forget to like and subscribe.
