[00:00:00] Matthew Johnson-Roberson: We don't live in a world where you look out your window and you see robots every day.

[00:00:04] Brian Heater: We're open to the possibility certainly that there will be a similar unexpected breakthrough in physical AI, but we certainly can't count on it happening.

[00:00:13] Matthew Johnson-Roberson: Now, for robots, it's like, well, what is the next thing I'm predicting? Is there something that's that sort of atomic, that small, that I can do and parallelize and put into a lot of GPUs and train against?

[00:00:30] Brian Heater: Matt, are you a vibe coder?

[00:00:32] Matthew Johnson-Roberson: A hundred percent. Not even am I a vibe coder, my whole desktop, if I were to show it to you, is 100 windows of coordinated agents writing software for me. I haven't directly written a line of code in three months. We are trying to use AI agents to do all of the stuff that sucks. I am 10 times faster. My ability to make decisions, think clearly. The biggest advantage that higher ed has over doing R&D in any other context...

[00:01:02] Brian Heater: Hello, everyone, and welcome to another episode of Automated. My name is Brian Heater. I am the managing editor of the Association for Advancing Automation. This week we are catching up with Dean Matthew Johnson-Roberson. The last time we spoke, he was the director of the Robotics Institute at Carnegie Mellon. He has since left that job and is helping to build a brand new College of Connected Computing at Vanderbilt University in Nashville. He's also not ever too far from startup land. He co-founded the last-mile delivery firm Refraction AI, and more recently, the healthcare startup Patients.app. Thanks to Matthew. Thanks to you as always for listening to the show. If you're watching this before June 24th, 2026, please come out and see the third ever Automated Live on stage with Ali Agha of Field AI at Automate in Chicago. And with that, please like and subscribe, and enjoy this conversation with Matthew Johnson-Roberson.

[00:01:55] Brian Heater: So I had a realization about 10 minutes before this conversation started because I associate you, or have associated you, with CMU. Obviously, the last time we spoke you were at CMU. You were there for a while. I was on a business trip earlier this week. For 27 hours, I was in Nashville. And now we're on the call and it dawned on me, oh, I could have stayed a little bit longer and we could have actually done this in person.

[00:02:35] Matthew Johnson-Roberson: Yeah, it's funny. I have a whole setup here. It would've been great to have you in the office, but it's great to speak to you virtually.

[00:02:41] Brian Heater: One of the questions I asked when I was over at Dexory - do you know Dexory? They're a London-based inventory company. They built what they say is the tallest AMR in the world, which I have no reason to debate 'cause it's 60 feet tall. They just opened up a large testing facility in Nashville. And one of the things I asked them is, what's the talent pool like there? Obviously, you're coming from Detroit, or I guess Ann Arbor, and then Pittsburgh - really deep pools there in terms of automation and robotics. So what's the community like over in Nashville?

[00:03:39] Matthew Johnson-Roberson: It's great. It's certainly different than what we had in those other two pools, but there's a pretty vibrant tech scene here. We've had some big corporations move recently, and like I was in the last two places, I'm at a university here, Vanderbilt, and I will say, that's the through line. If you meet sort of kids that are getting into college today, they're the best talent pool that I've ever seen. Certainly I don't think I would get into college now, given how challenging and how accomplished undergrads are these days. But no, it's a great question. Nashville's been one of these places where healthcare has obviously been the dominant technology for a long time, much like robotics became dominant in Pittsburgh. And so we've been thinking about what does it mean to have a tech-focused ecosystem, or at least part of one, here in Nashville. That's the question before us, and the hope is that what I'm trying to build here with the university can be the accelerant on what is already a growing and more vibrant AI and computing hub.

[00:04:51] Brian Heater: That's interesting as well because I'm now realizing that healthcare has also been kind of a through line for you. Pittsburgh, again, you've got Pitt over there. You've got world-class healthcare facilities. And your current startup, you actually have a healthcare startup right now.

[00:05:10] Matthew Johnson-Roberson: That's exactly right. We have a startup called Patients.app. It's an AI healthcare startup designed to help clinicians review lengthy medical records and make medical decisions faster and better. I guess the commonality is that my wife is an anesthesiologist, and finding employment for both of us in the same city means it has to be strong in healthcare and strong in tech. And either fortunately or unfortunately, the number of opportunities to do that is pretty narrow. So I kind of bounce between places with world-class healthcare facilities and growing and ever-improving technology scenes.

[00:05:52] Brian Heater: So is she connected to the startup in some way?

[00:05:56] Matthew Johnson-Roberson: Yeah, she serves as a medical advisor. She's helping us out there. Really the genesis with this startup was actually watching her workflow. She'd come home and then have to do a bunch of charting in an electronic health record. And as opposed to hanging out with me and my daughter, she'd be sitting there finishing charts from that day. And I thought, there has to be a way that AI can make this better. So very selfishly, that was why we decided to build this startup. At least my motivation for building a startup was I'd like to reclaim some of that time to spend with my family as opposed to sitting there entering stuff, watching her enter stuff in EHR.

[00:06:36] Brian Heater: You had that infomercial moment - there has to be a better way?

[00:06:39] Matthew Johnson-Roberson: That's right. Exactly. I don't know if you have any physicians in your life, but to watch what it is for them to have to use an interface that looks like it is from 1992 and click through about 100 to 150 screens just to close out one patient, I thought, this can't be the future.

[00:06:59] Brian Heater: I'm hoping that - obviously I follow automation really closely, and over the past - actually, the last time, when you and I really first started speaking was right around the pandemic, and we were seeing the ways in which the pandemic was transforming automation and accelerating a lot of things. I would like to believe that the pandemic has had a hopefully somewhat positive effect on, I guess I would say, the American healthcare system. That's a sticky subject right now. But we've seen a little bit of automation there. Beyond robot surgery, there are maybe a handful of interesting companies in the space right now. Has there been progress, technological progress?

[00:07:52] Matthew Johnson-Roberson: Certainly some, right? If you ask me, I think that the biggest thing, at least over the last couple years coming out of the pandemic in terms of transformative technology in the healthcare space, was sort of AI scribes. That is an area that grew really quickly, has taken a ton of land, and now is becoming the standard of practice for almost all the major academic medical centers and increasingly in smaller facilities as well. So that I think is gonna help. But one of the challenges we realized is that even if you can get the entire - so the way this technology works is that your doctor will whip out your phone or their phone, and they'll put it down and say, are you okay with me recording? They'll record, and then everything you say to the doctor gets transcribed and hopefully turned into something they can use as a note to register this encounter. We used to do this with humans, traditional scribes, and now there are these AI versions of them. But one of the things that you still see is that much of the healthcare system is filled with inefficiencies that it's not that we don't have technological solutions for, but it is just a much knottier problem in that faxes are still the dominant way we exchange information between healthcare facilities. And that's crazy to think about. It's not as if we haven't built better technology than the fax machine, but regulation and risk aversion and concerns about privacy mean that fax machines still dominate the way we move records around. So I think there's a lot of headroom for technological growth in the healthcare space. And in fact, one of my biggest lessons coming out of the pandemic was that building a hardware company is just hard. It is a grind. And I thought, whatever the next thing I do is, it's gotta be something where we can flip a couple switches on AWS and then have it take off, as opposed to grind it out, building generation after generation in a factory.

[00:09:52] Brian Heater: Sure. I mean, obviously hardware is hard. Obviously robotics are incredibly hard, but you are as familiar with anyone as the kind of 99.99999% problem because you come from autonomous driving. You had your startup in the space, and then again, you come from Ford, Detroit - you come from that world as well. And I assume that healthcare, especially when we're talking about AI in healthcare, has a very similar 99.99999% accuracy issue.

[00:10:25] Matthew Johnson-Roberson: Exactly. This is the reason why it felt like an interesting place for us to do some experimentation in next. Coming out of self-driving, one of the things that became really clear is that, to your point, you need to have a very, very reliable system that handles edge cases out to many decimal places. But secondly, I think a lot of the lessons learned for me - and I think that we've learned as a community in the self-driving car space - is that in safety critical technologies, there's both the technological challenges, but there's also a lot about acceptance, about safety, about working in a heavily regulated industry. And that knowledge, I think, is really important for anybody that wants to innovate in any of these spaces. I'll include transportation in that, in the same way I would include healthcare. When people's lives are on the line, it's really, really important to bring an ethos of responsibility. And people that work in these legacy industries in many ways have come up with that as the way that they think about everything. So it's not just transportation. You can look at mining, you can look at oil and gas. These are all industries that have factory work. These are all industries where you put on your high-vis vest, you go through a very detailed safety onboarding just to take a factory tour. And that level of rigor is obviously something that I think is really important, and certainly really important when you're dealing with a safety critical industry.

[00:11:54] Brian Heater: Legacy - and maybe I'm overstating this, but the word legacy or the phrase legacy industry kind of strikes me as being a situation in which they might perhaps be almost hesitant or reticent to adopt a new technology.

[00:12:12] Matthew Johnson-Roberson: Yeah. If I think about self-driving, it was a great example where it really forced the auto industry to think deeply about, well, what degree of our workforce, what degree of innovation is necessary for us to remain relevant as a company? And I think that actually pushed - even for consumer cars that aren't really focused on self-driving features - it pushed those platforms and those companies forward probably a decade faster than they would've gotten up to speed otherwise, and I think that's a great lesson. Healthcare is such an interesting one because there are lots of things we could do to make our healthcare system better. Some of those are technological, and many of them are policy and much more knotty and complicated. But I think at the end of the day, all these systems have a lot of commonality, and I'll put education in the same bucket as well. They involve a lot of people. If you're building cars, that involves a ton of people. You have a massive workforce. The warehouses have a ton of people in them, even today, even with the levels of automation we're seeing in car assembly and manufacturing. And healthcare is probably the greatest example where that is the sector of the economy that's still hiring while other sectors have slowed down, and that's really because the delivery of healthcare is still a very human business. We can help with technology, and we can do other things in robotic surgery, but there are eight, 10, 15, 20 people in the room when you're doing a surgery. There are people bringing stuff in, bringing things out, monitoring different subsystems, all kinds of stuff. So it's just a very labor-intensive business, and the same has been true for education. We've gotten all kinds of things to help us teach better and educational tools, but my biggest bottom-line expense at Vanderbilt here is still people. And all these efficiencies have made my people better, more effective - my faculty better researchers - but we're still using people all the time. And so I think that's sort of the through line. There's a lot of value in what we can do with technology in these different areas, but I think it's still incredibly dependent on lots of humans. And so that's been a big rate limiter for change.

[00:14:19] Brian Heater: You touched on something interesting in there as far as the auto industry and self-driving cars. I'm curious to get your take on this, as far as whether the large automakers, whether or not they had specific success in the self-driving space, whether their efforts still moved them forward individually.

[00:14:49] Matthew Johnson-Roberson: I mean, this is just my opinion. I don't know. But what I would say is that we looked at who they started hiring during those periods when they got really worried about the degree to which self-driving was gonna eat into their core business. And they brought a lot of people in who I think had a very - I guess for the auto industry - sort of radical mindset, and thought a lot more about rapid iteration, thought a lot about how technology can be an accelerant. And we've seen also internationally just an incredible amount of pressure on Western automakers from the rapid expansion and advancing of Chinese automakers. You just go to China now and you get into any car and it feels like you're 10 years in the future from what you can buy off the line here. And I think all those pressures have forced the auto industry, to some degree, to think a lot about what type of people do we need to hire? What type of executives? What type of group leaders? What type of engineers? And in the era of AI, I'm hoping that all comes together in a way that pushes that industry forward with the same velocity and the same urgency that the first boom of self-driving cars in the last 10 years did. I think it still remains to be seen, but you can go look at any major Chinese automaker and you can see the impact of self-driving on the entirety of that vehicle stack - everything from payment systems to personalization to entertainment to navigation to even the self-driving features. And it's very impressive.

[00:16:40] Brian Heater: I've gotten a much better insight into manufacturing than I had in previous positions. And something that I've learned, which is now very obvious in hindsight, is when introducing new technology into both manufacturing and logistics, there is hesitancy and there is pushback. Basically they want you to disrupt things with minimal disruption. Minimal downtime. A factory wants to continue to be able to produce things while introducing a new technology. And I'm wondering if there's an analog there for all these other industries that we're talking about.

[00:17:23] Matthew Johnson-Roberson: Yeah, it's a great question. I guess the best example I had of this was when I got to tour some of Amazon's fulfillment centers. They were on their seventh or eighth generation of AMRs driving around - the Kiva robots. They were able to show me the whole history of how they had gone. And to me, it was the greatest distillation of when a robotic system is clearly really useful and it drives the economics for a company, that you can bring a very disciplined and iterative approach, but you can keep the factory floor humming while bringing out what are eight or nine generations of a vehicle and have them swap in and continuously improve. So companies that are doing this really well have brought that continuous iterative design principle to the physical world. That I think is really what's necessary to both continue to innovate but also keep the lights on. Because it's not an example where you can shut everything down, turn a fulfillment center off or a factory off or a warehouse off for six months to upgrade it. That's really not viable when you're talking about the amount of capital they have sunk into these things. And so again, I think anybody working in the hardware space - there's just a real power to being able to have something run, learn what you can from it, and then upgrade it as it continues to proceed.

[00:18:54] Brian Heater: I know that all these things tend to be sort of cyclical and they ebb and flow, but it strikes me lately - I do a newsletter and I've been interviewing a lot of founders and a lot of CEOs, and it seems like last mile delivery companies have been getting a lot of press again. I just interviewed Ali from Serve, and I interviewed the CEO of Coco. You were doing that, or trying to do that, a long time ago. In a certain sense, was Refraction just a little too early?

[00:19:30] Matthew Johnson-Roberson: I think so. I don't know. It's a great question. You try to do the postmortem and figure out what you would do differently. I think it just comes down to there's a really hard problem in scaling. Let's use the last mile companies as a good example. Even if you look at companies that are incredibly well-capitalized in the full-size self-driving space, like take Waymo. Think about how slow and deliberate their rollout is and how long it takes them. They're actually driving around here in Nashville now, but they're not offering rides. They've been driving around at least for the last six months, maybe a year, I guess just doing mapping presumably, in preparation for a launch here. But that's a lot of investment and time, resources, and people to maybe field - I don't know what they're gonna have here - let's say a couple hundred cars.

[00:20:23] Brian Heater: And to back up and say, they are a Google company, so they have the resources to do that.

[00:20:28] Matthew Johnson-Roberson: Essentially almost infinity resources relative to anybody else playing in this space. And even there, they're talking about bringing on - I see probably 50 or 60 cars floating around here - but that's still a pretty modest number when you think about the size of the ride-hail business. And then look at the last mile delivery space. We just haven't seen or figured out exactly what the formula is to scale to tens of thousands. I'll give a shout-out to Starship. Starship was certainly the furthest along of any of these, and they were pretty widely deployed on campuses. I don't know if they still are. But that to me was an example where even there we were talking thousands of robots, not hundreds of thousands. And probably in an average day here there'll be - I don't know, I got 10 things delivered this morning. I got these mics delivered this morning. Amazon came to my house three times. I'm honored. So there's probably hundreds, if not millions, of deliveries happening, and the question becomes, well, how do we get to that level of scale? I'll keep going back to when I saw the level of complexity and the rollout of those Kiva robots in these Amazon fulfillment centers. That's a 15-year process for them to get to the point where they have a million of them driving around. And I just don't know what the arc is gonna be for last mile. I'm still a big believer - there's no reason to think we can't do this. It's just it always feels like it's a couple years away.

[00:22:03] Brian Heater: Yeah. There are so many unknown unknowns when it comes to that, especially when it comes to urban spaces and interacting with humans. That's the biggest unknown of all. Was it Serve, I think, that's trying to operate in downtown Chicago? It's a very densely populated area. It's really hard to plan for that.

[00:22:31] Matthew Johnson-Roberson: Yeah, I think it's one of the big challenges. This is the thing I think about a lot - we don't live in a world where you look out your window and you see robots every day. And I don't know how long it's gonna take until we get there. I think the biggest example of this now is that in many cities, you look out your window and you see a Waymo or a Zoox or pick your favorite full-size self-driving car company. And I think that's a big step forward. So I'm hoping that's part of a larger arc where we will see robots moving around, doing useful things in the next couple years. But again, until that takes hold, a lot of what we're seeing is sort of experimentation. And I just hope that one of these lands and hits and that then we really get to see the scale that we've been promised.

[00:23:29] Brian Heater: You said something similar to that one of the first times that you and I spoke, and that has stuck with me, especially lately because one of the key things we've been talking about - we're now using the term physical AI. I think we all sort of settled on that as the phrase du jour. And as I'm interviewing all these physical AI companies, our kinda recent pet topic is the flywheel issue. And that I think speaks to that really well. We're in this weird space right now where we have the problem of we don't have enough robots deployed to collect enough physical data to train them well, but we also don't have good enough physical data to deploy that many robots.

[00:24:19] Matthew Johnson-Roberson: Right. Yeah, look, I think this is a good example of where there's this big gap in what we're promising and what we're gonna be able to deliver on the physical AI axis. I have two schools of thought on this. The first is that it is incredibly promising as a narrative, and even as hopefully a reality, that we would be able to move away from the world where we need to, for lack of a better term, hand-design every algorithm for every problem we have in robotics. We just say, okay, well, here's the thing. Let's get a team of engineers on it. Let's work through it. Oh, that doesn't work. Try this, try that. The promise of this idea that we'll get enough training data and then we'll have a general purpose intelligence for physical systems that will allow us to solve many problems is incredibly promising. And I think that's become more promising in that we've kind of gotten there and been able to achieve that with language models. Certainly not all the way there, but the capabilities of these language models relative to what we thought they were gonna be capable of, or what we kind of were explicitly training them to do, has been incredible. So that's been game changing in many ways. And that's the promise for physical AI companies. But you kind of highlight where the big gap is. We got onto some flywheel with language models where we had enough data, we had enough of an understanding of how to train over that data with an objective function that expressed something that we could optimize against and get better at.

[00:25:59] Brian Heater: I would have the major caveat there that that data was all preexisting, 'cause we had libraries and books and the internet.

[00:26:05] Matthew Johnson-Roberson: Agreed. And look, it's not that that's the same, but one of the things that I struggle with when I hear people say, well, we don't have enough robot data, is I could take a robot, put it on the street today, walk it around for six hours or 12 hours or 24 hours, and have terabytes. Hundreds of terabytes of video data, of encoder data. So I would put the problem more specifically on we don't have an objective function that we can optimize over that tells me how to use that data efficiently or effectively. It's not the volume of data in my mind, but it is, what's the utility of that data and how do I build something that learns on it and learns on it well? And no one has answered that question to my satisfaction where they say, oh, here's the objective function, and we just need to go gather a bunch of data to optimize that objective function. It's not clear what that objective function is for the physical world, and it's also not clear that even if we had it, that it can be sufficiently optimized against. So that's the big open question in my mind there.

[00:27:23] Brian Heater: Can you expand on that a little bit? When you say objective function, what do you - I mean, obviously it's hard to give an example because we don't have one yet.

[00:27:31] Matthew Johnson-Roberson: I'll try to make it really concrete. Not to oversimplify the last decade of incredible work in large language models, but essentially you can predict the next word that's gonna come. And that's a very simple objective function to express and to optimize over. And it turns out if you do that over an internet size of data, you get things that are really smart. That's a very simple distillation of a lot of the work that happened over the last decade. Now, for robots, it's like, well, what is the next thing I'm predicting? Is there something that's that sort of atomic, that small, that I can do and parallelize and put into a lot of GPUs and train against? There are examples of what we're doing. Can I predict where this arm should be, where this leg should be, joint angles? Can I predict the next frame coming out of the camera? Can I predict X or Y or Z? But I promise you, none of those are as atomic or as easy to represent or as easy to test against as is the next word. And so we just haven't come up with that thing, and so it's not clear to me that more data is gonna get us out of that problem. Until we have that thing - whatever the next best word problem is for robotics - until we have that, I'm not sure that all of the trees that we're setting on fire to burn lots of GPU hours to train models is gonna bear fruit. But I've been wrong many times before in my career, and I look forward to being proven wrong in this case as well.

[00:29:10] Brian Heater: Well, the burning planet conversation is a good one, but probably one for a different day and a different podcast. That's interesting, 'cause the next word problem - it's universal and it's consistent, and it's something that you need with every single kind of large language application.

[00:29:34] Matthew Johnson-Roberson: Exactly. Despite all the complexity and everything we're doing, if you go to the frontier model companies - go to Google, you go to Anthropic - they spend much of their resources just improving the models on that specific problem. And then they do post-training and lots of other cool stuff after that. But what they spend whatever, $400 million training last week on, is trying to get a better next best word predictor, for lack of a better term.

[00:30:00] Brian Heater: It's like a solved problem that's still being refined.

[00:30:04] Matthew Johnson-Roberson: Exactly. They are continuing to bear fruit in making the models better at doing that. And it feels to me that we're still casting around for what that problem is in robotics. And until we put our fingers on it, we may be going down rabbit holes that don't actually end up getting us our universal robotic brains that we're hoping for. Now, if we stumble on that and we get the same kind of unlock that the language models did, that's game-changing. So I think it's still a worthy place to spend lots of time, resources, and have lots of smart people chasing after. It just feels pretty non-linear to me. If we find it, we can spend billions of dollars training against it, and if we don't find it, we can spend billions of dollars on training and get models that don't really do anything terribly useful.

[00:30:54] Brian Heater: Listen, I was a creative writing major at UC Santa Cruz, so a lot of this stuff is foreign to me, or at least I'm coming at it from a very different angle. But I always appreciate, especially with things like LLMs, the surprise. Even the people who are working on them - the whole black box thing, and the fact that they're like, oh, these work a lot better than any of us expected them to. And that to me is a great example of that Arthur C. Clarke quote that everybody loves, that sufficiently advanced technology is close to magic.

[00:31:38] Matthew Johnson-Roberson: Agreed. I use LLMs every day to do all kinds of things now, and I continually am delighted and surprised by the degree to which they seem to, A, be very intelligent in whatever way you wanna define that, but really probably more importantly than that, very capable on tasks that, in my mind, I am pretty confident they have never been trained on. I'm coming up with problems that are just in weird parts of the space that I know there's just not a lot of training data on the internet for. That's the nature of university research. We're trying to do something new, and their ability to solve problems that I know are unsolved in the literature is really exciting. And I guess that is what kind of fills me with renewed optimism that even if we haven't settled on what the solution is for physical AI yet, we're in such a better world when it comes to tools for exploring what that could and should be. Stuff that used to take me a team of graduate students - I'd put all of them on it, and they would spend all their time implementing stuff and spend very little time thinking - that is totally flipped on its head. Now you can have coding agents implement all the stuff you want, and we can spend all the time thinking about, well, what's the next thing to try? Is this actually working? Why isn't this working? And even getting to spend the mental energy or the time to focus on that, I do think is gonna bear long-term benefits for research, for innovation. I don't see how it can't.

[00:33:21] Brian Heater: So I guess if I have to put a really fine point on it, then it would be that we're open to the possibility certainly that there will be a similar unexpected breakthrough in physical AI, but we certainly can't count on it happening.

[00:33:41] Matthew Johnson-Roberson: Yeah. I think that is the exact distillation. If we're expecting that it's gonna follow the exact same path as language models, I think we're gonna be disappointed because there's already enough data and indicators that it's on a different trajectory. Now, not to say that it can't follow some nonlinear path that we haven't seen yet, but the idea that it's gonna be easy and it's just gonna be another thing that we do, like what we did with language models, I think is rapidly becoming not the case.

[00:34:12] Brian Heater: I got the sense, the first couple times we spoke - it was pretty soon after you were transitioning from startup world to getting back fully into academics - that it was sort of like, all right, I'm never doing that again. And as I'm hearing you describe this decision to get into the app world, to get into the AWS world, it's sort of like, all right, I'm getting into that again, but I'm easing myself back into it.

[00:34:44] Matthew Johnson-Roberson: Yeah. Well, look, two things have really stood out to me. The first is that I love robotics research, I love building physical things, and I love testing them. The lesson in my mind is that that is super fun in the academic research space because the stakes are very low, and it is all imagination and excitement and not terribly stressful - to the degree that any startup can be not stressful, which I think is basically none of them, but to the degree that it can be less. Scaling a hardware business is two types of stress at the same time. The way I've been kind of bifurcating my brain lately is that because I think there's this problem we have yet to still solve - which is, let's figure out what the objective function is for the physical world for training AI models - that seems like a super interesting and very exciting question to try to answer with a bunch of smart people and a whiteboard in an office. It seems like an incredibly stressful and incredibly challenging question to answer on the fly when investors have given you lots of money and they would really like to see a return on their investment in the short to medium term. And so I'm really excited about trying to take technology that's here, particularly things like language models, and figure out how they can really greatly advance all the industries that we know. So not just healthcare, like we're working on with the startup, but also here at the university. Here in the College of Connected Computing, we are trying to use AI agents to do all of the stuff that sucks about being at a university as a professor or a student or a staff member, and so we're trying to automate all of that. We're building a new college from scratch, which means we have all this blue sky to just make everything new, and that's been great and very fun and exciting. But on the research front, we are deep in trying to think through these really hard problems, which is, well, if we want to train an AI model in the physical world, what are the ways that you gather data to do that? What are the ways that you look at that data and try to understand, is the model successful? Is it working well? And how do you do that without it being incredibly expensive, both from a time perspective and also from a risk perspective? So I like that kind of bifurcation where I think startups are great for scaling. Startups are harder to answer open-ended questions that could take a decade. So my only advice to any founder is, if you are trying to solve a problem that you don't know the solution to already, or no one knows the solution to, that's a naughty one. Good luck. If it's a decade off, a startup is probably not the right vehicle to pursue it, would be my pitch.

[00:37:36] Brian Heater: So you're saying there isn't the same inherent stress with publish and perish as there is with learning?

[00:37:41] Matthew Johnson-Roberson: There is, but it's nowhere near. Look, it's small potatoes over here on the academic side of the house. You know what I mean?

[00:37:48] Brian Heater: Also tenure - I feel like tenure probably makes a bit of a difference.

[00:37:51] Matthew Johnson-Roberson: Tenure is an incredible mechanism. We have a system that is designed to allow you to - the biggest advantage that higher ed has over doing R&D in any other context is that you can come in and basically say, I'm gonna work for 30 years on this problem. And yeah, you gotta publish along the way, and you gotta get tenure and all this stuff, but when I'm hiring new faculty here and they tell me, I've got a 40-year problem I'm working on, I'm like, great. You're never gonna run out of stuff. You're gonna come here, you're gonna teach students. That's a great economic mechanism to allow us to keep the lights on, and you can just think deeply about your problem, and there'll be new students every year that'll wanna think through that problem with you. And 40 years from now, maybe you crack it, maybe you don't. But there are just few other places in the world where you can at the outset say, oh, over the next 40 years, I'd like to solve X.

[00:38:47] Brian Heater: Well, okay, here's a hypothetical analog to a startup. Let's say that you go from being a dean at one of the most respected robotics schools in the world to starting up an entirely new school. Is that not similar to the stresses of a startup?

[00:39:12] Matthew Johnson-Roberson: It really is. And it suggests that despite any of my previous learnings, I don't learn from my past, and I just wanna keep trying the same thing over and over again. No, look, I love the early stages of a startup. I love the early stages of building anything. That is where - there's a lot of specialization of labor. I think where I have the most value to add as a human being while I'm on this planet is that I can help people that are trying to take an idea from nothing and get it going. And then you'd have to ask other people, but 10 years down the road, I'm not sure how good I am at being helpful, or certainly not as helpful. But I think when there's nothing -

[00:39:50] Brian Heater: They keep asking you to do it, so there must be something there.

[00:39:51] Matthew Johnson-Roberson: They keep asking me to do it. I must be doing something all right. But exactly. In my mind, it is this balance between trying to say, okay, what works about what we're doing now, or what works already that's out there? But what can we throw out and what can we start over with? And I think bringing that mindset to higher ed, and particularly at a place here like Vanderbilt, where they've been so open to that, has just been amazing. Higher ed is not known for its flexibility, nor its dynamic ability to change rapidly. And I've been continually impressed with the degree to which they're letting me take risks, letting me try new things. We're saying, hey, everything's on the table when it comes to reconsidering what it means to go to college, to teach in college, to get a PhD, to get an undergrad degree, to learn. It's been a challenging time for higher ed in the last couple years, as I'm sure you're aware. But I think if there's ever a time for us to consider what should we be changing, what should we be keeping constant - and then for computing as well, particularly for computer science education. Since we last spoke, the entire way that I program has changed completely five times. It started out I was using autocomplete and tools that helped you finish off the line of code you're writing.

[00:41:11] Brian Heater: Matt, are you a vibe coder?

[00:41:16] Matthew Johnson-Roberson: A hundred percent. Not even am I a vibe coder, my whole desktop, if I were to show it to you, is 100 windows of coordinated agents writing software for me. I haven't directly written a line of code in three months.

[00:41:30] Brian Heater: Are you a vibes dean?

[00:41:32] Matthew Johnson-Roberson: I'm a vibes dean. The way we're running the college, I have agents helping me make decisions. Look, if you're not using these tools, there is no way you are gonna be competitive with people that are. I am 10 times faster. My ability to make decisions, think clearly, all of that. Now, you gotta apply some human in the loop on this stuff. But while we've been on this podcast, my agents in the background have written 700,000 lines of code. Even if I were to do nothing but do that, I couldn't get that kind of output or efficiency, no matter how I split my day up or what level of coffee I drink.

[00:42:17] Brian Heater: Yeah, I've got a complete aside and something that I could easily just type into ChatGPT, but do you have any idea why the paper is called The Hustler?

[00:42:28] Matthew Johnson-Roberson: No, I don't, actually. That's a great question. I know why we have so many nautically themed things here at Vanderbilt. The nickname of Cornelius Vanderbilt was The Commodore, so that shipping magnate energy has pervaded lots here. But I actually don't know why it's called The Hustler. It's a great question. I should look that up.

[00:42:54] Brian Heater: Are there a lot of Lionel Richie themed - you know, Music City.

[00:43:00] Matthew Johnson-Roberson: It's entirely possible. I don't know if Lionel's recording here, but I can look out my window and see.

[00:43:04] Brian Heater: He kinda had that country thing going on. Anyway - 10 recording studios. I ask this 'cause I was looking into the origins of the College of Connected Computing. It seems like the ball got rolling before they brought you on. They were doing the search for dean. But it can't be overstated, the fact that this is the first new undergraduate school added since the Blair School of Music in 1981. So the first one in like 40 years. That's a huge deal.

[00:43:46] Matthew Johnson-Roberson: Agreed. I think it speaks to the moment we're at, where you can look at really any measure - news coverage, the economy, the stock market, whatever you want - and computing is so central to just everything now. And honestly, I think it was pretty prescient for Vanderbilt to try to kick this off, 'cause this all happened really before we were in the full swings of this AI era we're in now. They started planning, oh, we need a college that's gonna be teaching computing, and hopefully doing that for all of our students across the university. So when I came in, one of the things that got me to take this job, obviously leaving CMU, which is an amazing place - the Robotics Institute was just an incredible home for me for such a long time - but what got me to take this was that I thought, if there was ever a moment to try to build something new in computing, it's now. And if there's ever a time to think about how are jobs, the economy, the future of what everything is gonna look like, when does that need to be reconsidered? And what do we need to be thinking about so that we're graduating students out into an economy with skills that allow them to be gainfully employed, to live long, rich, and fulfilling lives? It feels like this is the moment for us to think about all those questions. My entire workflow of how I do everything in my life has changed so dramatically in the last year that it highlighted to me that somebody who's 18, 19, getting started in their career - if we could give them the ability to go out there with knowledge of how things are changing, with a resiliency to the speed and pace at which technology is changing the way jobs are done, what jobs even exist - that seems like such an important mission. And so that's what the College of Connected Computing is trying to do. It's saying, how do we prepare you if you're an English major, if you're a journalism major, if you're an art history major? Whatever you're studying, how do we prepare you for the fact that computing, and particularly AI, may have upended that industry in many ways that we can't even appreciate now by the time you graduate in four years?

[00:46:06] Brian Heater: The more I think about it, the more I realize that the startup analogy is a pretty good one, except I guess the one kind of wrench in this conversation is tenure. Because I've had this conversation with a lot of people of, hey, why did you leave your job as an executive at Google to launch a startup? You have a kid, and this must have been a really difficult conversation with your partner. So obviously, you had this conversation with your partner where this is a little bit more uncertain, and we're moving cities. Although you were able to move cities to a place where she was able to also continue her career. But it still had to have been a very difficult decision to leave Pittsburgh and CMU.

[00:46:54] Matthew Johnson-Roberson: I loved Pittsburgh. I loved CMU. It is really hard. The thing I keep coming back to - and this is the thing I say to young people when they come into my office - is, life is both long and short, but the short bit of it is that you have very few opportunities to really take risks and to try new things, and when those opportunities present themselves, I think you have to be open to them. Because the thing more than anything else that can undermine you in what you wanna do in the world is just not fully taking the risks that seem interesting or exciting to you, or just not figuring out a way to keep learning and growing. One of the most frustrating things about the way our society's set up is, as you get older and more senior and you have more knowledge and wisdom in your career, you're actually put in a position where you can take less risks. You get to be less creative, and I think that's ultimately really frustrating. And so for me, I thought, this is an incredible opportunity to continue to grow as an academic, continue to grow as a person. If you're not growing and learning and changing, I don't know what we're doing. And so, yeah, again, I've got a four-year-old. We had to uproot her, put her in a new daycare, all of these things. It's not easy. But I come to work every day so excited. I have a whole host of new problems in starting a new college that I just never had at CMU, 'cause that was a college that had been around for 50 years at that point.

[00:48:42] Brian Heater: And just to be fair and to be certain, that's an upside.

[00:48:46] Matthew Johnson-Roberson: Agreed. I come in fired up because every day, what do I get to solve today that I'd never thought about before? And as a leader, man, it gets me in the office and so excited. I turn up. I check what my agents have done. I see what results they have for me, and then I'm ready to go. I'm ready for an exciting day, and that to me is worth all the challenges. But there's a human cost to all this. Look, I don't wanna undersell - and you talk to any founder that's founding anything - it certainly has this cost. It has a cost on your personal life. It has a cost on your emotional well-being, on your free time. But there is nothing like coming in and just not knowing what the day holds but being excited for it. Knowing that you're gonna have six things thrown at you that you've never thought about before, but being so amped that if you can figure them out and you can solve them, that - and to bring a team along with you. That's the other thing - to have a bunch of people here. This is a college now that has about 650 students already, and about 50 or 60 faculty. To try to get them to come along on this journey and say, hey, we're building something here. We're trying to triple in size in the next three years. That's sort of unthinkable in the academic sphere, and it really feels like the classic startup journey. It's a lot of risk for everybody, but if you can convince them and align them all to the same mission, it just feels like the best of humanity to have everyone come in and be firing on all cylinders trying to build something together.

[00:50:24] Brian Heater: That's a really good point, and it's something that I hadn't considered, 'cause I was just coming at this from the point of view of, okay, it must have been difficult to recruit the CMU dean to come over here. But now it's the former CMU dean's job to say, hey, how do we get the potential CMU, MIT, et cetera, students to come to this entirely new school and to try something new? How do we convince this new graduate to come to our startup instead of Google or Meta?

[00:50:58] Matthew Johnson-Roberson: Yeah, that's one of the big challenges. If you're starting something new, you both have the challenge that nobody knows who you are necessarily, and that you're asking them to take a risk themselves. But for us at least here at Vanderbilt, that's the perfect confluence of factors, because the kind of students we want to recruit here are ones that really are saying, hey, I would like to build something new and actually take a risk. There are amazing students all across the country in a bunch of computing programs, and I think there are a ton that those other schools, including CMU, MIT, all these other places, are the right fit. What you get there is really innovative education, but also a real sense that they've been doing it this way for the last N years, and that you're becoming part of that culture. So you leave becoming part of whatever the existing computing culture is at those places. The students that we're trying to reach here, it says, look, you can build something. You get to build a brand new computing culture. I loved many things - I was an undergrad at CMU as well. I love many things about that place. But if you had dropped me in and told me I could build a culture, I think I would've built one that was different than the one I inherited when I started there. That to me is what we're saying to students. There are blue chip computing schools you can go to. There are startup opportunities aplenty. But this will be one of the few opportunities you have to be in on the ground floor with your early equity shares - though you don't get to make any money. None of us make any money, so that's the critical thing to mention. Nobody gets rich out of this.

[00:52:30] Brian Heater: You get 50% of zero.

[00:52:32] Matthew Johnson-Roberson: Exactly. But you get to be part of that inaugural class. One of the things that's blown my mind about higher ed is that people still come back. Our football team did really well this year for the first time in a long time, and alumni came back that had graduated 50 years ago. So in my mind, we're building something where I hope that we have students that come back 50 years from now and say, I was part of the first class of the College of Connected Computing.

[00:53:20] Brian Heater: I was employee number five.

[00:53:22] Matthew Johnson-Roberson: Exactly. And they get to tell that story, and then when their grandkids graduate from a very well-established, well-heeled computing program, they can try to convince them that, look, I had a real hand in being part of this coming together. I'm sure all of us older people will be ignored. But I do think that's an incredible opportunity for young people.

[00:53:46] Brian Heater: It's funny - as you may or may not be aware, before I really started covering robotics, I was a hardware editor at TechCrunch for a long time. I was covering consumer electronics, and something that became very clear is, when it comes to being innovative in the hardware space especially, it's a lot harder to do at a large company because it's a lot harder to take risks. And it sounds like there is a real equivalent to be drawn in the academic world.

[00:54:23] Matthew Johnson-Roberson: Oh, 100%. It's the same thing. And I get it. Look, if you're at a big company that's making money and successful, or you're at a big university that already has a program that - CMU's program is ranked number one. So there's nowhere to go but down. The question is, to what degree is it even smart to allow somebody like me to come in and change a bunch of things? Where are we gonna take the program? So more than anything else - people always ask the question, how do startups end up innovating? Why are they allowed to be innovative? But more than anything else, if you have the ability to fail, that is just a gift because it allows you to take such risks. And nowhere to go but up. We are creating something new. There's nowhere to go but greater reputation, greater success. It is an incredible luxury or gift to be able to try to build something new. I would just encourage everybody that has the opportunity to take any of those risks in any aspect of their life. I don't think it has to be necessarily at their job, but to take risks and try to build something. It feels like much of what the modern era of malaise and AI slop and all these things that we get fed - there's obviously a lot of bad stuff happening. But the opportunity to try to make a better world is just one of the unique pleasures of being human, and I think anybody that can avail themselves of that, that has the opportunity to do so, I think they should.

[00:56:08] Brian Heater: Well, a rare episode of the podcast ending on a positive note. Thank you so much for joining us.

[00:56:12] Matthew Johnson-Roberson: I really appreciate you having me. Thanks.

[00:56:18] Brian Heater: Thank you to Dean Matthew Johnson-Roberson and the folks at Vanderbilt University for helping to set that up. If you've been enjoying the show, please like and subscribe. We are just a hair under 1,000 subs on the YouTube channel. Please subscribe, tell a friend, and please check out the newsletter over at automated.fm. If you're watching this before June 24th, 2026, please come out to our third ever Automated live on stage with Ali Agha of Field AI at Automate in Chicago. It's gonna be a great conversation about physical AI and more. The guy was building Martian helicopters. It's gonna be super interesting. Okay, that's about it for this week. Stick around. We will be back just about this time next week with another episode of Automated.