[OPEN]


VO: As far as we’ve come with AI and robotics…. There’s still a huge, huge gap when it comes to combining the two.  


AI excels in the digital space… and in the physical world, robots are often pre-programmed. 


That’s where physical AI comes in. It’s critical for things that can’t tolerate the kinds of mistakes that are common in today’s statistics based AI… like self driving cars… or managing the power grid. 


At the center of this work is the 2025 I-triple-E Edison Medal Recipient,
Daniela Rus. She’s the Director of MIT’s largest research lab - C-SAIL, which is the Computer Science & Artificial Intelligence Lab.




AXX Daniela Rus (00:00): If your computer vision system makes mistakes 10% of the time, and you put that computer vision on a car and you ask it to help guide the car, you cannot afford to make mistakes 10% of the time. You see, it's not just robots that are in the physical world and that need solutions, that understand the physics of the world. So many other applications that are critical to our economy and to our everyday life operate in the physical world. And so to make the most of ai, we really need the kind of solutions that do not make mistakes, and also the kind of solutions that can run in situ on the body of the machine rather than in the cloud.


VO: I’m Jennifer Strong, and this is SHIFT. 


[THEME]


VO: This episode… It's the latest installment of our oral history project.
Daniela Rus (00:49): My name is Daniela Rus. I am the director of the Computer Science and Artificial Intelligence Laboratory at MITA, professor in the Department of Electrical Engineering and computer science, a robotics and AI researcher, and the co-founder of four companies, including Liquid ai. Well, as a child, I, I grew up in Romania, and Romania was under a very harsh regime. It was the communist regime of Nikolai Charles School. There was no tv, there was hardly anything to buy yet we had a happy childhood and a childhood full of dreams and fantasies. And in my case, I spent a lot of time reading science fiction and kind of fantasizing about new ideas, about invented better worlds. And so I was a huge and very passionate reader of the Verne novels, the kinds of things that used to belong to science fiction back then, but are becoming reality today. And in fact, I find this really extraordinary about the moment in time we are experiencing today. I find it really extraordinary that many things that used to belong to our reading fantasies and to the movies are now becoming reality. And they're being enabled by machines, by robots, by computers, by ai, by advances in mat, in materials, and by advances in algorithms and problem solving with computing.
Daniela Rus (02:34): I did my schooling, most of my schooling in Romania. I was good in math and in physics, but the education in Romania was really very broad. So we studied history and geography and art and languages, just about everything. And I really enjoyed school. I really enjoyed learning. The other thing that was really interesting about the education I had was that one month every year I had to work in a factory. All school kids did that. And, uh, I worked in a factory that produced spare parts for locomotives. And so as part of that, I learned how to use a machine shop. I learned how to use a lathe. I made a screw from scratch. This was, this was an experience that actually didn't seem so useful back then. But now, as I reflect upon this trajectory, I feel like this experience really gave me the, the tools of, of building in the physical world. And as the math that we studied, uh, in school got more and more abstract, I knew that I wanted to do something massive, yes, but also something with a physical component. So between visual fantasies, math and physical components, I, I somehow converged towards an undergraduate degree in computer science and mathematics along with an astronomy minor.
Daniela Rus (04:01): And then I went to work on my PhD. And at the time, my thesis advisor, John Hopcroft, told me, well, you see Daniella in computer science, most problems are solved. What he meant by that was that a lot of the problems that the founders of the field developed and proposed as important problems to solve, had solutions. And then John said, now it's time for computation that interacts with the physical world. So the early part of computing was all about computation inside the computing box. But as we expanded how we view computing, how we imagine computing interaction with people, interaction with the physical world became all very important. And these are very different from computation that happens purely in the digital space. And in fact, it's a very interesting, uh, moment in time today because, uh, I see a kind of a parallel as we think about where AI is today. We see AI having tremendous impact in the digital space, but not so much in the physical space. And so I've started talking about the importance of developing physical ai, AI that can understand online information. And with that it can make the, the physical machines more intelligent, but also AI that doesn't just operate based on statistics, but also operates by understanding the laws of physics, understanding the rules of the world, understanding common sense, which today's traditional AI solutions have trouble with.
Daniela Rus (05:47): Computing is a very young field. Maybe 60 some years ago, our founders at MIT had the crazy idea that two people might use the same computer at the same time. That was a dream back then. And so can you imagine in, in, uh, 60 short years, we moved from two people dreaming about using the same machine to a world where everyone computes. It's extraordinary. Also, consider that 20, 30 years ago, computing was something that belonged to the experts view. Uh, computers were large and complicated, and you really needed expertise in order to know what to do with them. And look at the world today, we have access to so much power, to so much capability that didn't exist just a couple of decades ago.
Daniela Rus (06:43): The field of AI was defined and formulated in, uh, 1956 when the founding fathers of AI led by John McCarthy and Marvin Minsky went to New Hampshire, and they spent a summer at Dartmouth, maybe one month, uh, during a summer at Dartmouth, they talked about where computing was going and what might be the next important academic field of study. And when they came out of the woods, they told the world that they had coined a new term, artificial intelligence, to denote the science and engineering of making machines with human-like characteristics in how they play games, how they see the world, how they move in the world, even how they learn. And we have been on the journey of understanding and developing these kinds of machines and in parallel understanding nature ever since. And this is also a very young field, if you think about it, 1956 to 2025 is not that long. So we've come a very long way in computing and in ai, and we have a very long way to go. The objective is to develop the science and engineering of intelligence. That means understanding natural intelligence and then understanding how to make machines that exhibit some of the properties of naturally intelligent systems. And we are making good progress, but we have a very long way to go still.
Daniela Rus (08:29): So right now in the field of computer science, we have two largely separate fields. We have artificial intelligence and robotics and artificial intelligence is amazing us with its ability to make predictions and to learn. But it is confined in the digital space. It is confined inside computers. Robots on the other hand, have a physical presence, presence, and the ability to execute tasks. But they are mostly not intelligent. They're mostly pre-programmed. And so physical intelligence aims to bridge the separation. Physical intelligence aims to bring the solutions from the digital space to the physical space. But this requires innovation because today's AI solutions are not ready for the physical space. They do not understand physics. They do not understand common sense. They do not understand the, the laws of nature. And so what we have is extraordinary capability from statistics based solutions. But we need more for the physical space.
Daniela Rus (09:39): We need more because today's AI makes mistakes, but applications in the physical world cannot tolerate mistakes. If your computer vision system makes mistakes 10% of the time, and you put that computer vision on a car and you ask it to help guide the car, you cannot afford to make mistakes 10% of the time, for example. But you see, it's not just robots that are in the physical world and that need solutions that understand the physics of the world. So many other applications that are critical to our economy and to our everyday life operate in the physical world. And so to make the most of ai, we really need the kind of solutions that do not make mistakes, and also the kind of solutions that can run in situ on the body of the machine rather than in the cloud. And these domains might be managing transportation, managing power grids, managing supply chains. There are so many different ways in which practically every field, every industry sector can be enhanced and improved by AI provided that we can get the kind of AI tools that are are useful.
Daniela Rus (11:07): I will say that we have a lot of capability from today's AI solutions, but there are new ideas we can bring to bear to create new kinds of ai. And so, for instance, in my research group, we have been studying natural intelligence, and we have been building AI solutions that use a different kind of mathematics, in other words, different kind of technology than the transformer based solutions, which are now the ubiquitous AI solutions in most industry applications. And with this approach that is inspired by nature, we have achieved extraordinary results. We have achieved AI solutions that can be analyzed using dynamical systems tools. In other words, using the kind of mathematical methodologies that are now used to guarantee the safety of, uh, just about any other machine on the planet. The solutions are also much smaller. They can run in situ. That means they're private. The solutions are also less energy hungry. That means that they will not consume, uh, so much energy and they could run on the body of the machine, on your phone, on your computer, on your smart glasses, on your fridge, and on who knows what other devices we haven't invented yet. And so I'm really excited about small ai.
Daniela Rus (12:46): The solution that we came up with, we termed liquid neural networks. And what's different in the liquid neural networks is essentially two things. The computation inside the liquid neural network is inspired by the mathematics of the computation inside the neurons of a small worm called sea elegance. And, uh, sea elegance has inspired a lot of work in neuroscience today, and the wiring between the neurons is a little bit different. And with these changes, we get phenomenal results. First of all, we get a provably causal, uh, solution. That means that these, um, liquid neural networks learn to associate cause and effect. In other words, they learn the task rather than the context of the task. And that allows us to apply them in different domains because you see, for most AI solutions that are available today, you train your solution and that becomes frozen. This means if you deploy it in the wild, it does not continue to learn.
Daniela Rus (13:49): If you want new capabilities, you have to wait for the next release because of what's happening inside the liquid neuron. Um, liquid neural networks continue to learn and adapt after training based on the inputs that they see. And this kind of adaptation and ability to learn a skill in one environment and apply it in a different environment is a very important part of intelligence. And so we have a lot of, um, opportunity and excitement, uh, around making a different kind of AI that can operate on the edge, and that can enable not just energy efficient solutions, but also solutions that are private that do not require the, um, the cloud, um, calls. And so the way, the way I see it is, of course, there is a lot of value in, uh, in the, the large models that, uh, we ha are building and we're deploying.
Daniela Rus (14:48): But there is also a lot of opportunity in making smaller models that can run with causality and with generalizability and adaptation within context, within given context on device. And this is really an important point, and it's at an important moment in time. First of all, AI is transforming everything. And then even within the field of ai, I feel like we are at an inflection point because for years, progress has been measured by larger and larger models and ever greater demands for data centers. And this is pushing intelligence toward industrial scale, uh, infrastructure. But I believe that the future of AI will not rest in size alone. The future of AI will be defined by capability, by efficiency, by adaptability, generalizability, ubiquity, in other words, by intelligence that runs everywhere, that runs instantly and that runs privately. And so this is the kind of vision that liquid neural networks and the company we co-founded, uh, liquid AI aims to bring to the world.
Daniela Rus (16:03): And, um, and this is called in technical to uh, terms, this is called private edge ai. It means AI that can run on edge devices. So edge devices are small devices that do not necessarily talk to the cloud or, um, you know, small devices like, uh, appliances or phones or, um, or um, uh, glasses. And, and so I believe that, um, that making progress on private a edge AI will open, uh, tremendous opportunities in the future, uh, because this will move computation from distance servers to our devices in our hands, in our homes, uh, in our workspaces, where the solutions could empower everyone, people, organizations, companies, um, and, and they will do so while reducing costs and while reducing the energy use. And that's good for the bottom line. It's also good for the planet. And the sort of shift could have a tremendous economic implication because if you think about a model where you decouple intelligence from, uh, centralized data centers, uh, you can imagine ways in which we could redistribute, uh, the value creation from a handful of infrastructure providers to any business, to many, many businesses, organizations, to communities who can put AI in their own systems. It's almost like what happened with a personal computer. The rise of the personal computer shifted value away from mainframe operators to a broad ecosystem of innovators.And so I think the rise of private edge AI will also democratize access to artificial intelligence and will spread productivity gains across the global economy. 
[MIDROLL]


Daniela Rus (17:49): So Jennifer, during the pandemic, I had some extra time because I wasn't commuting back and forth to work. And also I had an opportunity to reflect about where we were as a, as a species at the beginning of the pandemic. It was really scary. And so I, I did a lot of reflection and I decided to work on two books to bring the robotics and AI ideas to the public. And when I got started, my, my dream was to really explain to the public how these technologies work, and hopefully to get everyone to think about a better world and a better life with support from machines, with physical support from robots and cognitive support from ai. And so this, this project turned into two books.
Daniela Rus (19:46): The first book is called The Heart and the Chip, our Bright Future with Robots. And this book talks about all the different ways in which robots can support us by amplifying our strength, by refining our precision, and by granting us the ability to interact with the world in ways we've only dreamed of as a way of overcoming our human limitations. And, uh, the book talks about what is possible today and in the future, also how the, the systems work and what kinds of things we need to be worried about. The second book is about ai, and the second book is called The Minds Mirror Risk and Reward in the Age of ai. And this book also talks about cognitive superpowers like speed, knowledge, inside creativity, foresight, mastery, and empathy, and explain what is possible and what will be possible. Explain how AI works, gives
Daniela Rus (20:47): An AI playbook for companies and organizations, is a set of questions and the kind of a framework to think about whether an AI solution is suitable for the company. And if so, how do you create value and what do you need to put in place before you push the go button? And then the third part of the book talks about stewardship, also, what could go wrong and what we should do about it. And so I'm very excited about bringing this knowledge to the public, and I hope it empowers people. If we want AI to reflect our highest values, then we can't leave its future to any one group. And so stewardship of ai, uh, necessarily requires participation from broad societal and organizational groups, participation from policy makers, participation from technologists, participation from business leaders, even participation from the public. And this future will be a collective journey, will involve the public developing AI literacy and AI skills for their respective trade.
Daniela Rus (22:11): And I, I think that's actually very important because knowledge empowers people. So some level of AI literacy, uh, is important. Not everyone needs to be an AI geek, uh, who can recite the mathematics of the artificial neuron. But understanding how AI operates when you can and cannot use it, uh, how you, how you should think about the probabilistic and statistical answers that come out of an AI engine, uh, is really important. And the public will need to know different things about AI depending on what they wanna use AI force. So if they lead with ai, if they use ai, if they develop ai, if they deploy ai, one needs to know different things depending on how AI can support, uh, your work or your activities as an individual. Business leaders also need to stay informed. They need to incentivize safety. They need to think about implementing and deploying AI solutions that benefit the stakeholders and society.
Daniela Rus (23:18): Technologists, uh, need to think about how do we test and evaluate the AI systems we deploy extensively? How do we develop technical guardrails? How do we build AI systems that embody the societal values? And policy makers need to stay informed about AI and they need to incentivize safety and innovation and consider how AI fits with existing laws, because we already have a lot of existing illegal infrastructure that could help govern AI and, uh, if necessary, they should create new laws with agility. All of this is very important as we think about moving AI forward. The final thing I will say, Jennifer, is that if we step back and reflect we are the only species living on this beautiful planet who are so advanced, so aware, so capable of building these extraordinary tools. But with this empowerment also comes the responsibility to ensure that the tools are used for the greater good, for the planet and for everybody and every plant and animal, we share this, this beautiful planet with.








[CREDITS]


VO: This episode was produced by me and Emma Silicons and mixed by Garrett Lang with original music from him. And Jacob Gorski, thanks for listening. I'm Jennifer Strong.