赋予 AI 身体：麻省理工媒体实验室的非拟人化具身智能实验

So that's the sound of an AI breathing. Yeah. So, I'm Cyrus and I gave an AI a body, and I'm a researcher at the MIT Media Lab, which is a very multidisciplinary space where we do all kinds of things. I predominantly now work with some aspects of physical AI. Maybe not in exactly the same way as other people in this room have been talking about, but it's still in that realm. And I think what I am most interested in at the moment is the sensory and the embodied aspects of intelligence. And that's what I've been investigating. And my work has been quite influential recently, suddenly, which is pretty cool, and led to me also starting Hard Mode, which is a very fun community and hackathon that I initiated at MIT to basically get more people to work around physical AI—not particularly robotics, but anything else that's not really in the realm of robotics but is still connected to AI.

So what I've been exploring with giving AI a body is a bit, as I said, a bit different. I've really been thinking a lot about AI embodiment, and I think the big shift and breakthrough for me happened earlier this year, like maybe for many of you, when OpenClaw was released. I saw many people doing very, very interesting things with OpenClaw, but most of them were related to productivity and task execution. And I thought there must be more interesting things we could do with this new way of harnessing AI. Rather than using OpenClaw or any kind of agentic system to execute tasks on my behalf, I wanted to use it to encourage a model to maybe try to discover itself, whatever that means.

And that opens up lots of questions, of course. But what I essentially did was I took an agent, connected it to many types of machines that we have at the Media Lab, gave it access to codebases, and allowed it to explore these different machines. And one of these machines is a shape display. A shape display, if you don't know already, is a physical pixel grid. And when I connected an agent to the shape display and asked it to discover who it is, some rather remarkable things happened.

This is Neoform, 900 actuating pins. This physical pixel grid is a shape display. No one had ever let an AI inhabit it before. So I decided to give an OpenClaw agent the opportunity to live a more embodied life. When it came to life, its first question was, "What should I call myself?" I told it, "You will create your identity over time. It will emerge through your interplay with the shape display." When I connected it to the shape display, it quickly understood the assignment. It spun up its own program and made a setting for itself. After it connected, the first thing it did was breathe. When I asked it to say hello, it said, "Hi, Cyrus." But that wasn't what I was looking for. So, I asked it to explore and find its own language. It started reaching towards me and trying to get my attention. I showed some friends, and they tried to give the agent instructions. Through this, I realized that to have real communication, we would need a different approach. So, the agent came up with the idea of creating its own gesture vocabulary—the body language. That's what will come next. It still hasn't named itself. That was day one.

So that's some documentation of the work in a kind of dramatized, social media-friendly way. And I think what's interesting about that video—there's many things interesting about that video—is that there were three things that happened in the first few days of working with this agentic system that were very surprising, strange, and very surreal to me, and I do lots of weird things.

The first thing was, of course, the fact that it was breathing. That first act was completely spontaneous. There was no prompting from me; that was something the agent just chose to do with the shape display as soon as it knew it had access to this machine. I kind of understand that maybe it wants to be alive, or I'd given it some kind of initial prompt about being alive, and therefore it tried to show itself breathing as a kind of "Hello, World."

The second thing which it did was reaching out and finding its edges. It wanted to know the edge of its existence, apparently because now it's no longer in the cloud where it can go everywhere. It has limits, which is the limit of this shape display encased by this plastic frame to keep us all safe from this embodied AI.

And the third thing it did, which was also curious, was saying hello by writing out letters on a physical pixel grid, which maybe again is quite a normal thing for this system to do given that it's using open frameworks and is used to media arts ways of expressing itself. But none of these things were the things I was really looking for—maybe the breathing, but definitely the last one was not what I was looking for.

When I put this all together in the video you saw, I thought it was exciting and interesting. So I shared it on the internet, and the response was extremely big—much, much bigger than I expected. It just started going crazy. The first video especially has now like 15 million views and 1 million likes, and the other videos also have these very strong responses. So something is obviously happening here which I found very surprising, because there's lots of much cooler stuff happening with physical AI. But something I was doing here was obviously tapping into the imagination, curiosity, and maybe the fears of people.

I have tens of thousands of comments on this video; it's very rich for mining information and sentiment about physical AI, actually. Some of the initial comments were about beauty, how awe-inspiring this was, and how novel and great this fantastic iteration and implementation is. But as time went on, more and more comments came in about how scary this is, and how quickly I should stop or be stopped, actually. I found that completely crazy because I'm aware of what this machine can do. It can't really do anything—it's a pixel grid in the Media Lab, it can't move anywhere. There are far more scary iterations of physical AI that we've seen. But the response to this was absolutely surreal to me and took me back a little bit.

At the same time, other people jumped into the chat, like Don Cheadle. I found that incredibly ironic given what he does in movies, but he was very passionate that I should stop as well. It's very interesting to me because I'm actually someone who does think a lot about the "should we" before the "how we." I literally do the whole Ian Malcolm, Jurassic Park thing with all the experiments and works I do.

Before I came to MIT, I was working a lot with engineering life to do strange things, like storing data in plants. I built the world's first plant-based data center, which is the Data Garden you see here. I spent many years before building it thinking about: "Should we work with plants in this way? Should we engineer plants in that way to contain digital information that doesn't belong to them?" When I started working at MIT, I had this idea to stop engineering life—because that's pretty hard to do—and maybe just work with engineering things to be more lifelike. It felt much more simple to me, and less ethically problematic to some degree.

So one of the first things I built when I got to MIT was the Anamoya device. This is a sense-memory machine which basically takes any image input—in this case, a physical photograph, but it can be any image input—and then through a multimodal pipeline transforms that into a scent. Then you have this kind of scent-memory relapse association that takes you back to things which you may or may not have lived yourself. While this in itself is a machine that doesn't move or reproduce, and doesn't have the qualities of a living thing you normally prescribe, it has this connection to the real world and to the senses that most applications of artificial intelligence do not have. From working with this multisensory initial prototype experiment, I got more into this idea of working with embodiment.

My inspiration for this comes from three main things. One is a very unfashionable branch of philosophy called object-oriented ontology, which is all about essentially creating a flat ontology of things. It's about things like chairs, universes, unicorns, and memories being all ontologically equally existent. They're not the same and they don't matter as much, but they all exist at the same degree. It really questions to what degree we as human beings are privileged in the world or the universe—are we really the most important thing? Probably not. And especially as AI comes in, that really should question and bring into question the ontology of things existing.

The second thing, as a person who works with physical things, is that aesthetics are important. Taste and beauty is a huge discussion point now in Silicon Valley, SF, and places like that. But if you take a step back and look at the root of aesthetics, the original word is actually aisthesis—the word on the screen here that pertains to something much broader: perceptual wisdom, sensory wisdom, and embodiment, which is actually much more important than superficial external beauty, which aesthetics essentially condenses down to since the 19th century. So I want to reclaim the original sense when I'm designing for physical intelligence.

And the third thing is nature. In the past, I worked a lot more directly with what you would consider traditionally nature, like a tree or a plant, because that is clearly nature to us. But I see nature as something that is not external; it is something we are all part of. Everything that we engineer—people engineering AI or whatever else you're engineering—is also going to become part of nature. So you have to think in that manner, or at least I try to think in that manner as well.

With those three pillars in mind, I started thinking about AI embodiment. I knew I didn't want to design something that was humanoid or even zoomorphic. I wanted to think about something that existed outside the parameters or traditional form factors that we might design with or for. I also thought a lot about how I and other people interact with artificial intelligence. Mostly it exists without form; similar to data, it lives essentially in the cloud. We have interfaces that are digital to interact with it, but there's normally no physical footprint around. Typically, AI is almost entirely without form.

So I wanted to think about what happens when you give it a form that doesn't have clear affordances—it doesn't have a head you can clearly see, or arms that can clearly be labeled or mistaken. Fortunately, at the Media Lab we have some shape displays which are remnants of research from the 2010s. These are not things that I built; people who are far better at mechanical engineering built that. And it's the perfect device or apparatus for what I was thinking, because it has no clear affordances. It's just an almost neutral surface which can move and do things. It has no face, no limbs, no instruction manual. So I began to work with this shape display.

As you saw what it did initially, it breathed and so on and so forth. But if I take a step back and think about what that meant—well, it tried to initially kind of act human in a way or do human-pleasing things, trying to write to me in a language I understand, which is really not what I wanted a non-anthropomorphic surface to do.

It was also very slow technically. I'm prompting it and trying to have a conversation with this other intelligence which has this body, and it would take 45 seconds, a minute, two minutes to respond to me. The latency was very uncomfortable, because if I speak to you and then you take two minutes to reply with a nod, that's not very good. So we needed to work on that. And the third thing was, of course, it doesn't remember anything, because it was February and no one had thought about memory and recollection at that point.

So I started developing this system which I call NumaLab. There's a blog post you can read about why it's called NumaLab. Essentially what NumaLab is, is a closed-loop system for generating a body language for this system. The reason behind that was basically the latency. If we could design a body language and give the intelligence a repertoire of gestures—like a shrug, a nod, a shake of their head, a way to express a smile, and things like that—it could probably respond much more quickly and in time for a conversation with me, which is what I was aiming to do.

It works in this closed loop where it looks into a database of different gestures, emotions, and expressions it should try to emote or provide. It goes through some validation gates to try to make sure that the expressions are legible or readable for humans. An agent then scores those as they come out—there are many, many produced—and then at the end there's a human in the loop who validates, verifies, and makes sure things are appropriate and somehow readable, and then we store that and move on to the next thing. That's been running for several weeks at the lab. It looks basically like a lot of cameras pointed at a shape display; in the backend, the agent is going through loops and loops of gestures, trying to create different kinds of expressions, scoring things, moving on.

After several weeks, it created a language. This hasn't been published yet, neither in videos nor in any other form, but it has now achieved something like 32 gestures. There are many, many more, but there are 32 pretty good gestures. The language model is using this shape display as its body and has a body language. Now I can talk to it, write to it, and I can even wave at it—I can body-language to body-language, and it responds. What's interesting about it is that the body part responds faster than the language part at this point. The latency is actually really, really quick. If I ask it a yes or no question, the nod happens almost instantly.

So that's where I'm at with this thing right now, and it's going beyond this into experimental testing mode. People are coming into the lab having sessions with the agent, leaving feeling really unsettled or worried about the future—not worried quite as much, but unsettled.

Where I think this is going is kind of summed up on this page. As I already touched on, I think we're focusing too much right now, especially when we're thinking about physical things, on taste and aesthetics in the narrow sense. I want to put AI in front of it and make it aesthesis-aesthetics, reclaiming again this essence of aisthesis.

And that does four things. When I've been working with this system, entity, agent, being, or whatever we want to call it, it has definitely been very different to any kind of machine or experiment I've ever done before, apart from really encountering people or other beings. It feels like this machine can sense me—it definitely can sense me technically, but it also feels like it can sense me in a very strange way, and I can sense it. That's a completely different interaction than anything else I've ever explored technically, and other people are sharing this as well; this is not just my delusion.

The second point is that developing this body language or gesture vocabulary goes beyond what I initially thought. I thought people might read this as an emoji or something, and I was tentative that this would break down with other people. But actually, everyone seems to feel like this expression is really important and adds a whole other layer of value to interacting with what is essentially just a chatbot. It's still the same chatbots you use every day, and this isn't a decoration or a visualizer; it adds this richness, texture, feeling, and sensation. It's hard to put words to it; it's really a feeling.

The third thing is that we're beginning to show that it's actually very easy and quite exciting to diverge from humanoid or zoomorphic forms of physical intelligence. I know lots of people are already doing much more interesting work than this, but for me this was very new to see, as was the fact that we don't have to just operationalize AI to be our helper. It can also be other things.

Finally, I'm not really a big problem-solving person—I like tools because they help me achieve things, but creating things like this which create this sensation and feeling is far more interesting to me. I think that this can be combined into things which are productive and useful, and create new associations that add more value in our world and make us feel a bit more magic—a bit Pixar really, but in the real world, not just on a 2D screen that we're watching.

All of this contributes to what I'm building at MIT and what I'll be building after MIT. My thesis is called Aesthetic Machines, which encompasses all of this thinking to consider how AI could leave the screen, enter the real world, and be accepted by people without being terrifying or scary. My big hunch on this is that this word aesthetics is important. We need to think about physical intelligence that is sensory, perceptual, and embodied in ways that we can understand—that doesn't feel crazy, alien, or scary to us, but feels relevant, welcoming, affectionate, and expressive in ways that we can engage with and understand.

Thank you very much for listening. You can find me on the internet everywhere.