ACM ByteCast - Peter Stone - Episode 84

Episode Date: April 16, 2026

In this episode of ACM ByteCast, Rashmi Mohan hosts 2024 ACM/AAAI Allen Newell Award recipient Peter Stone, Professor at the University of Texas at Austin and Chief Scientist at Sony AI. He received the award for significant contributions to the theory and practice of AI, especially in reinforcement learning (RL), multiagent systems, transfer learning, and intelligent robotics. As a leading figure in AI research, Stone has fundamentally advanced how autonomous agents learn, plan, and collaborate. His groundbreaking work on RL algorithms has enabled robots to acquire skills through experience. He is an ACM, AAAI, AAAS, and IEEE Fellow, an Alfred P. Sloan Research Fellow, and a Fulbright Scholar. At UT Austin, he is the founder and director of the Learning Agents Research Group (LARG) within the Artificial Intelligence Laboratory, as well as Founding Director of Texas Robotics. In the past, he also worked at AT&T Labs - Research and co-founded Cogitai, Inc. (acquired by Sony). Peter explores the intersection of professional research and personal passion, detailing how his lifelong love for soccer fueled his involvement in RoboCup, where he aims to develop humanoid robots capable of competing at a World Cup level by 2050. The conversation highlights his leadership as the Chief Scientist of Sony AI, focusing on landmark projects like GT Sophy, an AI that mastered the complexities of Gran Turismo, and the development of FHIBE, an ethically sourced dataset designed to mitigate bias in machine learning. Throughout the interview, Stone emphasizes the importance of ad hoc teamwork—the ability of autonomous agents to collaborate on the fly with unfamiliar partners. He also shares his passion for undergraduate research and advocacy for AI education at all levels.

This is ACM Bythcast, a podcast series from the Association for Computing Machinery, the world's largest educational and scientific computing society. We talk to researchers, practitioners, and innovators who are at the intersection of computing research and practice. They share their experiences, the lessons they've learned, and their own visions for the future of computing. I am your host, Rashmi Mohan. We've all heard the warnings about AI taking up. over the world. But today's guest is more interested in whether AI can win a championship. He is a mastermind behind the digital athlete, creating machines that can navigate a 180 mile

per hour pair pin turn with more finesse than a pro racer and lead a squad of robotic soccer players to a world title. But his work isn't just about games. He has been the architect behind many complete intelligent agents for many decades. Dr. Peter Stone is the Trushard Foundation Chair in Computer Science at the University of Texas at Austin, the founder and director of the Learning Agents Research Group within the AI Lab, director of Texas Robotics, and the chief scientist of Sony AI. From leading the Global Robo Cup Federation to developing G.T. Sophie, the AI that outraced world champion drivers in Grand Tourismo, his work focuses on the ultimate challenge,

creating autonomous agents that can navigate the messy, unpredictable reality of the physical world. Peter is also an ACM fellow and has most recently won the ACM Triple AI Alan Newell Award, amongst many other honors. We are so excited to speak with you, Peter. Welcome to ACM Bikecast. I'd love to lead with a simple question that I ask all my guests,

which is if you could please introduce yourself Tell us about the work that you do and give us some insight into what drew you into this field. Yeah. Thank you for inviting me. It's a pleasure to be here. My name is Peter Stone. I'm a professor of computer science at the University of Texas at Austin and also chief scientist at Sony AI. And my research has always been in artificial intelligence. So from far before, it was as cool as it is now. In fact, when I got into the field in the early 90s, many people thought that its time had come and gone. But what drew me, to the study of artificial intelligence was really the mystery of how the brain works and what is the nature of intelligence. I'm not the first to say in the AAAI presidential speech where Kambapati laid out what he thought was the three big challenges of our time, scientific challenges. One is how did the universe begin? How did life get started on Earth? And then the third is,

what's the nature of intelligence. And that's what piqued my interests, starting from when I was in high school. And so when I went to University of Chicago as an undergrad, I explored different ways to try to get at that question. And I went to some neuroscience classes, some psychology classes. But neither of those really hit the mark for me. In the neuroscience classes, we were learning how, you know, about axons and dendrites and neurotransmitters and sort of the molecular questions of how the brain works at that time. And I sort of thought, but yeah, but that's not telling me anything about what happens in my brain when somebody tells me their telephone number and I remember it or something like that, which was a thing back in the day.

Now we don't have to remember anybody's telephone number. We just put him into our phone. And so I went to some psychology classes. And there it was very, very interesting to learn about what people knew about human behavior, what caused people to, you know, under what circumstances people would behave in different ways. I learned all about the work of Kahneman and Tversky. And it was fascinating. But again, that was too high level as opposed to the neuroscience, which seemed too level for me.

It still wasn't telling me what's going on in the brain, when I remember something. And so then I started going to the computer science classes and was introduced to artificial intelligence. At the time, there was actually no computer science major at University of Chicago. I was a math major, but took all the computer science courses I could, including some graduate courses in artificial intelligence.

And that's where it sort of, to me, felt like there was a way to try to get it, what is the nature of intelligence by trying to recreate intelligence. They have theories about how people may be doing intelligent things and try to recreate those in computers. and that fascinated me. In some sense, the rest is history, but it's been a long path since then.

I did go on to graduate school at one of the top places for artificial intelligence time still, Carnegie Mellon University. I was fortunate to have a fantastic PhD advisor who gave me the freedom to explore the ideas that most interested me. That's Professor Manuel Velosso.

And then when I graduated, I went to AT&T Labs research, a fantastic team, the best people in the world working on reinforcement learning and other forms of artificial intelligence, at the time, was there for about three years, and then in 2002 came to University of Texas at Austin as a new assistant professor. Wonderful. Thank you for sharing that. And I mean, I think you

already called this out for, you got into this field so much earlier than the excitement of today. I know that your work is significant. You've made a lot of investments as well as progress in multi-agent systems and interactions. That has been sort of the cornerstone of your research. Why did this area in particular fascinate you, Peter? And like sort of what has been the progression of thoughts and ideas that you have seen in this field? Yeah, that's a great question. Thank you for that. Yeah, I usually say my research is sort of at the confluence of three different sub-areas of artificial intelligence. Multi-agent systems is one of them, for sure, and then also machine learning, specifically reinforcement learning, and robotics and putting those three together.

But multi-agent system is the way I got into it. It was actually, just after my first year as a PhD student, during the first year, I was immersing myself in the area of artificial intelligence planning, There was a system at Carnegie Mellon called Prodigy, that my advisor was key participant in. So I was working on sort of ways to do better and more robust planning. And then I went to the main AI conference in the summer of 1994. I was in Seattle. So this was AAAI, 1994. And there there was a demonstration by Alan McQuartz Lab from University of British Columbia and his student at the time, Michael Sohota.

And they were showing a one-on-one soccer match. And it was two robots on a ping pong table playing soccer, trying to knock the ball into goals on either side. They had, Alan at the time, I think in 1993, had published a paper, putting forth robot soccer is a great challenge task for artificial intelligence. And so they were showing this demonstration in connection with that. And I looked at that. I was on my soccer team as an undergrad. I've played my whole life. I was on the varsity team at the University of Chicago.

and I got there and I was immediately captivated. I was like, oh, robots playing soccer, that's fantastic. It could merge my personal interest, outside interest, with my professional interest in artificial intelligence, but it felt like something was missing. One-on-one to me did not make a soccer match, did not make a soccer game. You need to have multiple teammates on each side.

And so I sort of made it my goal to expand what they were doing into at least three-on-three, eventually up to 11-on-11, which is what people do. And that got me deeply into the sort of burgeoning field at the time of multi-agent systems where people were thinking about, you know, if you created a single autonomous agent, that wouldn't be, you know, complete enough. They would always need to interact either with people or with each other. There were people thinking about how could autonomous agents communicate. They were thinking, people thinking about what would happen if they were competing. And soccer sort of had this dual aspect of multi-agent systems, both teammates that you collaborate with and adversaries. that you compete against. So it's both collaborative and adversarial at the same time. And so as I dove into that and learned about the literature in multi-agent systems in the mid-1990s, I learned that people were making a strong case that really resonated with me, which is that a lot of intelligence, a lot of human intelligence, is focused on our ability to communicate

and interact with other agents. We can look at other people and predict what they're going to do. we make assessments about whether they're good to cooperate with or not based on their body language or what their facial expressions are. People were starting to make the case that multi-agent systems and interaction is one of the main seats of intelligence and might be one of the keys to unlocking artificial intelligence. So I became sort of involved in all aspects of artificial intelligence, not just the ones that were motivated by robot soccer. I've been involved in other types of autonomous agents like autonomous bidding agents,

autonomous traffic agents, you know, autonomous cars are also in a multi-agent system. And they're neither collaborative nor adversarial, by the way. They're just in a system with other self-interested agents. And so I've sort of, through my career, looked at many different aspects of multi-agent systems, and it continues to fascinate me. I could talk later about ad hoc teamwork, which is something that's area of multi-agent systems that I helped introduce and which has fascinated me over the last 15 years or so. Oh, absolutely.

I mean, this is fascinating. I was thinking through, I was trying to do some sort of research on your work, and I saw your deep involvement in Robo Cup, and I was wondering how that engagement began. You know, you're one of the leading figures that participates and develops strategies for this. I'm wondering, when you have so many agents that are all sort of interacting in this system, you know, as we do as humans, right? We're optimizing, in the case of a soccer match, trying to optimize for the same goal, what are the sort of key considerations that you need to keep in mind as you are trying to design this system?

It's a very complex question. There are many things. There's no, and one of the fascinating things about Robococor and Robot Soccer is that there's challenges at so many different levels. So first of all, I should say, yeah, I was very fortunate. I told you about how I became interested in multi-agent systems based on this demonstration I saw at AAA 924 from Alan McQuart. You know, I became captivated by it. My advisor gave me the freedom to explore in that direction. And it just so happened.

You know, I was in the right place and the right time. There were people from around the world that were thinking about trying to turn robot. soccer into really a global challenge problem and connected with some colleagues in Japan, Hiroaki Kitano, Izki Notam, Minora Asada, and others who were thinking of trying to create a robot soccer community, the competition, which they did. And I was one of the, you know, sort of, I was, I think the first participant in the United States, along with my advisor, along with Manuela. And it sort of turned into something that is now a real global community. There's thousands of people around the world who have become captivated by the challenges that we look at in Robocop.

One of the sort of main challenges, the founding challenge that Hiroaki Kitano put forth to the community was to create a team of humanoid robots that can beat the World Cup champion soccer team on a real soccer field by the year 2050. We're more than halfway there from when he first stated that, I think, in the late 1990s. Robocup has expanded since then to include things like disaster rescue and general purpose service robots at homes, there's a RoboCup at Home, RoboCup at Rescue. But to answer your question, what are the key challenges in robot soccer? There's, you know, they sort of span the gamut from starting at the low-level hardware issues, just how do you get a robot to move quickly and

agilely to control a soccer ball? There have been, you know, a lot of Robocuff historically has been with wheeled robots, but now there's a lot more focus on humanoid robots. But even, you know, with wheeled robots, how do they intercept the ball and move it in the direction that they want to? just an individual, how can it sense perceive and a sense decide and act on its own? That's on top of the hardware once you have the low-level capabilities. How do you intercept the ball? How do you kick the ball? How you walk fast if you've got legs? And then you get into the strategic and multi-agent issues. How do you coordinate with teammates? How are you flexibly play different positions?

So not all the robots are just chasing the ball like with five-year-olds playing soccer. So how do you organize the team strategically? then how do you play differently based on the opponent that you're playing against or the situation in the game or there's challenges at many different levels. Part of the reason that I've stayed involved and stayed fascinated by it throughout my career is I've had the experience that it has spun off many interesting research questions. I'd say probably I've had more than 25 students graduate with PhDs from my lab now and I think at least half, if not more, have been inspired by some challenge in robot soccer. None of them wrote a thesis that said,

goal is to solve robot soccer or that they did. It's still an unsolved problem. But they found some problem, some sub-problem that motivated them. For instance, one of my first students, Mohan Sriturran, worked on the computer vision aspects of robot soccer. And just at the time, you had to calibrate the cameras of the robots very carefully based on the lighting in the room. And then if the light turned off or a bulb burnt out, all of a sudden the robots wouldn't be able to see the ball anymore. He wanted to make robots more robust to that. So a thesis that went very deeply on the problem of robot perception, robot vision, that, you know, ended up not being at all really about robot soccer, but was motivated by a problem that he came across there. And that's just one

example. There's been, I'd say, you know, like I say, more than 10, probably 15 pieces from my lab that have been motivated by those problems. And I think there's still many open problems in robot soccer that they're going to, that are PhD thesis worthy. And so hopefully as a community, I'm certainly not the only person working on it. I have a great set of colleagues internationally that we're working together in some sense, but using a competition every year as our meeting point where we sort of benchmark what our progress is every year. And I hope that collectively we'll be able to reach that goal of 2050. I still don't know if it'll be possible. I look forward to that. But I think a lot of things that you say, you know, make a lot of sense

in the sense that the principles of interaction that you see within a robo soccer game can be applied across so many different applications, even if not all sort of working towards the same goal, but interacting with each other, multiple sort of systems or agents interacting with each other. One question I had around that, Peter, was what happens if there is an error, right? A robot makes an error. How do you recover from that? Is there a coach? Is there some learning that is happening or coaching that is happening on the fly? Yeah. So, first of all, the robots make errors all the time. But one of the challenges with robotics is it's a lot harder to figure out to trace back an error that a robot makes than it is a traditional laptop or desktop. It's hard to debunk things because it's hard to understand

what's the internal state of the robot without compromising its computations and its performance. But I think you're asking about, look, yeah, is there that next level of sort of oversight of coaching that humans have, which I didn't, I left out when I was talking about all the challenges of robot soccer, but you're absolutely right. In fact, in Robocop, there was for several years a coach competition where all the teams were given sort of fixed players and what they're job was to come up with a coach that could analyze the current team that they were playing against and impart high-level strategic advice in a formal language. The question was who could sort of guide these teams of agent to perform the best. And yes, that requires them to be adaptive,

to be able to change their behavior based on the commands that, or the advice that the coach would give. And then, you know, sort of the other dimension of your question is just can the robot soccer players learn? And there's opportunities at two different granularities. Early in my career, I worked a lot on the robots learning before the games, some sense, learning skills and learning strategies well before, when there's no restriction on it has to happen quickly during the course of a game. You can take lots and lots of data. An early influential paper I wrote was a collaboration with Rich Sutton, who recently won the Turing Award along with Andy Bartow for their pioneering work in reinforcement learning. And he and I created a multi-agent

reinforcement learning agent to play the game of robot soccer keepaways. So you can imagine like three, and this was in the RoboCup simulator, not with real robots, but you could imagine three simulated agents trying to just keep the ball away from one in a small space in a square, just figuring out when to pass and where to pass. This is a drill that human soccer players do all the time. And so we sort of took that and used reinforcement learning to have the agents learn to cooperate, to form a team, to be able to the keepers, the ones that are trying to keep the ball away from the other, to have them figure out where to move, to when to pass, where to pass.

And we started them with making those decisions randomly, but over the course of 24 hours of just, you know, working on their own and experiencing what works and what doesn't work, we were able to find that they could keep the ball away from the take of her, I think, two and a half times longer than random. They improved dramatically. This all happened, like I say, sort of at a single game time scale. This was at the scale of, you know, at least a day with a lot of computation

power. And there's been other examples with Josiah Hanna. We used a human that we used machine learning to create a humanoid robot, the now robot, to have it learn how to walk faster than any other group in the world had allowed that, had taught that robot to walk. And again, that was using machine learning, a sim to real kind of approach where we learned partly in simulation and partly in the real world, a method that we introduced called grounded simulation learning. But again, that all took place not during the course of a single game. Then there's, as I've been hinting, also the granularity of what can you do during the course of an individual game. There's a lot less opportunity for adaptation. People do that. I don't think there's been lots of examples yet of

autonomous robots being able to quickly figure out what the opponents are doing. I've had a little bit of work where in my PhD thesis, I created some extreme scenarios where the opponent, for some reason, only plays on one side of the field or the other side of the field and tested to see whether the agents, my team could on the fly learn to adapt its strategy accordingly. And also this learning on the fly very quickly is one of the main challenges of an area that I mentioned earlier called ad hoc teamwork. And ad hoc teamwork is motivated by people's ability to do things like playing pickup games of soccer or basketball and to cooperate on the fly in other ways. So for example, if I, you know, I'm a soccer player myself, if I go see other people playing soccer in a foreign

country. I've never seen the people before. I may not even speak the language. I can jump onto the field and start playing with them immediately. And I need to adapt my strategy based on quickly assessing, you know, am I the worst player on the field? Am I the best player on the field? There's nobody who knows how to play defense, so I should do that. This sort of collaboration with teammates on the fly, I introduced the term ad hoc teamwork to describe that. And that requires very quick adaptation, very quick modeling of what your teammates can do and then very quick assessment of how you can make the team shine. People can do this all the time, right?

If there's a disaster on the street, if there's an accident and a bunch of strangers see two cars crash into each other, even if they're all strangers, they'll immediately self-organize into a team and have one try to rescue the victims, one call 911, direct traffic around. People can just, you know, can do this. Traditionally, robots have not been able to do that

because they've needed to train together to become a team or they've had to have hand-coded paradigms or protocols given to them to become a team. And so I've been very fascinated by this second dimension of multi-agent learning, of doing it quickly on the fly with new TVs. Thank you for sharing that. So many interesting concepts in there. You also have a very unique take on just reinforcement learning, but with this new age of LLMs and the vast knowledge that is now available,

how do you see those two coming together? Yeah, that's a great question. And, you know, I think the crux of reinforcement learning is learning from experience and being able to change behavior based on trying in action and seeing what happens and seeing if your expectations for what would happen are met or whether they're going to change. And at least the initial LLMs and, you know, sort of the ones that have gone viral don't do that, right? They're trained offline. They first through pattern recognition, traditionally with transformers. There is a fine-tuning phase that people call reinforcement learning from human feedback, but it's not after deployment.

It's still in the lab back before it gets deployed to the real world. This concept actually of reinforcement learning from human feedback in some ways was introduced by one of my early students, Brad Knox. He created a system called Tamer that also learned from people saying, you know, yes and no, or good job, bad job to an agent, found that it was very powerful. So it's really nice to see that's playing a role in, the training of large language models. But once they're deployed into the real world, they're not learning anymore, right? They're fixed entities. Their weights are there. Of course,

people are working on changing that, but I think that's the missing piece for the sort of initial large language models is getting them to learn from experience in the way that reinforcement learning agents do so that they can improve their performance based on trying things and seeing what works it does it in the world. So I'm excited to see what the confluence of these areas will bring, the power of large language models for pattern recognition, in language and vision, and now in some sense in actions for robots,

with the ability to try things through trial and error, see what works and see what doesn't, and be able to change behavior, change weights, change the internals after deployment. ACM Bytecast is available on Apple Podcasts, Google Podcasts, Google Podcasts, Podbean, Spotify, Stitcher, and Tune In. If you're enjoying this episode, please subscribe

and leave us a review on your favorite platform. I want to talk a little bit about your role as chief scientist at Sony AI. You obviously had a very long and successful career in academia, but you've also had a lot of interactions with industry, and now you have a role very much in a corporation. What led you there? Tell us a little bit more about. about the work that's coming out of that rule?

I never really considered myself entrepreneurial at all. I saw myself as an academic through and through. I really love my role at the university and working with students and doing basic research. I think it was AAAI in Austin, around 2014 or 2015 sometime. I was approached my longtime longtime colleague of mine, somebody who I met in graduate school, Mark Ring,

who pioneered the concept of continual learning, which is very related to reinforcement, learning, but is in some sense, in some ways goes the next step of not just learning a single task from experience, but learning an ongoing stream of skills and knowledge and tasks over the quote lifetime of an agent. So learning continually, not just, you know, learning to achieve something and then say, now I know how to do it. You know, he approached me with the idea of founding a startup company to take that kind of idea to industry to see where we could go with it. We were joined also by a longtime colleague of mine, Satinder Singh.

Well, we met early on. We were colleagues at AT&T Labs research back in the late 1990s and early 2000s. So the three of us teamed up to found this code we called a Kojitai, a startup company, and continual learning and hired up a fantastic team. And then ended up in 2019 being acquired by Sony. And this is what led to the founding of Sony AI. So Tinder and Mark both went different ways by either before or right at that time. But I continued on with many of the employees that we had brought together at Kogitai, some of my star former students.

We sort of seated a new organization within Sony called Sony AI. And that is in collaboration. The leaders of that have been Pete Werman, who was a part of Kodgatai and is one of the founders of Kiva systems, the fantastic robots that transformed Amazon warehouses. So he has been a part of the team since the beginning. And also Mika Spranger, been within Sony for many years. and Hiroaki Kitano, who I've already mentioned as one of the founders of RoboCoh. And so, you know, we all came together to form Sony AI with our initial challenge of taking

the technology of continual learning and reinforcement learning to Grand Turismo, the really high-fidelity racing simulator that is a very successful PlayStation game of Sony. You mentioned this in the introduction. We eventually, three years later, so we started, I guess, we officially launched in 2020, And then by a couple years later, we had an article on the cover of Nature featured G.T. Sophie, our agent, that is able to outrace the very best human drivers in Grand Tresmo. People who've put in their 10,000 hours to become experts. People take this very seriously. There's international competitions.

People use it as a training ground to become professional race car drivers. And it wasn't known if an AI agent could ever do better than the best human participants at this real-time control. task. AI agents that could beat people at chess and checkers and poker and jeopardy and other games, but it hadn't really be done in such a physically realistic real-time control task. And so it was, I think, a real landmark result. That's why it was featured on the cover of nature. And it was also very important to one of the products of Sony. So it was a perfect project in some sense that it was a landmark scientifically and also useful to the company. And so that's is one example of the kinds of things we try to do at Sony AI to do really ambitious projects

that require some of the top researchers around the world. We've been fortunate to hire a fantastic team, now a couple of hundred people working on several different problems, each of which either is it succeeds is ambitious enough to be a real scientific landmark, the kind of thing that could lead to an article in science or nature, or is directly useful to the company, but ideally both. As chief scientist, I serve as the scientific advisor to all of these projects. They span the gamut, though the GT Sophie one was our first one that we were able to go public with. The second one was featured on the cover of nature just in December of 2025, so just a month or two ago from this conversation. And it was from our AI ethics team that's led by Alice Jean.

It was the description deployment of a system called Phoebe, which is a data set of, of images that are collected in a completely transparent and ethical way in the sense that all participants are compensated for appearing in the data set. The labels that describe them are all self-described. So it's not an external person saying what the race or nationality of a person is, but the person who's shown in the picture self-identifying. It's very diverse in the sense of people from, I think, 81 different countries around the world, different ages, genders, everything balanced in the way, and it's meant to give a blueprint for how can we create ethical data sets for artificial intelligence, as opposed to just scraping the web and using whatever

happens to be out there on Reddit and Instagram and things like that. And then also to sort of identify biases in existing methods. It's not a large enough data set to train a new vision model, but it is large enough to evaluate existing models, and that's how it's used in the paper. So that's a second project from Sony AI that was able to land on the cover of nature. We have another project that has actually been accepted as a paper in nature. It'll be, it should appear within, probably around March, March or April in this year of 2026. Watch for that. Another, I think, really big landmark result. And then we have about, you know, five or six other projects within Sony AI that are all, you know, very ambitious scientifically and also, you know, are tied to

the interest of Sony, which is an entertainment company that, you know, in many ways. Sony does a lot of different things. There's hardware cameras, but there's also Sony music and movies and PlayStation. A theme is trying to bring fulfillment and happiness to people who interact with technology. I really resonate with that message. The business isn't about getting people to click on ads and buy things,

but really about sort of having a fulfilling and enjoyable time in life, looking for ways to bring cutting edge artificial intelligence to that aspect of humanity. Yeah, that's amazing. It's fascinating. I think the work that you just spoke about in terms of Phoebe and building ethical data sets, I mean, such wide impact, right? I mean, it serves as a great example. It serves as a way for people to sort of model and mimic that and, you know, continue to build on that. It's amazing that it is also in line with what the organization, that's, you know, that's an actual business that has to sort of meet its goals. Thank you for sharing that, Peter. Yeah, and Sony really prides itself on being it as a, has a, a rating as one of the most ethical companies in the world for several years in a row. And I think, you know, this is this aspect of wrestling with how can we use artificial intelligence to advance technology, but also well keeping people in control of, you know, the creative process and doing it in a way that keeps humanity centered. That is really a core principle personian.

You know, this isn't just an add-on for the company. It's really important to be able to think about how to use artificial intelligence in a way that magnifies humanity. I think also the hallmark of a great research lab is one. You're exploring brand new ideas that can actually set off a series of other research questions and explorations while also finding a way to actually apply this to real world situations and products and services that people want to use. One other area that I wanted to talk about a little bit is you strongly believe in undergraduate research. I recently saw a video of yours. You're fueling a lot of programs within the university for that target audience.

Would you like to talk a little bit about that and also areas that you think, you know, students should be focused on? I also read a little bit about your recent paper on AI literacy. So just curious more about your engagement on those aspects of your work. Yeah, I'd be delighted to. Thanks for the opportunity. In today's world, AI education is critical for everybody at different levels. There's many different populations for AI education. When you talk about undergraduate research, I've been, had the good fortune to be involved in a really innovative program here at UT Austin called the freshman research initiative that introduces

first year students in the spring semester of their freshman year to research in sort of a class setting but introducing them to the skills they would need to do research in a particular lab. And then the students who are successful go on to do research projects in the fall of their sophomore year and then to mentor the next cohort in the spring of the sophomore year. And the ones who are successful end up then having another couple of years to be engaged, in research for their undergraduate career, which is very different than, you know, in many places where undergraduate research happens in the last semester or last year of the career. So this is a way of nurturing and making undergraduate research really a large part of the undergraduate experience.

This happens across the College of Natural Sciences in UT Austin, so not just in computer science, but there's streams in physics and chemistry and biology. I've led one in computer science on autonomous robots since 2007. It's now being run by Justin Hart, but, you know, it's been very successful at identifying some really amazing students. That's within the computer science majors. Connected to that, we also launched recently a honors program in robotics that students can apply to from high school that allows them to get into this freshman research initiative with some new streams in robotics and also to take a minor in robotics.

And we had an immense success. It launched for the first time this year and for the 50 spots. So I think we had more than 3,000 applications. So we're hoping to be able to expand that. We have also an online master's program in artificial. intelligence that was launched recently. We actually have three online master's programs that service right now almost 5,000 students, one in computer science, one at data science, and now one in artificial intelligence. This is allowing us to a very affordable price of $10,000 for the whole

master's, so $1,000 a course, and it's allowing us to really at-scale, educate professionals who are trying to go back and upskill in the technical aspects of artificial intelligence. I teach the reinforcement learning course in the program. We have all kinds of courses there in that. been very successful. And I should say, you know, one of the unique properties of UT Austin is our size, our scale. So we're a top 10 department, very high quality, but we're the largest top 10 department, I believe. So quality at scale, it's the biggest combination of quality and quantity. And so we do that both at the undergraduate level. We have more than, I think, 2,400 undergrads currently, and then this online master's where we're scaling up to thousands of students. So those are for the computer

science majors and master's students. But nowadays, I think it's, you know, as an educator, it's imperative that all students, regardless of discipline, whether they be in the humanities, in fine arts, social science, that they all have at least minimal literacy in artificial intelligence to be able to understand what's realistic, look at the headlines, and think what's height and what's grounded. And so together with colleagues, especially Jouillejit Biswas and Don Fasel, we launched a course on the essentials of AI for life and society, which is designed as an undergraduate course for anybody. It scales up to hundreds of students. We hope it'll actually go to thousands in the coming years. All of the materials are online. And I think they're appropriate, by the way, for, you know,

anyone from upper high school to adult to lifelong learners. I think, you know, there's no programming background needed. You don't need to have come in with any prerequisites. It's an introduction to what is this term artificial intelligence, what's the history of it, what have been the different paradigms, what are the societal impacts. There are some homework assignments that are all online, but they're not, you know, they don't require programming, as I said. I've been really excited about this as, I think, a responsibility. I hope that we get to a point where every university student for sure, and maybe even every high school student, has at least one course in artificial intelligence.

And we've designed this course to be that one course. If it's the only course someone ever takes in artificial intelligence, this should be the one, I think, because it'll give them that grounding and perspective about what the field is all about. And it can also serve as a launching point for deeper interest if somebody does want to go on then take more in-depth courses. So all, like I say, all the materials are online from my webpage. It's easy to find them. I would be delighted if anybody can take them and repurpose them for their own courses. I'd love to hear about that directly from people doing that. We've written

an article about the course. A couple articles now in EAAI, the Educational Advances of Artificial Intelligence Symposium presented it just last week, the second paper on it. And so we're trying to, you know, not only did we create the course, but we're taking big efforts to make it so that it can be impactful and useful for everybody, and I hope people will take it and run with it. Oh, fantastic. I mean, I think, you know, one, reducing that barrier for entry into artificial intelligence, because there's a term that's bandied about an espit for somebody who's not even in technology, it can be intimidating, and for actually providing, you know, material and ways in which people can take this and sort of propagate it even further. This has been an absolutely

riveting conversation, Peter. Thank you so much for taking the time to speak with us at ACM Bycast. My pleasure. Thanks for. inviting me. ACM Bipecast is a production of the Association for Computing Machineries Practitioners Board. To learn more about ACM and its activities, visit acm.org. For more information about this and other episodes, please visit our website at learning.acm.org slash bikecast.

That's learning.acm.org slash B-Y-T-E-C-A-S-A-A-S-E-A-S-E. Thank you.