OpenAI's Caitlin Kalinowski on Robotics, Physical AI & Hardware | AI Rabbit Hole 2025

These are great. Awesome. we are between you guys and lunch, so we will try to keep this lively and engaging. or else. That is so funny. That's not my picture. Oh god, that is But it looks pretty close. So, we'll go with it. It's a bad facsimile. You're much better looking. Thank you. I am so excited to be chatting with you, Caitlin. We have known each other for basically a decade, and have talked about essentially every topic on Earth, but rarely in a work context, never, which is we we met, essentially through a mutual friend, as motorcycling buddies. And so here we are. I run an early stage deep tech venture fund. we unfortunately don't have the privilege of investing in the startup that you are most recently involved with which is OpenAI. but I am so excited for you and your career. And for anyone in the audience who doesn't know, Caitlin has had probably the most illustrious career at the intersection of game-changing design, novel technology, kind of what the future will look like in the next 10 to 15 years, and increasingly geopolitical strife, and we'll get into that in a minute.

but you were the head of hardware at of VR and then AR at Meta and worked on gamechanging products like the Quest. And then before that, you were at Apple and helped bring the MacBook, the original MacBook Air to to light, which certainly set for me the baseline of like what is exceptional hardware, what we should expect from functional hardware. I don't know if you guys remember this. I'm dating myself, but the air fundamentally changed my opinion of what electronics in the hand actually should feel like, function like in enable us to do. And now you're the head of robotics at OpenAI, which kudos. What an exceptional role. I don't know whose phone is ringing, but and it's a great ditty. I'm super happy for you, but maybe you could shut it off. Thank you. what? No. Ah, there we go. no worries. what I am so interested by is that you have spent the better part of the last two decades building exceptional hardware.

Robotics have always fallen short in my opinion as an investor and former roboticist myself of actually landing as exceptional hardware in the market. What made you think that now is the time? Well, thanks for the kind intro. it's great to be here and thanks all for having me. why robotics now is I think a contribution of several trends that appear to me to be hitting all at once. it's unclear when they're going to all take off, but right now a few things are happening which which is kind of interesting. the most tactical is that you can start to really get custom robotics hardware from at scale. this is the beginning of the time where you could say, "Hey, I want to go buy an actuator. I need these specs like and someone could actually build it for you for the first time right now." So, that's that's great.

but in terms of AI, RL is changing how those actuators move the body of the robot. And so up until now we used classical control systems to you know write out differential equations and figure out how to make movement but we're at a point with with RL in AI where we can actually look at movement in the real world and copy it in a way and you see kind of I don't know if anybody saw the Unitree video the Unitree dancing or the dancing robot dog but a lot of this is is RL and that's really fascinating because we can actually approximate movement types and there's a few benefits of that. One is you actually don't need quite as fancy components because you you can use basically mimicry to make the the device move the way you want. but also because and we'll get into this I think a little bit later there's there's acceleration happening and so there these acceleration curves.

One of the the third thing is the sim-to-real ratio changing. And for folks who aren't roboticists or AI people, I think everybody here is an AI. So you guys are probably already know this, but like how much simulation does it take to make a robot do what you want in in in a computer versus real in real life and how much training in real life that ratio is changing. So there's more and more and more sim and less and less and less real required. this is an acceleration curve. So there's all these acceleration curves that appear to me to be adding up. And I think I've told you this, but I have historically as a consumer electronics person worked on a linear curve of of improvements year-over-year. And this is my chance to hop on a exponential curve like all my friends and like you and AI and and silicon world.

So I'm excited about that. May May we all be on the right exponential curves. I want that on a t-shirt. I I think you said something really fascinating. I kind of want to double click on this point because I remember basically the better part of a decade ago, well two decades ago I was working on driverless cars back at Carnegie Mellon and we kind of thought that we might be in a simulated data environment. Then the better part of a decade ago the first iteration of Waymo Zoox basically came out of the gate saying we can simulate the real world. We don't have to drive all these miles. And then turns out Tesla kind of has beat the pants off of everyone in automate in kind of actual autonomous driving because they had the miles driven. And so you're saying now that we're at we're starting to hit that inflection point where the simulated data is starting to get sufficient such that we don't have to have a basically 10 to one real data versus simulate data.

That's that's like that's the reason to jump in on this, right? this embodied AI if you can actually simulate the data because I don't know how we can continue justifying PhDs sitting in warehouses folding t-shirts which is wild to me and if you guys are folding t-shirts and kudos but maybe we should make some videos doing that instead. Exactly. And like the I I don't want to get into details but the perturbation of the simulation is just yielding more and more benefits than than than we thought. So if we believe that we can kind of virtualize these systems such that we're building more clever systems without having first to see them see an an embodied instantiation in real space. how does how does that actually translate to the hardware? Like how do you design hardware or what's different about how you're designing hardware now or thinking about it in a world where theoretically your system's getting smarter as it's out in the world and operating?

Yeah, you want to try to make the hardware smarter in advance. So this is tricky. You're you're like like if you're if you're designing a computer, you kind of know the silicon that's you're going to put on it. Then you figure out the display. Then you figure out how all the pieces go together. But with robots right now, you have to like the thing that the robot can do is going to materially change. There might even be new kinds of AI that don't even exist today that you want to put on it. So, you kind of want to try to future-proof your design as much as possible. And I think the way I think about this is actually my hope is our hope is that if we were to design a robot now and we were able to teleoperate it to do whatever we wanted that our hope is by the time we want to ship it that it will be able to do those things autonomously and that's the way that we're approaching it.

So basically we're trying to create robot hardware that can do everything we would want it to do in a teleoperated sense and then shift to to having it run on AI. by the time it it gets released as much as possible. Yeah, makes sense, right? Because otherwise you're in a lagging cycle and kind of can't afford to be. I I'm deeply pessimistic that this is all a race to the hardware on the race to the bottom on the cheapest possible hardware and this construct that smarter models will make up for worse systems. But as a recovering engineer myself, I just don't see I I I feel like there's a bottom line on how like basic, incompetent, irreputable the hardware can be. What's your take on that? I mean, I'm I'm a hardware engineer by training and like everybody's got hardware that's broken on them because it wasn't good enough or it didn't like it wasn't thoughtful enough or it didn't like solidifi like harden the right parts or whatever.

So, I I don't think there's a race to the bottom on on hardware components. I do think that RL gives us a chance to use less sophisticated high high high high-end components and merely high-end components and that the difference in the delta in cost is quite high between those two. But I don't expect we're going to have like you know $200 robots that can like do things that you'd want them to do other than maybe vacuum and and sweep in which case there's pretty good vacuum robots out there. Our mutual friend Rodney Brooks has been building them for a long time. I I guess that there's one trajectory of can we build the idealized system but then maybe the more interesting question is in the long arc of these systems proliferating our existence or factories or whatever the use case is you probably need a lot of embodied AI which means we need to figure out how to build a lot of these systems and in a world where we're used to building a lot of systems in China and we may not have that privilege for much longer how do you think about about the supply chain problem both from a subcomponents perspective, an assembly perspective, maybe even an engineering partner perspective.

I mean, this is what keeps me up at night. Not the way how we rob design our robot, not the AI actually, although those are very difficult both. It's the supply chain and Oh, they got my picture right. Thanks, guys. it's the supply chain and how do we do final assembly and where do we do final assembly and then where do we get the components? And my sense right now is tariffs are going to have an effect on this and it's unlikely that tariffs will not affect us even in you know one to five years there will be a tariff impact but there's also you know increasing geopolitical tension with China and in my history in my career the last 20 years we've made almost everything in China that I've made and so my guess is we're going to start to look at hybridization we have historically looked to China to help us think through our final designs and we've typically paid them if we're not going to go to volume with them something called NRE to to pay for their engineering efforts.

My guess is that will continue actually in many cases and the production may move out of China and and we may work with the same companies or different companies to produce those the the final assembly or or the components outside of China if we have to. But my guess is in some cases we'll just have to pay more for those components. So in my in my memory, you're probably the best positioned of almost anyone on earth because I think it was in 2019 that you were trying to get the Quest shipped and I don't know if you guys remember this, but in early 2020 we had kind of a global snafu in terms of our capacity to operate crossborder and ship things like you guys still shipped on time and on budget. Yeah. Yeah, the Quest 2 we did. So, we were in like halfway through the engineering builds December of 2019 and then of course we had COVID and we had to we had to ship anyway without being able to go there.

And then part of the fun is our display supplier was in the Philippines. The Philippines shut its borders. They shut down. Nothing happened in the Philippines for the next I think six months. So we had to move our our factory from the Philippines to China to make those displays and like we still we still had our schedule. So these things are possible. I think CO helped us realize that they're more possible than we ever thought because if you told us beforehand we would have said it was impossible, right? So I do think that there was some pressure from CO that we learned a lot about what can be done in supply chain. I hope that informs this this next phase. Do you think this kind of manufacturing ever moves to the US? I'm pessimistic on this construct. This is a great question. I mean, there is there is another exponential that's not as near to now and I don't know when this is going to happen, but I do think that there's a world in which we have industrial robots.

So, they're fixed robots, robot arms, the kind you see making cars, make robots. And then you really have some interesting questions. if you're not paying as much for human labor as like let me back up before COVID we were making the quest and then after CO we met went back and checked the factory and the factory land that had had a hundred people down it making putting components together turned into a line with four people and you had these cabinets and you had these automated things putting putting the product together and going down the line and it was incredible to see and it it happened all during co when we couldn't go there and By the time we got back, it wasn't I think it was the Quest 3 actually. It wasn't my product at the time, but we got to see it and it was automated and it that that was actually one of the catalyst for me come ending up where I am now, which is a realization that manufacturing is going to change significantly.

The benefit that we have though that's not going to change is the constellation of all the other stuff we need. And when we made the Mac Pro, we moved it to Austin, Texas from from China. And I called a guy on a Friday at 4 p.m. and I was like, "Hey, I need some custom screws." And he was like, "Well, I'm going home right now, but you know, maybe in 3 weeks we can make them for you." And like in China, if I called somebody at 4 p.m. on Friday, I had the screws the next day cuz they were like across the street, right? And so there's a there's a lot going on here where there's this constellation of part of the reason Shenzhen is such an incredibly successful area is you can do everything vertically right there. Yeah. So not only do we have to think about labor costs and labor availability is not just labor costs are people willing to do this in America right and I think in many cases they won't be.

So you got to automate that but then you have to think of the constellation of the components coming in and that's where I think it will take time to change supply chains. I think that's the challenge, right? It's that I I used to work in the aerospace industry and it's it's never the constraint of exactly how do you assemble the thing. It's how do you get the rated screws so that you can get to the tent and negative safety because every subcomponent has to be tested. Obviously, we're talking about probably different levels of safety requirements here. But similarly, I think you make an excellent point that the these all these systems are integrated and that for all of the advances that AI will enable us, including robots building robots. Yeah. we're still constrained to physical space. And maybe those constraints, whether they be tariffs or supply chain challenges, breed the kind of creativity that we need to see.

I think they will. I just don't think it will be very fast. Yeah. How traditionally, and again coming at it from my background as as a former engineer in the space and now an investor, we spend a lot of time trying to figure out how do you build the most efficacious team to go and run at the problem that you're trying to solve. And there have traditionally been basically a der of incredible talent in certain subspaces like frankly firmware engineers. How, if at all, has AI changed your hiring plan as you think about building a team here? Yeah, it's a great question. it hasn't changed my hiring plan. We've got many roles, including software roles, coming open, I hope, in in about a month. nothing's changing there, but I do expect that AI will change our ability to get more done with those people. And like I don't like right now so I come from consumer electronics into robots.

So I'm using OpenAI o3 to learn about all the things that I just hadn't picked up yet about robots, about how actuators work, how gearboxes work, trade-offs, all of that sort of thing. And I have definitely probably 4xed my ability to learn from using AI and I expect that will be the big difference. but you know, in hardware, you still got to make everything. You still got to make every component. You've still got to do the CAD. You still got to do the board layout. And AI hasn't come for our for our field yet. in a lot of ways, but on the software side, I would expect you can see a lot more efficiencies that way. So, smaller, better teams doing more faster. I mean, you really want agency and you want taste. Yeah. You want agency. You want somebody to be able to do things in a really efficient way, but you want the taste to know of all these ideas, which is the right one.

And those are the two things we're really looking for in addition to technical excellence. Yeah, I I think the point that Jobs made back in the kind of height of his time at Apple that, you know, the aesthetic choice is the correct choice. I that that appears to be even more true now that that kind of curation of decision- making is the critical difference in the outcomes. well, we're almost at time here. So, I think I will just say that I for one am wildly excited about what you may or may not be building here. And to you know, dispel, unfortunately, we can't speak specifically about what Caitlin is up to at OpenAI, but I know that I'm waiting with baited breath because you've only produced exceptional things in the past. And so, no pressure. Exactly. Thanks for we'll do our best. Expect nothing more from than that from you in the future.

and thanks for taking the time and thanks for being my friend and thanks for the work you're doing in the world. You're the best. Thanks and thanks for having me everybody. I think you guys are off to lunch. Yeah, I think it's lunch time.